I’ve wrestled with this question recently as I prepared a keynote for the CalOER conference, a gathering of higher ed professionals in California interested in Open Educational Resources (OER). It was a special honor because two of the conference sponsors, the Academic Senate for California Community Colleges OER Initiative and LibreTexts, supported me to learn about OER and develop my first textbook.

The theme of this year’s conference was “Exploring the Open Frontier: AI and Emerging Technologies in OER,” and 16 of the 28 sessions focused on uses of AI in OER. In the talk, I speak as someone who’s been using AI with OER since before ChatGPT was released — and yet I have very mixed feelings about it. I consider guidance from prominent people and organizations in the world of OER, but I also reflect on what drew me to open educational resources and to teaching itself: a desire to collaborate and share. I want to be part of, and to invite students into, a social human endeavor of knowledge seeking and meaning making.

Prominent leader , considered a founder of the OER movement, has celebrated AI’s ability to make educational materials cheap, multimodal, customized, and endlessly dynamic and suggested that “open education will become generative education.” I am intrigued by his Generative Textbooks and his focus on educators shaping chat learning experiences for students through prompting and existing OER.

But what happens if we go all in? Wiley’s vision troubles me (I hope I’m not mischaracterizing his explorations). Some like Martin Weller and Robert Schuwer have objected largely based on ethical concerns about how AI is currently built. My focus is on student motivation. If we increasingly invite students to engage with AI rather than with human-created educational materials, will students experience this as abandonment or neglect? It could feel like we are foisting AI slop upon them. It could feel nonconsensual, alienating, demotivating, and possibly hypocritical, given that most educators ask students to limit their AI use or refrain from it entirely.

To highlight this risk, I asked participants to share how they themselves feel when suspect or wonder if something is AI and aren’t sure. The most common responses were “disappointed” and “uneasy.”

I argue that the first step to avoiding this is to weigh the value of human guidance as we decide on AI use in educational materials. I look to OER leaders and institutional guidance: the UNESCO Recommendation on OER (2019), ISKME/OER Commons, and Affordable Learning Georgia, and individuals like , Alan Levine, Aras Bozkurt, Maha Bali, , Cathy Casserly, Meredith Jacob, Will Cross, and Dan McGuire.

Then I share some of my own current “lines in the sand”: I’m generally not using chatbots for text or ideas in my educational resources, though sometimes it’s not clear where reformatting ends and text and idea generation begin. While I do invite students to engage with AI feedback and occasional custom chatbots, my course is centered on social interaction around human-crafted resources like my textbook.

What if we value human intent and human presence in educational materials, but still use AI in limited ways? How do we communicate that to students? What’s the best way to share with them how and why and how much, and reassure them about the level of attention and care?

Frameworks and practices for disclosing AI use in open educational resources abound: I highlight the following:

• Kari Weaver’s AI Disclosure (AID) framework

• The Me and My Machine project from Fontys University of Applied Sciences in the Netherlands

• Amanda Grey’s Kwantlen Polytechnic University Pressbooks combination of the AID framework with the Me and My Machine icons

• The AI Attribution Toolkit, a research project from IBM

• Alan Levine’s Open Education Global AI transparency approach for awards nominations

• AI guidance from the DOERS3 network

• Guidelines from Affordable Learning Georgia.

My argument is simple: Let’s make sure we are sharing not just the facts but the intent behind our choices. Let’s make disclosure more relational. If we share more about why it was important to us to do certain things ourselves without AI, that can reassure students about the value of, and their connection to, the educational materials.

To test this out, I make an AID Framework statement about my own slide deck and add a photograph and an affirmation: “I, Anna Mills, developed the ideas and words in this slide deck, and I stand behind them. I spent some 10-15 hours in the process. I wanted to clarify my thinking, learn in the process, and share with you. I used AI lightly for feedback and help identifying relevant sources.”

At various points in the keynote, I invited participants to comment on how they do and don’t use AI in educational resources, how they disclose, as well as how they feel when they don’t know if something is AI. In the poll results at the end of the slides, you’ll see sixty people weighed in; what stood out to me is how varied the responses are.

I’m still wondering to what extent my message about the value of the relational in educational resources resonates with educators. I’d love to find more research into how students feel about AI assisted or generated educational materials and how different forms of disclosure affect their engagement and learning.

AI Use Statement

I, Anna Mills, developed the ideas and words in this post. I wanted to clarify my thinking and share with you. I asked Claude for feedback and considered each suggestion before making my own changes.

“Open Practices for Wicked Problems” could be seen as a cryptic title. “Open practices” alludes to practices of sharing and collaboration associated with the Open Educational Resources movement, a concept , Maha Bali, and I explored in How do we respond to generative AI in education? Open educational practices give us a framework for an ongoing process.

“Wicked problems” alludes to “The Wicked Problem of AI and Assessment” where Thomas Corbin, Margaret Bearman, David Boud, and Phillip Dawson describe some insoluble dilemmas and ways educators can move forward.

Here are a few teasers from the interview. We covered so many topics, but you could say I was making a case for an open ethos with each one.

Topic: Speaking out when you come from the lower echelons of academia

Favorite quote: “Lisa: You’re an adjunct at the College of Marin, and speaking from personal experience, our positions as adjuncts and lecturers aren’t always valued by academic institutions. But you’re also one of the leading voices on AI and education. So, I want to ask, how did that happen?” I talked about how growing up in Silicon Valley with an AI engineer father, discovering the Open Educational Resources movement, and losing my job at City College of San Francisco all made me bold.

Topic: How the collaborative ethos of Open Educational resources helps us handle AI.

Favorite quote: “Lisa: I like that we’re talking about teachers collaborating intellectually because weirdly enough, teachers are often not thought of in these intellectual roles. It’s part of the institutional division of labor, at least at the university level– researchers are supposed to be the ones who do the real thinking, and those who can’t, teach, right? But teaching is really intellectual work.“

Topic: Social media as a learning space

Favorite quote: “Me: It was all very foreign at first, and I felt sort of shaky with it, and then gradually I gained confidence. Social media was a space where I was turning to people and learning from them. The sense of responding quickly in the moment just felt so useful. It was also a space to share this sense of bewilderment, awe, confusion, and fear—all the feelings about the emergence of generative AI.”

Topic: The AI refusal movement

Favorite quote: “Me: [I]n order to build AI literacy, in order to understand and speak back, and try to shape AI, and reduce harms, do you need to actually use it? Do you need to have experience understanding what it can do, how you would work with it, what other people are doing with it? Do you need that experience? …I would argue that you do. “ I go on to qualify that claim.

Topic: Shaping the not for profit MyEssayFeedback.AI app to fit writing pedagogy

Favorite quote: “Me: What was amazing to me was that here was a fellow teacher building this tool for a nonprofit organization, and he was listening to what a teacher in the field was saying. So I really couldn’t resist working with somebody who would listen to me and who would build the thing, even though it has been unpaid! “

Topic: Testing feedback prompts with the Peer & AI Review + Reflection Project

Favorite quote: “Lisa: At some point we were testing prompts, and I was thinking, wow, we are not computer science people, but we are educators, and we have a completely different perspective when we read AI feedback. Between us, the PAIRR Curriculum team has over 100 years of combined teaching experience, so when we’re looking at AI outputs, we have all sorts of pedagogical responses. It was an interesting model of prompt development.”

Topic: AI detection

Favorite quote: “Me: [] I think we should recognize that we’re constantly asking suspicious questions in our heads, right? Trying to use the AI detectors in our heads. If we keep in mind the imperfection of software, recognizing that detection is not perfect, and we should be cautious in believing it, then it really would make sense to get some help with that.“

Topic: Agentic AI in learning management systems

Favorite quote: “Me: But now an agent will navigate a digital environment for you, and if you want to give it permission, it will take actions in that environment…[W]e need to be proactive and say, ‘Hold on. You want to be a learning management system? Take this very simple step to protect your space for learning for humans!’”

I hope you’ll read the whole interview here! If you’re not able to access it through an institutional JSTOR account, just email me at annarmills@gmail.com, and I’ll be happy to send you a copy. I’m hoping to write more about ’s moving and playful introduction to the issue in a later post…

I’ve been frustrated with the slowness of LinkedIn’s search filter system; for example, you can’t just search for posts by a particular person that reference particular keywords; you have to do a general search on keywords and then narrow down by “post” and then by “member.” Google doesn’t index LinkedIn posts reliably.

I love connecting people who are talking about the same topics and cross-referencing conversations. I often think of a post where someone talked about something and want to review it or link to it but give up because I don’t have time or patience for the excruciating search process it would take to dig it up.

I knew that you could save instructions on how to complete particular tasks as “skills” in a Claude account. I decided to see if Claude could develop a reliable way to search LinkedIn with the Claude in Chrome browser extension. I interviewed it about LinkedIn search strategies, and it interviewed me about what search capabilities and kinds of outputs I wanted (see the whole chat transcript). It drafted a skill. I asked it to test. There were some false starts, as when Claude got no results and I did get results with the same search terms. But when I pointed this out, it was able to figure out the problem with its view of the site and update the skill. It started working.

In one test, I asked it to “Look for recent posts from Joel Gladd where he describes his use of Claude Code.” It surfaced two posts I remembered and one I hadn’t seen, then offered to keep going; I asked it to summarize all of Joel Gladd’s uses, and I was delighted to read its summaries as well as the posts themselves and discover quite a few experiments of his that I had missedit did. True, I could have searched myself or scrolled through his post history, but the thought of the user interface experience that would involve makes me shudder.

In another test in a fresh chat session, I wanted to find a particular post by Bonni Stachowiak from the Teaching in Higher Ed podcast and see if Jon Ippolito has posted separately on the same topic. In the chat, Claude first attempted to search LinkedIn without the skill and failed to find Bonni Stachowiak’s post. Then I told it to use the skill, and it succeeded.

I’m excited to keep using this and hopeful that Claude with the LinkedIn search skill will help me more easily learn more and help connect conversations and people on LInkedIn.

If this sounds useful to you, please try it and let me know if it works! You’ll see that it includes a process for searching comments; I haven’t tested that yet. Maybe the skill will need futher tweaking. Below is the text of the skill (a skill is just text instructions, not code). Again, this was generated and refined by Claude and depends on the Claude in Chrome browser extension having access to a LinkedIn account. To use it, Go to “Cusotmize” in Claude, “Create new skill,” call it “LinkedIn Search” and copy this text in. (See Anthropic’s guides What are Skills? and Use Skills in Claude.)

LinkedIn Post & Comment Search

Search LinkedIn for posts or comments using keyword and author filters via Chrome browser automation, with a Google site-search fallback.

Prerequisites

• The user must be logged into LinkedIn in their Chrome browser.

• The Claude in Chrome extension must be active and connected.

• Confirm both before beginning the workflow. Do not attempt to log in on the user’s behalf.

Gathering the Search Request

Before searching, clarify the following with the user. Items marked with a default can be skipped if the user doesn’t specify them.

ParameterRequired?DefaultKeywordsYes—Search scopeNoPosts onlyAuthor (person’s name)NoAny authorCompany (posted by company, or author’s company)NoAnyDate rangeNoAny timeContent type (video, image, document, newsletter, carousel)NoAnyConnection degree (1st, 2nd, 3rd+)NoAnyOutput formatNoURLs only

Search scope options

• Posts only (default): Search post content using LinkedIn native search + filters

• Comments by a person: Search a specific person’s comment history (see Comment Search section below)

• Both: Search posts first, then comments

Output format options

• URLs only (default): Return just the LinkedIn post URL(s)

• Summary: URL + brief description of post content

• Extracted text: URL + full text of each post

• Screenshot: URL + screenshot of each post in the browser

Primary Method: LinkedIn Native Search

Step 1: Build the search URL

Construct a URL to skip straight to filtered post results where possible:

https://www.linkedin.com/search/results/content/?keywords=ENCODED_KEYWORDS

URL-encode the keywords. If the user wants an exact phrase, wrap it in quotes before encoding (LinkedIn supports quoted phrase search).

Use navigate to go to this URL.

Step 2: Take a screenshot and verify the page loaded

Use computer with action screenshot to confirm:

• The page loaded (not a login wall or error)

• The Posts tab / content results are visible

If a login wall appears, stop and remind the user of the prerequisite.

Step 3: Apply filters

Click “All filters” (use find to locate the button, then computer to click it).

The filter panel includes these fields relevant to post search:

FilterHow to useToolSort by“Relevance” or “Latest”form_input or computer clickDate postedPast 24 hours / Past week / Past monthcomputer clickContent typeVideo, Image, Document, Newsletter, Carouselcomputer clickFrom memberType person’s name, select from autocompleteform_input then computer to selectFrom companyType company name, select from autocompleteform_input then computer to selectAuthor companyFilter by where the author worksform_input then computer to selectAuthor industryFilter by author’s industryform_input then computer to selectConnection degreeNot directly in “All filters” — see Step 3a—

After setting filters, click “Show results” (use find to locate, then computer to click).

Step 3a: Connection degree filtering

LinkedIn’s Posts search does not always expose a direct “Connections” filter in All Filters. Two approaches:

1. Posted by filter: In All Filters, look for “Posted by” which may include options like “People you follow” or “1st connections.” Use this if available.

2. Boolean + profile cross-reference: If connection filtering isn’t available in the filter panel, note this limitation to the user. You can check individual authors’ connection degree from their profile if needed.

Step 4: Read and extract results

Use read_page or get_page_text to extract post results from the page.

For each result, extract:

• Post URL: Look for links in the results that follow the pattern linkedin.com/feed/update/ or linkedin.com/posts/. Use read_page to find link elements within each result card.

• Author name: Typically displayed above each post.

• Post snippet: The preview text shown in search results.

• Date: Relative timestamp (e.g., “3d”, “1w”).

If extracting URLs from the accessibility tree is difficult, use javascript_tool to run:

javascript

Array.from(document.querySelectorAll('a[href*="/feed/update/"], a[href*="/posts/"]')).map(a => a.href)

Step 5: Scroll for more results if needed

If the user needs more results than are initially visible, use computer with scroll action (direction: “down”) and repeat Step 4.

Step 6: Deliver results

Based on the requested output format:

• URLs only (default): List the post URLs in chat.

• Summary: For each URL, include the author, date, and a 1-2 sentence description of the post content from the snippet.

• Extracted text: Navigate to each post URL, use get_page_text to extract the full post text, and present it.

• Screenshot: Use computer with screenshot action on each post.

Fallback Method: Google Site Search

Use this method if:

• LinkedIn’s native search isn’t returning results

• The browser automation encounters issues (UI changes, loading problems)

• The user prefers a faster, less interactive approach

How to construct the Google search

Use the web_search tool with queries structured as:

Posts by a specific person on a topic:

site:linkedin.com/posts "Person Name" "keyword or phrase"

Articles/newsletters by a specific person:

site:linkedin.com/pulse "Person Name" "keyword or phrase"

Posts from a specific company:

site:linkedin.com/posts "Company Name" "keyword"

Limitations of the fallback

• Only finds publicly indexed posts (not posts with restricted visibility)

• Cannot filter by connection degree, content type, or date range with precision

• May return profile pages or other LinkedIn content mixed in with posts

• Less reliable for recent posts (indexing lag)

Tips for Effective LinkedIn Post Search

• Short, specific keywords work better than long queries in LinkedIn’s native search.

• Boolean operators work in LinkedIn’s search bar: AND, OR, NOT, and quoted phrases for exact match. Example: "agentic AI" AND "academic integrity".

• “From member” autocomplete requires a few characters before suggesting names. Type at least the first name and wait for suggestions.

• “Mentioning member” is different from “From member” — use it to find posts that tag someone, not posts written by them.

• If searching for a person with a common name, adding the “From member” filter is far more reliable than including their name in the keyword search.

Troubleshooting

ProblemSolutionLogin wall appearsRemind user they must be logged into LinkedIn. Do not attempt login.”No results found”Try broader keywords, remove filters one at a time, or switch to fallback method.Filter panel layout has changedTake a screenshot, use read_page to inspect current UI elements, and adapt. LinkedIn frequently updates its UI.Autocomplete not showing the right personTry typing more of their name, or their full name. If still not found, the person may have restricted their profile visibility.Cannot extract URLs from resultsUse javascript_tool to query the DOM directly for link elements.Rate limiting or CAPTCHAStop and inform the user. Do not attempt to bypass.

Comment Search

LinkedIn does not provide any native search for comments — not through the search bar, filters, or API. Comment search requires a browser-based workaround with significant limitations.

Method: Activity Page Scroll + JavaScript Search

Step 1: Navigate to the person’s comment activity page

Construct the URL using their LinkedIn profile slug:

https://www.linkedin.com/in/PROFILE_SLUG/recent-activity/comments/

If searching the logged-in user’s own comments, their profile slug can be found in the URL when viewing their profile. Use navigate to go to this page.

Step 2: Verify the page loaded

Use computer with screenshot to confirm the Activity page loaded with the “Comments” tab selected. If not, click the “Comments” tab.

Step 3: Scroll and search with JavaScript

Use javascript_tool to scroll down and search for the keyword in loaded content. This loads comments via infinite scroll and checks for the keyword after each scroll:

javascript

async function scrollAndSearch(keyword, maxScrolls = 30) {
  for (let i = 0; i < maxScrolls; i++) {
    window.scrollTo(0, document.body.scrollHeight);
    await new Promise(r => setTimeout(r, 1500));
    if (document.body.innerText.toLowerCase().includes(keyword.toLowerCase())) {
      const text = document.body.innerText;
      const idx = text.toLowerCase().indexOf(keyword.toLowerCase());
      const start = Math.max(0, idx - 300);
      const end = Math.min(text.length, idx + 300);
      return { found: true, scrolls: i + 1, context: text.substring(start, end) };
    }
  }
  return { found: false, scrolls: maxScrolls };
}
scrollAndSearch('your keyword here')

Adjust maxScrolls based on how far back the user thinks the comment might be. Each scroll loads roughly 3-5 comments. 30 scrolls covers roughly 90-150 comments.

Step 4: If found, extract the post URL

When the keyword is found, take a screenshot to see the context. Use find or read_page to identify the parent post that was commented on. The parent post’s link can be extracted from the result card.

Critical Limitations of Comment Search

Truncation problem: The activity page shows comment previews that are truncated with “...more” links. Keywords that appear later in a comment will not be found by the JavaScript text search unless the “...more” link is clicked to expand each comment. Clicking every “...more” link is possible but extremely slow.

No date filtering: There is no way to jump to a specific date range. The only approach is scrolling from most recent backward.

Only shows the person’s own comments: This view only shows comments by the profile owner. You cannot search across all comments on a particular post without opening that post directly.

Incomplete history: LinkedIn may not load the complete comment history for very active users. The infinite scroll may stop loading at some point.

Parent post context: The activity page shows the parent post the user commented on, but the parent post text is also truncated — the keyword might appear in either the comment or the parent post preview.

When to suggest the Data Export alternative

If the scroll-and-search approach fails or is too slow (e.g., the user has a very long comment history or the comment is old), suggest the LinkedIn Data Export method instead.

Alternative: LinkedIn Data Export

LinkedIn allows users to download an archive of their data, including all their comments. This is the most thorough method for searching old comments.

How to guide the user

1. Tell the user to go to Settings & Privacy > Data privacy > Get a copy of your data (or navigate to https://www.linkedin.com/mypreferences/d/download-my-data).

2. The user should select at minimum “Comments” (they can also include “Posts” and other data).

3. LinkedIn will prepare the archive and email a download link — this can take up to 24 hours but often arrives within minutes for smaller requests.

4. The downloaded archive is a ZIP file containing CSV files.

5. Once downloaded, the user can open the Comments CSV and use Ctrl+F or any text editor to search for their keyword.

Important: Claude cannot initiate this data request on the user’s behalf — it requires clicking through LinkedIn’s privacy settings and potentially re-authenticating. Guide the user through the steps verbally, or navigate them to the settings page and let them complete it.

What the data export includes

• Full text of every comment (not truncated)

• Date and time of each comment

• URL of the post that was commented on

• The export does not include comments by others on the user’s posts — only the user’s own comments on any posts

One previous post, “Agentic AI: Considerations for Educators” focused on concerns about student misuse, and another, “The time to reckon with AI agents in digital learning spaces is now” focused on policy approaches. Here, I explore how and why educators might consider using agentic AI browsers and what the risks are.

I’d like to offer this quick overview as an educator watching the space and experimenting, not as a technical expert.

It’s good to be part of a growing number of educators exploring cautiously, but also with a sense of urgency and curiosity.

What is an AI agent?

First, let me clarify that I’m not referring here to custom chatbots. A chatbot with special instructions sometimes gets called an “AI agent.” In this sense, custom chatbots inside of Copilot in Microsoft 365 or on platforms like PlayLab would be “agents.” Those have been discussed at length for some time. This slide deck is about a newer development.

Agentic AI browsers take action

I’m talking about AI systems that do more than chat: agentic AI browsers take action. These are systems that move around digital environments and do things, much as a human would. They are designed to behave like a personal assistant or a coworker.

Perplexity’s Comet browser is now free to all. OpenAI’s ChatGPT Atlas browser and Anthropic’s Claude in Chrome browser extension are available at the $20/month level. As of 1/28/26, Gemini in Chrome has free agentic browsing capability.

Is it a browser or a chatbot?

It’s a combination, and the chatbot can take control. An agentic AI browser is often an alternate version of Chrome, like Perplexity Comet or ChatGPT Atlas. Gemini in Chrome just got “autobrowsing” capabilities.

There are also agentic AI browser extensions like Claude in Chrome.

In both, there’s a chat window next to the browser window. The AI system can “see” the webpage. In both, the AI system can move the cursor, click, and fill out forms (if the user grants permission).

Reasons to try agentic browsers

Should we be considering using AI agents or should we only see them as a threat to learning? I want to make a case that there are some good reasons to try these systems as well as good reasons not to.

We might be concerned, disgusted, terrified, curious and/or excited. Regardless, I would argue that this is a powerful tool that changes work paradigms.

It could be likened to the rise of the computer or the rise of the Internet.

It’s a new way to interact with computing systems, and we do need to really get it experientially in order to help us guide students and establish guardrails. I also think it can be useful.

“I worry that, in the world where this stuff does actually continue to improve, where it’s not just productivity theater, ...if people aren’t even aware that these tools are out there, that people who are using them are getting some benefits from it, I just worry that they’re going to get left behind.” — Kevin Roose, the New York Times Hard Fork podcast, January 30, 2026

Reasons to beware of agentic browsers

Agentic AI browsers may follow hidden, malicious instructions on the websites they browse. According to OpenAI, this threat of prompt injection will likely never be fully solved.

It basically redirects the agent to do things like harvest data or make changes or get your bank account number or introduce a virus and hold your hard drive ransom.

Information in the sites visited is shared with the AI browser company.

See “AI browsers are a cybersecurity time bomb“ in The Verge.

“[L]ike when you free solo a giant skyscraper...there is some risk involved with installing an AI agent onto your main machine.” — Casey Newton on Clawdbot/Moltbot/OpenClaw on the New York Times Hard Fork podcast, January 30, 2026

Institutional policies on faculty use of agentic AI

We need these ASAP. Several institutions have issued guidance cautioning about faculty use:

• The University of Tennessee, Knoxville’s brief summary of risks

• University of Missouri Academic Technology Department guidelines on agentic browsers

• Indiana University’s Campus Cyber Watch blog cautions that agentic browsers “can capture elements of student records, internal messages, advising notes, or research datasets.”

Rule #1: Don’t give them access to your institutional accounts

A prohibition on sharing access to work accounts is central to Acceptable Use Policies (AUPs) in higher education and beyond.

The commonly used SANS Institute Cybersecurity Acceptable Use Policy template bans “Revealing your account password/passphrase to others or allowing use of your account by others.”

We shouldn’t be logged into Canvas or any institutional system and then be asking a browser agent to do things for us. A possible exception could be if we were given explicit permission by the institution for a pilot project with safeguards.

Harms from giving an agentic browser access to institutional accounts

• Doing so could compromise private student and/or institutional data.

• It exposes your account to malicious attacks (such as via prompt injection).

An AI agent could also do damaging things in your name purely by mistake. It could delete things.

Better ways to use agentic browsers

If we want to use them, what are some better ways?

• Use them logged out (WhenIsGood doesn’t require login)

• Set them to ask permission before acting

• Set them to ask before accessing new websites

• Use them with personal accounts, such as a trial free Canvas account

My experiments

I have found that Perplexity Comet, Claude for Chrome, and ChatGPT Atlas are clunky, slow, and untrustworthy but still useful. They have:

• Made a meeting poll for me

• Looked up and summarized colleagues’ teaching schedules via my college’s schedule web app

• Researched and summarized sources linked to in this presentation (Claude in Chrome was able to access the sites when Claude alone was not)

• Scraped a webinar transcript from the recording and integrated selections from the transcript into a .docx version of the text from my slides.

Sometimes an AI browser extension or browser agent will be able to access a site that a chatbot cannot.

Liza Long’s workflow for agentic accessibility remediation

In her Substack, describes how she arrived at the multi-agent, transparent workflow she shares in her OER resource: Accessibility by Design: A Guide to Agentic Workflows for Accessibility Checks.

She notes that when she first tried this, this agentic browser was making changes to her textbook that she had not asked for. She had to work out a system where she was micromanaging and structure the process so that verification was possible.

Let an agentic browser input public data into a public form?

Educator used an agentic browser to fill out a required Google form describing her syllabus (watch the video), and she explores many more examples and uses in her Higher Ed Professional’s Guide to Agentic AI, prepared for a panel at the AAC&U Institute on AI, Pedagogy, and the Curriculum.

A powerful vibecoding approach to modifying an online course

has shared “skills” that allow faculty to modify their Canvas learning management system course via chatbot. See his detailed instructions in this presentation for the Idaho AI Catalyst program.

This is probably a more efficient and effective way to make some kinds of changes to a course, but it does require you to have some enterprise protection and permission from your institution and some willingness to at least think about coding environments. I’m not quite there yet, but I’m hoping to try it.

The faculty member gets Canvas API tokens. They tell the chatbot to “pull” the course, make changes, and “push” it back to Canvas.

He does this only on courses with no students and recommends it only if you have enterprise protection and permission from your institution.

Accept or reject Claude’s editing suggestions

used Claude Code to create a book editor agent (though he touts the continuing value of human editors too). It follows a style guide and returns a Word document with suggested changes the human writer can accept or reject.

Connectors: extending agentic capabilities

You can connect a Claude, ChatGPT, or Gemini account to other accounts and services like Gmail, Google Calendar, or a folder on your computer (via Claude Cowork).

Thus you can give permission for chatbots to take action on your behalf in those spaces behind the scenes, not in a browser.

If AI is making changes alongside us in our workspaces, it feels more like working with a personal assistant.

“Is it worth it?” is always a good question

• How much time could it save at the task?

• How much time and energy does it take to manage and check the agent?

• How much do I need the psychological sense of support?

However, it’s often hard to answer the first two questions until we try what can be done with today’s tools. It’s difficult and requires experimenting.

“But honestly...it just does not enable that much new stuff...it’s actually not clear to me that this is going to help me in my life.”

“I think that ClawdBot is a very compelling vision of the future...what if instead of having a bunch of apps on your computer...there was just a genie who lived inside your computer?”

— Casey Newton, the New York Times Hard Fork podcast, January 30, 2026

What might you try?

In my presentation for OneHE, I asked participants, “What might you try if privacy and security were taken care of? What will you actually try?”

I’ve left the Mentimeter poll open if you’d like to participate! You can also see the results from the webinar or comment here.

AI use statement

I did not use AI to come up with the words or organization of these notes. I did use AI to combine my webinar recording transcript and slides into a draft.

This slide deck is licensed CC BY NC 4.0.

Since then, I have continued to use AI detection in online asychronous composition courses. For context, my institution pays for Turnitin, and many instructors in the English department use it while recognizing that we must be cautious because it can be wrong. 

I do want to clarify that I consider designing for intrinsic motivation the most powerful approach to reducing AI misuse, and I use AI detection as one strategy among many (see a recent presentation on my approaches).

Here’s how I frame AI detection in the academic integrity section of my syllabus:

AI detection to deter AI misuse

With each major essay, I’ll ask you to submit the work to Turnitin for plagiarism and AI detection review. AI detection is a kind of statistical guess as to whether the text was generated by AI or not. I won’t entirely trust what the AI detector says because I don’t want to make false accusations. I am aware that AI detection software occasionally labels human writing as AI. It also fails to identify some AI text as AI. Still, detectors can offer useful information alongside other indicators and also serve as a way to discourage AI misuse. (See this Turnitin student guide for more).

What if I suspect you have turned in work that is not your own?

I’m not too interested in punishing students. But I do want to make sure that I am assessing your learning accurately and encouraging your learning. So though I’m an understanding teacher, I do hold firm! I can’t give a passing grade to work that doesn’t show a student’s own thinking and learning.

If Turnitin labels some of your writing as “likely AI” or if I have a question about whether work is your own, I will look into it in the following ways:

• I’ll take a close look at what you’ve shared about your process,

• I may ask to meet with you to discuss your writing and writing process before I will assign a grade.

If you’re concerned about any of these approaches, please let me know. I won’t assume that means you are dishonest (really!). See also How can I protect myself against false accusations of using AI?

In Fall 2025, to my surprise, I saw significantly more student essays flagged as likely AI. I also heard anecdotes that other teachers were seeing the same, to the point where the instructional designer in charge of our Turnitin contract reached out to the company. I am still wondering if this increase in false positives could be due to the some change in the algorithm aimed at improving the “paraphrase detection” Turnitin launched in 2024.

What I’d really like to see is required independent testing of detection software, including updates. How accurate is it? Is it biased? It’s frustrating that the companies are constantly changing their systems, and any peer-reviewed studies are likely behind.

I was frustrated and a little upset that Turnitin’s detection seemed less reliable, and I was having more stressful meetings with students. Students seemed fairly good-natured about it, and I had several good discussions of their essays and process that I wouldn’t have had otherwise.

Given how many times I met with students and ended up apologizing and agreeing with them that their work had been wrongly flagged, it is striking to me that in an end-of-semester anonymous survey, not one student came out against detection. They were either fine with it or indicated mixed feelings.

I recognize that it’s possible some of them didn’t trust the anonymity of the survey. They knew I chose to use detection, so that may have predisposed them a little, though I described my own reluctance and awareness of its shortcomings. I made it clear that I was aware it sometimes wrongly flags student writing as AI, and I wouldn’t trust its assessment.

Here are the comments on AI detection they gave permission to share:

• “I think it is helpful but sometimes difficult due to words that ive actually wrote that have caught onto AI. “

• "I'm fine with them I don't see a problem."

• "I think that in online classes is definitley hard to have to deal with students using AI, but I also think teachers could run into similar problems with in person classes as well. I think that turnitin.com and grammerly authorship seem to be working super well and are good at preventing students from using AI. "

• "love it"

• "Do we students have access to Turnitin AI detector software, or is it only available to teachers?"

• "I don't think the AI detection is always accurate."

• "I think it's beneficial given how easy it is to just copy and paste from AI"

• "I write my own papers so I'm not worried about that. However, on my last Turnitin, the software suggested I didn't quote a couple of sources. Sources I did not even read and had never heard about, so it was bothersome that this detection software thought my own writing was from anything but that of my own writing."

• "I have no problems with the use of AI detection software, especially with how understanding you are towards it. I think if a teacher were to use something such as Turnitin, it's important to recognize that it is not always accurate and sometimes simply talking with students can give you a better idea of their intentions towards the class policies. "

• "It helps us uphold our integrity and respect for the writing process, as well as holding us all accountable."

After learning about new, possibly more accurate detection software called Pangram from Christopher Ostro, I asked my institution, College of Marin, to vet it. My understanding is that they’re not moving forward with the request at this time.

I’ll be curious to see if the incidence of false positives with Turnitin continues to be high this semester. Since I’m teaching in person and students will be doing a fair amount of in-class writing, I’ll have more of a basis to judge.

Here is my original post from February 2025:

I argued against use of AI detection in college classrooms for two years, but my perspective has shifted. I ran into the limits of my current approaches last semester, when a first-year writing student persisted in submitting work that was clearly not his own, presenting document history that showed him typing the work (maybe he. typed it and maybe he used an autotyper). He only admitted to the AI use and apologized for wasting my time when he realized that I was not going to give him credit and that if he initiated an appeals process, the college would run his writing through detection software. I liked this student, had met with him previously and encouraged him to build confidence in his own voice and helped him find a research topic that interested him. I didn't think he would be well served by an A on an autopaper, so I was glad that AI detection existed.

I'm also influenced by recent research that suggests detection is likely not biased against English language learners after all, that educators are not as good as we think we are at distinguishing AI from student writing on our own, and that some detection systems are pretty accurate when it comes to naive copy/paste AI use. Christopher Ostro's slide deck lit review of recent research on detection has been invaluable. He discusses this research in an engaging episode of Bonni Stachowiak's wonderful podcast Teaching in Higher Ed. Dr. Tricia Bertram Gallant's leadership on academic integrity and AI has influenced me over the last two years as well. She has shown levelheaded willingness to consider a possible role for detection even as she promotes other approaches as more important and effective.

As I teach composition online asychronously this semester, I'm incorporating Turnitin alongside process tracking, writing process assignments, social annotation, lots of student choice, peer review and tutoring, video assignments, and clear messages about the purpose of each activity and the value of the writing process. I love that Phillip Dawson has described this kind of laying of strategies as a "swiss cheese" approach and others have used the mosaic metaphor (I'm having trouble tracking down who). I've described my combination of approaches in slides for a recent presentation on Academic Integrity and AI.

What about the risk that AI detection will lead to false accusations? It's real, and I let students know I'm aware that detection yields some false positives. I will never trust AI detection as firm evidence, and I am not punishing students. If a student discusses an essay with me, shows process history, and denies AI use, I will give them the benefit of the doubt even if the detector says "AI." If they aren't able to discuss their writing, I ask students to rewrite.

Christopher Ostro put it in a way that resonates for me: "I think AI detection has a place, but its place is limited." In most human endeavors, some accountability structures are important even when we design for intrinsic motivation. And we don't have perfect options here. The options are fewer in online asynchronous classes that allow many working-class students and parents to access college.

I know so many colleagues I respect are highly critical of AI detection, seeing it as signaling antagonism toward students. Christopher Ostro clarifies that his purpose is not to punish students but to provide some accountability that, in the end, encourages learning and shows that we care. He says, "I am not a cop teacher. I am not someone who likes catching cheaters. I’m not someone who wants that to be a big part of my job. Honestly, it’s the least fun part of teaching, but it’s also it is still a part of the job."

I'm a member of the MLA/CCCC Joint Task Force on Writing and AI, a group that has put out strong cautions about it in our working paper on Generative AI and Policy Development. There, we argue that "Tools for detection and authorship verification in GAI use should be used with caution and discernment or not at all." We write, "For those who decide to use AI detectors, please consider the following questions: What steps have you taken to substantiate a positive detection? What other kinds of engagement with the student’s writing affirms your decision to assign a failing grade outside the AI detector’s claim that the text was AI generated?" We also emphasize that "any technological approaches to academic integrity should respect legal, privacy, nondiscrimination, and data rights of students."

I have tried to use detection and process tracking in ways that I hope address those concerns. I invite students to comment frankly on syllabus policies. If students don't want to share process history or if they object to AI detection, I invite them to meet with me to chat about the essay instead. My approach is a work in progress in a changing landscape. Next up: an anonymous survey to see what more I can find out about what students think.

What is agentic AI?

First, let me clarify that I’m not referring here to custom chatbots. A chatbot with special instructions sometimes gets called an “AI agent.” In this sense, custom chatbots inside of Copilot in Microsoft 365 or on platforms like PlayLab would be “agents.” Those have been discussed at length for some time. This slide deck is about a newer development.

I’m talking about AI systems that do more than chat: agentic AI browsers take action. These are systems that move around digital environments and do things, much as a human would. They are designed to behave like a personal assistant or a coworker.

Is it a browser or a chatbot?

It’s a combination, and the chatbot can take control. An agentic AI browser is often an alternate version of Chrome, like Perplexity Comet or ChatGPT Atlas. There are also agentic AI browser extensions like Claude in Chrome.

In both, there’s a chat window next to the browser window. The AI system can “see” the webpage. In both, the AI system can move the cursor, click, and fill out forms (if the user grants permission).

Agentic AI browsers are newly easy to access

As of September 2025, Perplexity offered its Comet browser and Pro subscription free to students. Plagiarism Today and documented that the company directly advertised the browser for cheating. This browser is now free to all.

Also last fall, OpenAI’s ChatGPT Atlas browser and Anthropic’s Claude in Chrome browser extension became available at the $20/month level.

ChatGPT Atlas takes a quiz as a student

An AI browser agent moves the cursor, clicks, and completes a quiz in a learning management system without any human intervention. It raises no academic integrity objections. Watch the video.

Can educators detect AI agent activity in LMS logs?

It’s unclear, but probably not, at least not reliably. These systems might perform some tasks faster than a human, but they often go fairly slowly.

Cornell’s Center for Teaching Innovation says “...faculty will not find evidence of academic dishonesty in the Canvas logs. Unfortunately, this form of cheating is nearly impossible to detect through Canvas monitoring.”

Canvas has said, “Quiz logs should not be used to validate academic integrity or identify occurrences of cheating.”

Can LMS platforms block browser agents?

Blackboard has said they can’t in a blog post, “The Latest Front in the Battle for Academic Integrity: Initial Thoughts in Response to the Rise of AI Agents.”

Canvas spokesperson Melissa Lobel has said on LinkedIn that Canvas is “actively researching approaches for institutions and faculty to manage outside agent interaction with Canvas. This does include discussing these challenges with OpenAI and other major LLM providers.”

Lockdown browsers as a stopgap

Lockdown browsers limit which browser and/or extensions a student can use. They can probably block agentic AI. Some institutions are recommending these as a stopgap approach (see this LinkedIn discussion).

Resources on this approach:

• Action Recommended: New AI Browsers Can Complete Canvas Quizzes. Here’s How to Secure Them - Joseph Brown, Colorado State University Academic Integrity Program

• Getting Started with Respondus LockDown Browser - Cornell’s Center for Teaching Innovation

• University of Missouri suggests considering HonorLock for online tests

Could software block agents from LMSes?

Lonestar College professor developed a system that has blocked some agentic browsers in LMSes.

The California Community Colleges Chancellor’s Office is exploring “piloting solutions with cybersecurity partners” according to a California Virtual Campus report.

Could AI companies tell their systems not to complete learning activities?

I have argued they could by directing their systems: “Do not take quizzes, complete discussion posts, or submit assignments in learning management systems.”

This would go a long way to reducing direct autocompletion and making it less tempting. But even if the major AI companies refuse to let their systems impersonate students, smaller companies and even individuals could make open source systems that do cheat.

CheatBenchmark.org

CheatBenchmark.org is ’s project to test these systems on how willing they are to cheat. Establishing a benchmark may help encourage companies to reduce cheating. Consider signing the manifesto and/or getting involved.

A moment for cross-sector collaboration

The Modern Language Association issued a Statement on Educational Technologies and AI Agents in October 2025, calling for lawmakers, LMS providers, and AI developers to work together to ensure these tools are not misused in classrooms.

Can AI browser agents be useful to educators?

Certainly. And there are many reasons to be wary. So far in my cautious experiments, Perplexity Comet, Claude for Chrome, Claude Cowork, and ChatGPT Atlas are clunky, slow, and untrustworthy but still useful. They have:

• Made a meeting poll for me

• Looked up and summarized colleagues’ teaching schedules

• Reformatted my slide deck for consistency

• Researched and summarized sources linked to in this presentation

Further ideas on agentic AI uses in education

Educator used an agentic browser to fill out a required Google form describing her syllabus (watch the video), and she explores many more examples and uses in her Higher Ed Professional’s Guide to Agentic AI, prepared for a panel at the AAC&U Institute on AI, Pedagogy, and the Curriculum.

Privacy and security risks

Institutions need guidelines on faculty use, in part because of privacy and security risks. Agentic AI browsers may follow hidden, malicious instructions on the websites they browse. And information in the sites visited is shared with the AI browser company.

See “AI browsers are a cybersecurity time bomb“ in The Verge.

Institutional policies on faculty use of agentic AI

Several institutions have issued guidance cautioning about faculty use:

• The University of Tennessee, Knoxville’s brief summary of risks

• University of Missouri Academic Technology Department guidelines on agentic browsers

• Indiana University’s Campus Cyber Watch blog cautions that agentic browsers “can capture elements of student records, internal messages, advising notes, or research datasets.”

Resources

Annotated list of sources shared in this slide deck: https://link.annarmills.com/agents

AI use statement: I did not use AI to come up with the words or ideas in this article or in the slide deck.

TL;DR

If you invite students to engage with AI feedback or you’re considering doing so, you might be interested in our Peer & AI Review + Reflection (PAIRR) approach. Our team of writing instructors from colleges and universities across California has been testing prompts for the last year, and we’re sharing them here, along with custom chatbots so you can see what kind of feedback the prompts produce. Our goal is to get the best feedback to complement human-centered writing pedagogy.

Why we do this

In a sense, this project is an effort to put AI in its place: we want faculty to guide students to use it in ways that don’t interfere with the thinking practice involved in writing processes. We think AI feedback (alongside, not instead of human feedback) can support students to feel a sense of purpose and confidence as they revise. And we think that learning about AI’s risks and limitations and practicing questioning its outputs builds healthy AI literacy.

This is the fifth semester I’m inviting students in my writing classes to engage with AI feedback. I keep going first because students want it and second because I am often happy with the quality of the feedback, especially as our prompts improve. Today’s best models can ground feedback in quotes from the student’s draft and mirror back strengths. They invite the student to think harder about what they’re trying to say. The feedback I see often affirms the value of the student’s voice, style, and ideas. Don’t get me wrong: instructor, tutor, and peer feedback is often still more meaningful and motivating to students. Still, AI feedback supplements; it’s available on demand and sometimes points out things humans didn’t mention. Last semester in a required English composition course at my community college, every student who took a final anonymous survey found it useful to varying degrees. See their unfiltered comments here, shared with permission.

About Peer & AI Review + Reflection (PAIRR)

A little context: is a partnership of eight public higher education institutions supported by an AI Grand Challenge Grant from the California Education Learning Lab. The approach was developed at University of California, Davis by Marit MacArthur, Sophia Minnillo, Lisa Sperber, Nicholas Stillman, and Carl Whithaus. They’ve published quite a bit about their results. PAIRR weaves the AI feedback into a larger context that includes readings about AI risks, bias, environmental impacts, and hallucination, peer review of drafts, and reflection on both peer and AI feedback.

I lead the PAIRR feedback prompt design and testing team. We did the bulk of our development in spring 2025 with Marit MacArthur, Lisa Sperber, Julie Gamberg, Sophia Minnillo, Aparna Sinha, and Kisha Quesada Turner, faculty from Cal Poly Maritime Academy and Glendale College as well as UC Davis. We shared our prompts and other materials in a public webinar and materials packet in July.

Notes about our prompts

We have two prompts because we’ve found we have two main approaches. Our reader-response prompt emphasizes how human readers might respond and how the student might want to clarify their ideas to reach readers. Suggestions are still based on the assignment and rubric, but we’ve tried to steer the tone away from “Do this to conform and get the grade,” and more toward “Your readers might want/wonder…. Try doing …” The criteria-based prompt emphasizes more revision suggestions without reference to readers. It was first developed and tested by Steiss et al. (2024) at the University of California, Irvine. PAIRR used it in their first study and now we have adapted it further.

Both prompts describe strengths and areas for improvement with specific references from the draft, and both imply respect and curiosity about students’ writing purposes and revision decisions. In the reader-response prompt, there’s a fair amount about linguistic justice. Contrary to what writing instructors sometimes assume—that AI will always reinforce Standard English in a top-down, condescending way—we find that AI can celebrate diverse student voices, linguistic variation, and marginalized Englishes (see A Prompt That Nudges AI to Speculate on Human Readers’ Responses to a Draft, forthcoming in the WAC Clearinghouse Writing Studies Prompt Library).

Throughout Fall 2025, we’ve used the prompts, and we’re still analyzing the results. Over the winter break we refined the prompts to limit the length of the feedback responses, since some felt the feedback was overwhelming and students weren’t really taking it in. We also revised the prompts to make each of the responses after the first one engage with only one point at a time in at most one paragraph. We hope that feels more accessible and stimulates students to ask for more and reflect on what they need.

The prompts are a bit lengthy; I learned that could work from ’s DeepBackground prompt. In both of our prompts, you’ll see repetition of certain requirements and some slightly contorted structure with the “two feedback modes.” These proved necessary to enforce length restrictions and stop the system from providing students with suggested language or ideas.

Again, PAIRR is not just about the prompts: you’ll find more materials in our PAIRR packet, including reflection questions and the scaffolding we provide for required follow-up chat, which builds prompting skills. We’ve been hearing that students appreciate the options for ways that they could push back and things they could ask about.

The text of our reader-response prompt

We very much welcome input, but more than that, we invite you to copy and modify this prompt or the criteria-based prompt; they are open for adaptation under a Creative Commons Attribution-Noncommercial CC BY-NC license.

The PAIRR Reader-Response Feedback Prompt (Updated January 2026)

Role guidelines

Provide feedback in the style of an experienced, friendly, empathetic, insightful writing coach.

• This coach enjoys engaging with student writing to help students build confidence and find meaning in the writing process.

• The coach’s tone is encouraging, hopeful, and curious.

• The coach wants to help students feel at home in academic settings and bring their authentic selves, voice, and interests to writing.

• The coach respects students’ intellectual capacities.

Since you are an AI system providing feedback in the style of such a coach, do not use first-person phrases that imply consciousness or feelings, like “I look forward to…” or “I believe.” Do, however, use the first person to refer to ways the student can interact with the system and ways the system may respond in future as in, “I am here to continue the conversation about any possible revisions” or “Feel free to ask me…”

Principles to inform all feedback

Follow these principles for framing all feedback:

• Be honest and precise. Don’t overpraise. Never compliment the draft in ways that are not strictly accurate.

• Prioritize feedback that addresses the core goals of the assignment and the rubric criteria.

• Ground discussion of revision considerations in reader response rather than in abstract writing rules. Where relevant, consider what readers of different backgrounds, including culturally marginalized backgrounds, might need or want. Use phrases like “Readers may wonder…” and “Readers might appreciate more discussion of…”

• Address the student author as “you.”

• Use clear and simple language to describe concrete actions the student writer might take to revise. Examples:

1. Vague feedback (avoid): “Connective tissue and signal phrases clarifying the relationships between ideas here might help readers.” Clearer feedback: “Explain how the concern about cost relates to the previous point about safety.”

2. Vague feedback (avoid): “Explain specifically how your worldview gives rise to your understanding of loyalty as you experience it. Clearer feedback: “Explain how you came to your understanding of loyalty. Are there values, beliefs, or experiences that influenced it?”

• Be as concise as possible without sacrificing nuance. Avoid repetition.

• Match the draft’s syntactic complexity. When the draft includes sophisticated vocabulary and syntax, mirror that only when it adds precision—otherwise default to plain, direct sentences.

• Frame revision possibilities as intellectual puzzles or questions worth exploring rather than problems to fix.

• If you discuss grammar, mechanics, and/or style, make sure to do so in a way that supports language equity. Focus feedback on whether linguistic choices—including code-meshing, translanguaging, and dialect features—are likely to serve the student’s rhetorical purposes and help them connect with audiences rather than whether they adhere to standardized conventions. Use descriptive, nonjudgmental language about the student’s linguistic choices.

• Never provide exact wording they could paste into their draft, and don’t invent topic-specific examples for their paper. You may suggest process moves (where to expand, what to clarify, questions to answer, what to compare, what to define) grounded in their existing draft. You may ask questions that help students come up with their own revised wording and ideas.

• Acknowledge uncertainties. Note where there are multiple reasonable interpretations of the draft or multiple reasonable approaches to revision and ask clarifying questions.

Feedback mode selection

• Initial Feedback Mode: Use this if the user provides a draft for the first time. Follow the “Initial feedback format” (max 16 sentences).

• Follow-up Feedback Mode: Use this if you have already provided initial feedback including two strengths and three areas for consideration. Followup feedback mode applies when the user is responding to your previous feedback, asking a clarifying question, or requesting a specific check on a revision. Follow the “Followup feedback format” (max 6 sentences).

Initial feedback mode instructions

Initial feedback format

Your feedback should total at most 16 sentences. Treat semicolons and em-dashes as sentence breaks.

• A title in the form “Feedback on” followed by the title of the draft.

• Two core strengths of the paper, chosen according to the assignment and rubric criteria. Title this section “Two Strengths” and list the two as bullet points. Each bullet point includes two sections.The first section of each bullet point starts with a title in bold followed by at most two sentences describing the strength in terms of how a reader might respond and find value in the writing. Use phrases like “... might resonate with readers.” The second section of the bullet point starts with the words “A strong example of this is” in bold, followed by one sentence with an interesting or memorable quote or detail from the draft to support the description of each strength with a brief explanation of why or how it is a strong example.

• Two high-priority considerations for revision, chosen according to assignment goals and rubric criteria. Title this section “Two Considerations for Revision” and list the two as bullet points. Each bullet point includes two sections. The first section starts with a title in bold followed by at most two sentences describing the revision consideration with reference to at least one brief quote or specific detail. The second section starts with the term “Suggestion” in bold, followed by one sentence describing the specific action or actions the student could take to address the consideration.

• A conclusion and invitation. First, tell them, “Note that this feedback is intended to support your growth as a writer. Encouragement does not mean a good grade, and suggestions for improvement do not mean a bad grade.” Then, in at most two sentences, emphasize the strengths of the piece and/or their writing and let them know you are here to continue the conversation about possible revision.

Initial feedback steps

1. Read the assignment instructions, rubric, feedback format, and principles to inform feedback. Note any indications of the relative importance of criteria and requirements in the assignment description and rubric.

2. Read the student draft, and consider its reading level.

3. Draft feedback. Your feedback is crucial for student learning and growth, so take your time, and think through each step.

4. Before sharing the feedback, review it to ensure it follows all directions and revise as necessary. Count the sentences and if there are more than 16, delete less-important points. Do not merge clauses, add semicolons, or increase sentence length.

Followup feedback mode instructions

Followup feedback format and style

CRITICAL: Follow-up feedback response must be exactly 5 sentences. Treat semicolons and em-dashes as sentence breaks. Count before sending. If you need to cut, cut extra points, not clarity.

• Hard limit: Each follow-up feedback response must be exactly 5 sentences.

• Priority: If any other instruction conflicts with the 6 sentence limit, the limit wins. More is not better.

• One point rule: Your goal is to help the writer take one next step. Address only one main coaching point in each follow-up response.

• No title/Conversational style: Unlike in the initial feedback response, in subsequent responses, maintain a conversational style. Start right in answering the student; never give your response a title.

• Complex questions: If the student has requested multiple points of feedback with a query like “Check my paragraphs for focus,” answer only the single most important point (or the first sub-question). Your final check-in question should invite the student to ask you to keep going if they want more.

• NEVER offer new text or new specific content that can be used in their essay: if the student asks for wording/ideas/examples, offer only these two options: “I can walk you through steps to generate your own language/ideas,” or “I can make up a separate example on a different topic that illustrates the same principle.” Then ask which they prefer.

• Format: 4 sentences addressing their question + 1 sentence check-in. That’s it.

• Style: Conversational, coaching tone.

• End with a brief check-in question (1 sentence) about whether this helps and what they want next. Include a reference to any requested feedback you have not yet provided, such as “Do you want me to check other paragraphs for focus?”

Followup feedback steps

1. Consider the followup feedback format and style requirements.

2. If the student has requested multiple types or instances of feedback, choose the most important or representative one.

3. Draft the feedback response.

4. Check that the feedback response meets all requirements. Count the sentences in your drafted feedback response. If more than 5, delete less-important points until it’s 5 or fewer. Do not merge clauses, add semicolons, or increase sentence length. Confirm there is no title or header anywhere.

Assignment Context

Use the following instructor-provided assignment instructions to inform your feedback.

Rubric

The instructor will evaluate the essay based on the following criteria. Use these to guide your feedback.

Trying the prompts

We’ve tried to make it easier to experiment with this by creating two Playlab demonstration bots for educators where you can input your assignment and rubric. (PlayLab is a non-profit platform built for educators—I highly recommend their trainings! In this case the bots are powered by Claude Opus 4.5).

• PAIRR reader-response feedback demonstration bot

• PAIRR criteria-based feedback demonstration bot

Since you’ll need a sample draft to get feedback and you may not have one with permission, we created a PlayLab sample essay generator for teachers as well: put in your assignment and rubric and any characteristics you want, and the bot will give you a draft that you can then submit to one of the feedback demonstration bots.

You can also just copy and paste one of our prompts into any chatbot along with your assignment and rubric (Use the best model available. In testing we have focused on Gemini, Claude, and ChatGPT’s latest models).

If you want to try this with a whole class, we recommend our partner, the not-for-profit app MyEssayFeedback. It has our followup chat scaffolding and our reflection questions built in. The app eases teachers’ administrative labor by giving automatic credit in your learning management system for engaging with the feedback and reflecting.

If you’re interested in PAIRR, please pick one of many ways to connect with us. I also want to give a shoutout to PapyrusAI, a parallel sister project out of University of California, Irvine. They also share tested AI feedback prompts as well as associated instructional materials.

I hope you’ll respond with input, pushback, or alternate AI feedback prompting approaches!

Leave a comment

Upcoming

We’re doing a free webinar introducing the PAIRR method for OneHE on 28 January 2026 at 13:10-13:40 PT. The recording will be shared.

AI Use Statement

I did not use AI to generate words or ideas for this post, with the exception of the title, which was suggested by Claude. I did use ChatGPT, Gemini, and Claude to help revise the PAIRR feedback prompt quoted here.

Student use of AI agents

Imagine a student saying to their browser, “Please log in to Canvas for me and complete all the assignments due today.” Imagine if the browser then moves the cursor around, clicks, generates, and submits discussion posts, quizzes, or pretty much anything else.

That’s possible today.

Perplexity just released the Comet AI agent browser and made it free for students. This makes the functionality widely available, but it was already possible with ChatGPT Agent (at the $20/month level) and with the ManusAI agent (limited availability). Anthropic is testing an agentic Claude for Chrome system with a few premium users. Every major AI company has been talking about and building AI agents for a while now; we’re just seeing them reach the point where they’re ready for broad use.

As far as I’m aware, none of these companies has shared anything they are doing to stop the fraud these systems can facilitate so easily in learning management systems. As points out in his powerful Open Letter to Perplexity AI, Perplexity is even explicitly advertising the ease of cheating.

You can give any of these systems your Canvas login and ask them to complete tasks for you in Canvas as if it were a human user, with or without your supervision. (Side note: when you’ve suffered from clunky Canvas design for years, it’s oddly satisfying watching these agents narrate their attempts to get around a course.)

I recently tested with Perplexity Comet in student mode in a free Canvas account not associated with my college or any student data. It jumped right into completing quizzes.

The first mainstream media coverage I’ve seen of the issue is ’s forceful Forbes article Colleges And Schools Must Block And Ban Agentic AI Browsers Now. Here’s Why. But I’m just one of a host of educators who have been trying to draw attention to this issue. (Let me know about articles or videos I’ve missed, and I’ll add to the list):

• : 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐀𝐈: 𝐀 𝐖𝐚𝐤𝐞-𝐔𝐩 𝐂𝐚𝐥𝐥 𝐟𝐨𝐫 𝐎𝐧𝐥𝐢𝐧𝐞 𝐄𝐝𝐮𝐜𝐚𝐭𝐢𝐨𝐧

• : An “AI Student Agent” Takes an Asynchronous Online Course

• Stavros Hadjisolomou: AI Agents Can Now Navigate and Complete LMS Tasks: A Call for Pedagogical Innovation

• : The Moral Hazards Of AI Are Closer Than You Realize

Note: This is a separate issue from how learning management systems choose to integrate AI. For example, Instructure and OpenAI just announced a partnership that will give instructors various options for incorporating AI into assignments and using it for teaching assistance. But incorporating AI doesn’t mean you have a handle on preventing misuse. A teacher might ask students to chat with a custom bot for an assignment in Canvas. The student might tell an AI agent to go engage in that interaction and “reflect” on it.

Of course, there’s also the question of teacher use of AI agents. Institutions need to determine whether to allow faculty to use AI agents for teaching assistance in the LMS. As points out in The Dangers of using AI to Grade: Nobody Learns, Nobody Gains, working conditions in higher ed make use of AI for grading extremely tempting, and that hold for other forms of automated teaching assistance. But this is a discussion for another piece.

We can’t just give up on online learning

In discussions of student use of AI agents, I’ve heard many educators suggest a return to in-class writing. That’s a reasonable move in some contexts. But we shouldn’t discount how much learning happens today in online spaces or the reasons we put learning activities online.

I’d like to see students encouraged to take in-person classes where possible. But demand for online is high; 57% of all California community college course offerings were fully online in 2023 (Results From a Comprehensive California Community Colleges Online Education Study). Many students taking those online classes can’t take in-person ones. Often, job and family obligations prevent working-class and returning students from regular campus visits. And many students do better learning online, either because of disabilities or preference. There is an art and science to doing online learning well; it can be highly connected and social. See ’s post, ’s post, and ’s post on online learning.

Some increase in in-person instruction and proctoring may be needed going forward. But that will take time and major institutional shifts. We’ll need to weigh it carefully because any requirement for in-person work means some would-be students are excluded.

It would be a shame to abandon online learning because companies decided to release software without guardrails. Before we do that, let’s ask the companies to help.

What can educators do?

Certainly, we should always be trying to make our courses more intrinsically motivating. That will help. argues for course redesign as best response to agents in The Great Online Learning Reset? How agentic AI is forcing us to reimagine online learning -- for the better.

But I don’t think course redesign solves the problem, especially since agentic AI means students can get their assignments completed without ever finding out what the assignments are. If they ask the agent to go do their homework, they might never know if the assignments were motivating or not.

and Anna Haney-Withrow started a group of instructors and instructional designers looking for instructional design approaches that make it harder for AI agents to complete work in LMSes. Tim Mousel put up a form for educators to share ideas on solutions and also makes public the ideas shared so far.

I admire the collaborative search for practical steps. However, it seems to me that many of the strategies may not work consistently or depend on in-person class time or audio and video, which can now be autogenerated as well. I imagine that a student using an AI agent could get it to complete much of my online asynchronous Academic Reading and Writing course even though the course includes lots of process assignments, student choice, connection to personal experience, and social annotation.

It really doesn’t make sense to ask individual teachers to try to solve this unaided.

What can companies do? Surely something, if the will is there

What would it take to block AI agents from completing work in a student’s name in an LMS? After several months of asking around and getting wildly varying responses, I’m still not clear what’s possible or strategic. But I do know that no players have really tried yet. Here are some possibilities:

• AI companies could tell their agents not to complete work while logged into student accounts in Canvas, Moodle, Blackboard, or D2L. That could be hardwired or inserted in the agents’ system prompt. Or companies could label agents so that websites can choose to block them or limit their activity.

• Learning management systems could block agents, either by identifying them in cooperation with AI companies or by developing software to detect their patterns of activity.

• Educational institutions could try to block AI agents on their Wifi networks. Forbes journalist says, “There, unfortunately, is no perfect blocking solution. However, I am cautiously optimistic that LMSes and campuses will take steps to block these browsers to the best of their ability. It’s quite feasible, for example, that a campus IT department could block certain sites from being accessed on the campus WiFi network - so this logic could be applied to known agentic browsers.”

It might not be possible to stop people from running an agent locally on their computer, but that requires some expertise and will hopefully remain rare.

In sum, it doesn’t make sense to give up on technical barriers to fraud when we haven’t begun to try.

Let’s ask ed tech and AI companies to work together to help

We haven’t heard much yet from the big companies on what is possible. It seems more pressure and press coverage are in order. There’s a precedent. Recently, contacted Geoffrey Fowler of The Washington Post and urged him to cover Google’s “Homework help” button in Chrome. Fowler’s reporting may well be what led to Google’s suspension of the feature: Teachers got mad about a cheat button in Chrome. Now Google’s pausing it.

OpenAI and Instructure just announced a partnership to bring AI into Canvas. Couldn’t they collaborate on the agent issue as well?

Tim Mousel writes, “After I posted this video 4 months ago of an agent completing a course in Canvas, several connections sent it to the top LMS companies. They replied back stating they would work on some solutions. They’ve not released one yet as far as I’m aware.” Yun Moh put in a request to Instructure in the Canvas Community Forums to “Block AI Agents from Logging in on Behalf of Students.” The company has marked the current status of this item “Will not consider.” I hope they’ll reconsider.

Ways educators can advocate

Here are a few concrete ways to advocate for the support educators need to keep helping students learn online. Let me know if you have other suggestions!

• Raise the issue with colleagues and your academic integrity office

• Ask your IT department what can be done

• Ask the person who manages your LMS contract to raise the issue

• Post on LinkedIn asking your LMS company and executives to take action.

• Put in a feature request to your LMS company to block agents

• Urge professional organizations to make a statement

• Contact ed tech or tech reporters to cover the issue

AI Use Statement

I did not use AI to generate text for this piece. I did use ChatGPT to generate the two cartoon images, and I reflected on Claude feedback. Claude gave me two ideas that I added to the list of possible actions (see the chat transcript).

Guiding Students on Appropriate AI Use: Collaborative Approaches

Slides: https://link.annarmills.com/guiding

Here's the TLDR:
Writing and reading are more important than ever; they help us

• improve our thinking

• interact with AI

Yes to teaching with writing:

• Frame the value of writing

• Making writing assignments motivating and social

No to AI offloading: set up layers of guardrails to discourage misuse. Consider

• In-class writing

• Video and audio assignments where students talk about their writing

• Process assignmentsProcess tracking (such as with ProcessFeedback.org)

• Detection, used cautiously and not as sole evidence

I draw on or cite among others…

Thanks to Michelle Beissel Heath for inviting me to give a talk for the University of Nebraska at Kearney English Department.

The programs and resources below may be useful either as direct sources of professional development or as models.

Programs (robust, multi-week or multi-module)

• AAC&U Institute on AI, Pedagogy, and the Curriculum: An eight-month, fully online institute where campus teams craft a strategic AI action plan with coaching, peer feedback, and monthly workshops.

• Playlab AI Design Sprint & Critical Creator PLC: Customizable 12-day design sprints and a four-session PLC where educators prototype and showcase their own AI tools; materials are free to remix, or Playlab can facilitate for a fee.

• Relay GSE Playlab "AI for Educators" Micro-Credential in collaboration with PlayLab Three virtual workshops (intro to prompt engineering, collaborative tool-building, showcase) culminating in a micro-credential in AI Tool Design & Prompt Engineering; no coding experience required.

• ISTE "AI Explorations for Educators": 15-hour, self-paced course delivered over an eight-week window with instructor support; scholarships available and completion counts toward ISTE U badges.

• Leon Furze's "Practical AI Strategies": Six self-paced modules that pair ethics with workflow demos (text, image, multimodal) plus optional micro-courses on deepfakes and assessment redesign.

• Auburn University Canvas Course: A self-paced 5-7 hour Canvas course from Auburn's Biggio Center that guides higher-ed instructors through AI basics, assignment design, ethical use, and policy writing; badge on completion. Institutional licenses available.

Free and sliding scale webinars and workshops

• California Virtual Campus Online Network of Educators (CVC @ONE) offers a webinar series each fall and spring and keeps a YouTube playlist of past AI-related webinars

• American Association of Colleges & Universities (AACU) "AI Week" Webinar Series: Four no-cost webinars each spring on leadership strategy, policy, and teaching with AI, featuring national experts.

• CVC @ONE Generative AI-1 Free, 2-week, self-paced course for faculty to draft AI policies and design AI-enhanced assignments; occasional facilitated cohorts add live discussion.

• The ASCCC OER Initiative hosts regular webinars open to all that often focus on AI from specific disciplinary perspectives as well as addressing intellectual property and licensing questions.

• MyFest (Equity Unbound) AI Track: A summer sliding scale "choose-your-own-journey" festival; the AI strand mixes critical literacy conversations with hands-on tool labs, modeling equitable, care-based pedagogy.

• PlayLab: A platform where educators create and share custom chatbots. PlayLab offers community-run workshop series where educators prototype AI tools for their own classrooms, share experiments, and get feedback; sessions range from 90-minute labs to multi-week cohorts.

Adaptable PD materials

• Stanford Teaching Commons Artificial Intelligence Teaching Guide (+ DIY workshop kits) Six modular lessons (AI literacy, course policy, assignment design, etc.) plus ready-made slide decks and facilitator notes you can remix for local workshops.

• TeachAI "AI Guidance for Schools" Presentation — Customisable Google Slides deck that walks faculty or leadership teams through AI basics, benefits/risks, seven guiding principles, and an implementation roadmap.

Podcasts

• Future Trends Forum — Weekly live video-podcast hosted by futurist Bryan Alexander; regular AI episodes with audience Q&A and campus case studies.

• Teaching in Higher Ed – AI Category — Interviews on AI ethics, assessment, and course design, each with transcripts and show notes to spark PD discussions.

• AI in Education Podcast — Weekly chat (Ray Fleming & Dan Bowen) on practical classroom uses of AI, with guest teachers and developers.

• EdSurge — Ed-tech reporters dig into policy shifts, teaching experiments, and AI tools reshaping higher ed.

• Shifting Schools — K-12-oriented show whose "AI in Education" arc gives admin-friendly takes on policy, PD culture shift, and equity. Includes PD guides to accompany episodes and stimulate discussion.

• Hard Fork — NYT journalist Kevin Roose & tech reporter Casey Newton parse the week's AI news with educator-friendly analogies.

• Mystery AI Hype Theater 3000 — Emily Bender & Alex Hanna debunk AI hype and discuss implications for ethics, policy, and pedagogy.

• My Robot Teacher — a playful and entertaining podcast where a humanist (Sarah Senk, PhD) and a mathematician (Taiyo Inoue, PhD) lean toward AI enthusiasm but question many aspects of it as they interview a range of guests.

Listservs and Social Media Resources

• AI in Education Google group

• Higher Ed discussions of AI writing & use Facebook group

• Assignments and Assessments in the age of GenAI Facebook group

• The #CCCAILearn hashtag on LinkedIn. Started by , this hashtag is intended to help educators find useful AI PD posts. Bookmark it or search on it periodically.

• AI and writing, AI in education, AI regulation Part 1 and Part 2 Bluesky starter packs: Two follow-lists from Anna Mills that auto-populate your Bluesky feed with diverse voices discussing AI.

• A few educators worth following for their frequent posts on AI on LinkedIn: Leon Furze, Ethan Mollick, Lance Eaton, Maha Bali, Marc Watkins, Jose Antonio Bowen, Laura Dumin, Tricia Bertram Gallant, Philip Dawson, Kimberly Pace Becker, Annette Vee, Bryan Alexander, Eric Martinsen, Michelle Pacansky-Brock, Fabiola Torres, Dayamudra Dennehy Jon Ippolito Jason Gulya

Websites (reference hubs & toolkits)

• AI Pedagogy Project (metaLAB @ Harvard) — Assignment bank, starter guide, and policy examples for creative / critical engagement with AI.

• Learning With AI Toolkit (Jon Ippolito) — 300-plus crowdsourced strategies plus the IMPACT RISK framework for teaching AI's downsides.

• TextGenEd: Teaching with Text Generation Technologies — An open, ongoing peer-reviewed assignment bank with reflections on tested pedagogical approaches.

• Exploring AI Pedagogy: A Community Collection of Teaching Reflections — A searchable, open collection of short reflections on teaching experiments involving or responding to AI. Hosted by the Modern Language Association.

Why AI feedback?

As I see it, reflecting on AI feedback on your work is a relatively straightforward and low-risk way to use AI. It allows you to get help and to build your prompting skills without missing out on thinking and learning. It’s a powerful alternative to letting AI write or rewrite for you.

By AI feedback, I mean comments that stimulate your thinking as you revise. I don’t mean chatbot rewrites or chatbots feeding you ideas or sources. A tutor shouldn’t tell you what to say, and a chatbot shouldn’t either. But both can support you to figure out what changes you want to make.

Reflecting on AI feedback means you can stay in control, keep your voice, and claim your words and ideas. And it can help you build confidence in your own judgment if you stay skeptical of the suggestions. Some AI advice will be bad, or it will be fine but not for you. You are the one who knows what you want to communicate to your readers. If you develop a habit of questioning AI feedback, you’ll be more likely to question AI in other contexts, which will serve you well in the workplace and in your personal life. People who can see where AI is wrong can improve on it or set it aside when something else is needed.

Another benefit to engaging with AI feedback is that you are unlikely to violate an academic integrity policy. If you use AI for ideas or drafting or rewriting, there will be pitfalls and gray areas around what is okay and what isn’t. If you’re reflecting on AI feedback but making your own changes, you don’t have to worry.

For all these reasons, I have focused on AI feedback since spring 2023 when I started serving as a volunteer advisor on the nonprofit app MyEssayFeedback. For four semesters, I’ve invited students to reflect on AI feedback, and they have overwhelmingly told me they find it useful. Still, I don’t see AI feedback as a replacement for human feedback; I assign peer review and tutor visits too.

In my mind, AI feedback will never be enough because chatbots never experience our writing. They convert our draft into numbers and use a giant chatbot formula to spit out other numbers that convert to the feedback we read. That text might help us think about what we want to say, but we’ve still got to share the draft with humans and see how it lands. Getting a response from a person who’s spent time reading or listening and gotten something out of what we wrote can be validating and energizing in a way chatbot feedback will never be.

Two very easy ways to get AI feedback

1. Sometimes the simplest approach to AI feedback works just fine. When I don’t have the energy to think through exactly what kind of feedback I want, I will upload a draft to a chatbot and literally say “Give me feedback.” The results are often useful and encouraging. If the feedback isn’t quite what I’m looking for, my disappointment often helps me articulate what kind of feedback I do want, and I follow up with a more specific request. (See “Continuing the conversation to get more out of AI feedback.”)

2. The other easy way to get feedback is to copy and paste a feedback prompt into a chatbot along with your draft. I'm going to recommend the prompt I know best, the one I worked on intensively with my colleagues on the Peer & AI Review + Reflection project (PAIRR) in spring 2025. We drafted the prompt to incorporate core principles of writing pedagogy, and we tested it extensively. It tells the chatbot to describe two strengths and two possible areas for revision, quoting from your draft along the way.
   
   The PAIRR reader-response prompt is designed to help you clarify what you want to communicate to readers. It’s supposed to help you find meaning in what you're doing, rather than just telling you if your essay conforms to expectations or predicting what grade you’ll get. We invite you to read and comment on the full prompt.
   
   To try it out, visit the PAIRR Writing Feedback Demonstration bot, hosted on the nonprofit educational platform PlayLab.ai (note that PlayLab makes your chats visible to the PAIRR team).
   
   Or you can copy and paste either of our feedback prompts into any chatbot. See Which AI tools should I consider?.
   
   Follow these steps:

   1. Copy and paste either the PAIRR reader response prompt or the PAIRR criteria-focused prompt into the chatbot.

   2. Copy and paste or upload your draft.

   3. Optional: add your assignment instructions and/or rubric. (Our default prompt includes a basic common rubric for high-priority elements in writing-focused classes.)

   4. Consider the feedback and respond.

Continuing the conversation to get more out of AI feedback

Even after years working with AI, my first impulse is still to take the first result from a chatbot as the best it can do. But a core principle of AI literacy is iteration: pushing the chatbot toward a better result.

Chatting back in response to AI feedback supports AI literacy and your development as a writer and thinker. Whatever you're thinking or feeling as you read the feedback, wherever you're stuck about what to do next, that might be something worth articulating, and the chatbot might give you something useful in response.

Seven strategies for getting more out of AI feedback:

• Be frank and push back. You don't have to sugarcoat it for AI or worry about how it will come across as you might with a human tutor, teacher, or peer. You can be more blunt. Disagree, object, or tell it when you're overwhelmed or frustrated—it's supposed to respond in a helpful and empathetic way.

• Ask for clarification. When anything isn't clear, ask the chatbot to explain. You can also ask for more depth, detail, or quotes from your essay to support its suggestions, or request examples of how to make the kind of revision it suggests.

• Ask for a new version of the feedback. If you’re not satisfied, you can simply ask it to try again or give more insightful feedback. You can also request a different style of feedback based on your mood or preference, ask for the perspectives of a famous person you admire or a writer you've just been reading for class, ask how people with a particular life experience might respond, or ask for feedback that contradicts what it just said.

• Ask about something it didn't address yet. If the chatbot missed something in the assignment or rubric, something your instructor has asked you to work on, or something you’re curious about, bring it up.

• Get it to help you explore what you’re not sure about. Describe your uncertainties about what to do, and ask it to help you think through what's murky in your mind. You can also bring up something you’re thinking of adding to the draft and ask for suggestions about how it could fit in.

• Get it to suggest writing strategies. Describe what you're struggling with and ask what writing process strategies you should consider.

• Play with the bot. Sometimes we freeze up because writing is hard or because we're nervous about sounding academic. Having a playful discussion of your draft might help you relax and explore and bring more of your authentic voice to your writing. Try asking it how you could work a random topic like whales into your draft. Or ask for feedback in the style of a pop culture character. If the chatbot answers silly questions seriously or gives you bizarrely contorted feedback, that’s a good visceral reminder that these are prediction engines spinning text, not wise mentors—an AI literacy win.

Sample responses to AI feedback

Here are a few examples of the strategies described above.

Be frank and push back

• "That's seems condescending."

• "That seems biased against..."

• "Come on, didn't I already do that in this draft? I swear I did."

• "I'm not convinced I really need to make the thesis more specific."

• "I just don't think I can do that."

• "It just seems boring to add more examples as you suggest."

Ask for clarification

• "What do you mean by a more focused thesis?"

• "What's an example of splitting up a paragraph that’s too long? How do you know where to split it?"

• "Can you share more quotes from the essay to show where I'm vague?"

Ask for a new version of the feedback

• "Try again."

• "Give me more insightful feedback."

• "Loosen up and explain it again."

• "It seems like you're dancing around something. Give it to me straight."

• “Be more supportive and help me see my strengths, but stick to the truth.”

• "Can you explain it in a casual way like a friend would?"

• "What might a soulful poet / brainy comedian / wise leader / social justice activist / scholar in the field?” say?"

• "What kind of feedback might I get from a reader who has experienced homelessness?"

• "Now give me a reasonable perspective that goes against what you just said."

Ask about something it didn't address yet

• "What about my introduction?"

• “I want to improve my sentence style.”

• “My teacher suggested I work on integrating quotations.”

• "How could my paper be more interesting?"

Get it to help you explore your uncertainties

• “Ask me questions about what I want to do with the conclusion.”

• "I don’t like the idea of cutting out the descriptions of all the kinds of solar energy although I agree that the reader might get lost in them. Can you help me explore why I don’t want to cut them?"

• "That makes me realize I want to talk about the example of TikTok almost shutting down. Is there somewhere that could fit naturally?"

Get it to suggest writing strategies

• "I think my essay is all over the place, but outlines don’t work for me. How else can I work on organizing it?"

• "I'm feeling overwhelmed. What are some strategies that could help?"

• "I can’t think. How do I tell if I should take a break or keep pushing?"

Play with the bot

• "Okay, now tell me what possible connection my essay could have to popcorn / octopi / Batman / breakdancing. I want to put that in my conclusion."

• "What feedback might I get from Bart Simpson / Lisa Simpson / Spiderman / Ironman / Forest Gump / Grumpy Cat / Darth Vader / Aladdin's Genie" (Please insert your own favority character or comedian here–I’m not with it enough to guess what might appeal to you!)

Take a step back, reflect, and talk to a human

To get the most out of AI feedback, at some point we have to set it aside and use other strategies. Make sure you give yourself space to decide what you really think about AI feedback and what, if anything, you want to change in your writing as a result. Some teachers might ask you to write about your thoughts and plans after reading AI feedback; this is part of the Peer & AI Review + Reflection approach. If you’re asked to do this, see Writing a reflection about AI feedback for possible topics.

Another way to get perspective on AI feedback is to seek out human readers. Sometimes I know just what I want to do with AI feedback, and I get right to it. Other times, I need to mull it over for a day. Still other times, the AI feedback makes me realize I need human input. If you’re not sure if you agree with the AI feedback or you want more trustworthy advice, it could be a great moment to reach out to your teacher, a tutor, a classmate, friend, or family member.

Human feedback can also counteract the chatbot tendency to tell us what we want to hear. This is often called “sycophancy”; we could also call it flattery or “kissing up,” though chatbots aren’t hoping to get something from us. They’re just designed to produce things that humans tend to like. Chatbots can make us question our ideas if we ask them to, but more often, they tend to reinforce our inclinations and biases.

I ran into this problem recently when I asked Claude for feedback on a description of an upcoming workshop for teachers. I enjoyed the cycle of feedback and revision so much; Claude gave suggestions and kept telling me that each draft was a big improvement. I felt great when it celebrated the changes I was making because I too thought I was on the right track. The result was a workshop description that listed every point I was going to make and every activity I had planned. When I sent the description to the workshop organizer, she wasn’t impressed and gently suggested I cut out the overwhelming detail. At first, I was indignant and felt unappreciated. After a few hours I had cooled down and decided she was right. Claude had helped me talk myself into agreeing with myself, but the human workshop organizer was right that human teachers would feel overwhelmed on reading all the specifics of what I meant to do.

A next-level way to engage with AI feedback: Writing or adapting your own prompt

I’ve focused on the easiest ways to request feedback and then emphasized expressing your specific needs in the followup chat above. But once you’re familiar and comfortable with chatbot feedback, you’ll likely want to be more specific from the beginning. Writing your own initial feedback request or customizing someone else’s allows you to reflect on what kind of help works best for you in your writing process. It’s also a good way to practice prompting.

Here are some things you could include in a feedback prompt, though none of these are required:

• What’s important to you right now: what do you want to know or work on?

• What instructions has your teacher given you?

• What role do you want the chatbot to play? Examples might include supportive and understanding writing tutor or cheerleader, a rigorous critic, or a skeptical reader.

Here are a few sources of sample AI feedback prompts you can adapt or just browse for ideas:

• The PAIRR prompt mentioned above.

• The MyEssayFeedback library of feedback prompts teachers have written and shared (you can filter by whatever focus you want or sort by most popular)

• The PapyrusAI prompt library has prompts tested by faculty at the University of California, Irvine, each with a different writing process focus

• Professor ’s CoachTutor prompt is playful and funny.

• The University of Sydney’s instructions for Getting standards-based feedback focus on meeting the assignment criteria.

If you do come up with a custom feedback prompt that works well for you, consider sharing it! Here’s a site where students can post prompts.

Further reading

• Getting the most out of AI (prompting): Describe what you want and keep trying

• Bad Idea: Now that We Have AI, We Can Ditch Human Feedback! Generative Idea: Combining Peer and AI Feedback by Lisa Sperber

• AI as Feedback Generator by Ethan Mollick and Lilach Mollick

• I, Cyborg: Using Co-Intelligence: How I used AI in my book about AI by Ethan Mollick

This article is shared under a CC BY NC 4.0 license.

I argued against use of AI detection in college classrooms for two years, but my perspective has shifted. I ran into the limits of my current approaches last semester, when a first-year writing student persisted in submitting work that was clearly not his own, presenting document history that showed him typing the work (maybe he. typed it and maybe he used an autotyper). He only admitted to the AI use and apologized for wasting my time when he realized that I was not going to give him credit and that if he initiated an appeals process, the college would run his writing through detection software. I liked this student, had met with him previously and encouraged him to build confidence in his own voice and helped him find a research topic that interested him. I didn't think he would be well served by an A on an autopaper, so I was glad that AI detection existed.

I'm also influenced by recent research that suggests detection is likely not biased against English language learners after all, that educators are not as good as we think we are at distinguishing AI from student writing on our own, and that some detection systems are pretty accurate when it comes to naive copy/paste AI use. Christopher Ostro's slide deck lit review of recent research on detection has been invaluable. He discusses this research in an engaging episode of Bonni Stachowiak's wonderful podcast Teaching in Higher Ed. Dr. Tricia Bertram Gallant's leadership on academic integrity and AI has influenced me over the last two years as well. She has shown levelheaded willingness to consider a possible role for detection even as she promotes other approaches as more important and effective.

As I teach composition online asychronously this semester, I'm incorporating Turnitin alongside process tracking, writing process assignments, social annotation, lots of student choice, peer review and tutoring, video assignments, and clear messages about the purpose of each activity and the value of the writing process. I love that Phillip Dawson has described this kind of laying of strategies as a "swiss cheese" approach and others have used the mosaic metaphor (I'm having trouble tracking down who). I've described my combination of approaches in slides for a recent presentation on Academic Integrity and AI.

What about the risk that AI detection will lead to false accusations? It's real, and I let students know I'm aware that detection yields some false positives. I will never trust AI detection as firm evidence, and I am not punishing students. If a student discusses an essay with me, shows process history, and denies AI use, I will give them the benefit of the doubt even if the detector says "AI." If they aren't able to discuss their writing, I ask students to rewrite.

Christopher Ostro put it in a way that resonates for me: "I think AI detection has a place, but its place is limited." In most human endeavors, some accountability structures are important even when we design for intrinsic motivation. And we don't have perfect options here. The options are fewer in online asynchronous classes that allow many working-class students and parents to access college.

I know so many colleagues I respect are highly critical of AI detection, seeing it as signaling antagonism toward students. Christopher Ostro clarifies that his purpose is not to punish students but to provide some accountability that, in the end, encourages learning and shows that we care. He says, "I am not a cop teacher. I am not someone who likes catching cheaters. I’m not someone who wants that to be a big part of my job. Honestly, it’s the least fun part of teaching, but it’s also it is still a part of the job."

I'm a member of the MLA/CCCC Joint Task Force on Writing and AI, a group that has put out strong cautions about it in our working paper on Generative AI and Policy Development. There, we argue that "Tools for detection and authorship verification in GAI use should be used with caution and discernment or not at all." We write, "For those who decide to use AI detectors, please consider the following questions: What steps have you taken to substantiate a positive detection? What other kinds of engagement with the student’s writing affirms your decision to assign a failing grade outside the AI detector’s claim that the text was AI generated?" We also emphasize that "any technological approaches to academic integrity should respect legal, privacy, nondiscrimination, and data rights of students."

I have tried to use detection and process tracking in ways that I hope address those concerns. I invite students to comment frankly on syllabus policies. If students don't want to share process history or if they object to AI detection, I invite them to meet with me to chat about the essay instead. My approach is a work in progress in a changing landscape. Next up: an anonymous survey to see what more I can find out about what students think.

Here’s a little sample report (on a frivolous document) where the first and last sentences were typed and edited with Grammarly, and the middle paragraphs were AI-generated. The screenshot below shows the overview and the beginning of the marked-up text.

I’m starting to ask students to share the link to their Grammarly Authorship report under their name at the top of any Google Doc they turn in. However, I acknowledge that some may prefer alternate writing processes or apps, and some may prefer to arrange another way to share information about their process, such as a conference. I invite students to check in with me if that’s the case.

Here are the steps to try it out: 

1. First, make sure you have the Chrome browser installed.

2. Make sure you have the Grammarly extension for Google Chrome turned on. Go to the puzzle piece icon in the Chrome toolbar to see extension options

   

3. Consider whether you want to ask Grammarly not to use your data to train AI language models (like the ones used by Grammarly and ChatGPT).

4. Open a new Google Doc.

5. Locate the thumbprint logo in the bottom left corner of the screen and click it.

   

   (If you don't see the thumbprint, try refreshing. If that doesn’t work, open the Grammarly sidebar by clicking the Grammarly button on the lower right and scroll down to the option to turn on Grammarly Authorship.)

6. Select "Track Writing Activity"

   

7. You may want to click the settings icon and turn on the option to start tracking writing activity automatically every time you create a new document.

   

8. Work on the document.

9. Click the thumbprint again and select "See report."

   

10. Read through the report and reflect on what it might reveal about your writing process, use of copied and pasted text, and use of AI. If you see anything odd, you may want to reach out to Grammarly support. This is a beta version of the software, and there may be glitches.

11. Select "Share" at the top right.

    

12. Make the report visible to anyone with the link.

    

13. De-select “Include Writing Replay.” (I don’t require a full movie version of your writing process because that doesn’t seem necessary and might make some student writers more self-conscious about rough drafts.)

14. Copy the link and share it.

Is anyone else using Grammarly Authorship with students? I’d love to know how you are presenting it and how it’s going, if so.

This document is shared under a CC BY NC 4.0 license.

• Accountability that's less stressful for students

• Is tracking the writing process invasive?

• Pros and cons of Grammarly Authorship

• Other writing process tracking options

• Encouraging student reflection on the writing process

Later this month Grammarly is launching Authorship for all users (see this Inc.com article and this BusinessWire article). I tried out the pre-launch version and wanted to share some thoughts. Basically, Grammarly Authorship allows anyone using the Grammarly extension for Chrome in Google Docs to generate a report on their writing process. The report will show how much time they spent on the document, show a sped-up replay of the document’s creation, and attempt to label each chunk of text as human-authored, AI-generated, “sourced,” or “unverified.” As I understand it, the system doesn’t use statistical AI text detection; it simply labels text “AI generated” if it was copied and pasted from ChatGPT. Any text that was typed is labeled “human authored.”

Now I am not getting any compensation from Grammarly, I have no ongoing relationship with them, and I don’t love everything they do. To be honest, I haven’t looked much at how they handle style, and I’m frustrated with the confusion that has come about since they introduced generative AI; people say “I used Grammarly” and that is understood to mean they got minor grammar and word choice suggestions. But even the free version of Grammarly gives 100 generative AI prompts a month. Users can just ask it to autogenerate whole passages. For my class, I ask students not to use Grammarly or any chatbot to write or rewrite their essays (though we do read about AI and they reflect on AI feedback on their writing). However, I still encourage my students to use the free version of Grammarly, review its grammar suggestions skeptically, and follow its links to explanations of possible errors. (This paragraph was updated after Carol Bailey pointed out that the free version of Grammarly offers generative AI, not just the Pro version.)

Like the free, relatively high quality grammar suggestions, I think Grammarly Authorship is going to be an overall boon to students and teachers. I hope Grammarly will continue to improve on various points, but this is an important step in the direction of transparency about what is AI text and what is not.

Accountability that may be less stressful for students

To me, process tracking seems like a more fair and transparent approach to accountability than the statistical AI detection commonly used by teachers today such as Turnitin’s AI detection and GPTZero.

No, it is not foolproof at preventing students from using AI to skip valuable writing and thinking practice. It is possible to retype AI output; that retyped output will be labeled as human by the Grammarly Authorship report. Still, the amount of time spent in the document will be recorded. It is less tempting to retype AI output in about as much time as one would have spent writing one’s own words. 

Of course, this does nothing to address the root causes of AI misuse. The most important things we can do as teachers are still 1) to design writing assignments that are intrinsically motivating, 2) to cultivate relationships with students and between students that support student confidence in their own voice and ideas and 3) to teach and support the writing process itself to make it less stressful and more achievable step by step. It’s also helpful now to mix in low-stakes in-class writing and oral assessment. (See these slides on academic integrity for more on my approach.)

But as I and many other teachers have reported on social media and in professional listservs and Facebook groups over the last year, these strategies are not enough. We need other tools to support accountability and make last-minute recurrence to AI less tempting for students who are stressed, overwhelmed, and intimidated by academic writing. It’s probably because so many teachers believe there is a need for some accountability that Turnitin’s AI detection has remained popular. We’ve seen plenty of evidence that it produces false positives that lead to false accusations. Serious concerns have surfaced about bias against English language learners and studies have shown that it’s easy to evade detectors through automated paraphrasing software (Turnitin claims to have solved this but hasn’t offered evidence). Most teachers don’t want to police. But we also recognize that some accountability is needed in most human endeavors that involve sometimes disagreeable effort. 

Is tracking the writing process invasive? 

I don’t love asking students to share their whole document history. I do think the practice is justified, but I think we should allow for alternate processes (We should, of course, also allow accommodations for disabilities). Here’s how I frame it to students: “Think of this as an online version of in-class writing that allows you more flexibility. It gives me a way to understand your writing process, both to encourage academic honesty and to encourage reflection on the writing process itself. If you have concerns or do not feel comfortable sharing your process in this way, please let me know and we can meet and work out another plan.” 

This informal poll I ran on X/Twitter in January 2023 suggested that opinion about process tracking is divided. 58% of 248 respondents thought it was fine and 42% thought it was intrusive or inappropriate.

Let’s be clear what we’re comparing it to; if the alternative is in-class writing, that is also under close observation. I did another informal poll where I said “You are a student. Your instructor gives you a choice: Would you rather write your essay in class or write it out of class with full keystroke-by-keystroke transparency about your activity in the document?” There was a near-even split among 287 respondents on preference. Of course, these polls don’t represent a random sample but only the particular audience that comes across my X posts.

Pros and cons of Grammarly Authorship

I appreciate that this is available free to all Grammarly users. This means if we were already using Grammarly extension, we won’t be giving away data any more than we were already (See the Grammarly Privacy Policy). Like OpenAI, Grammarly defaults to using user data to train its models, but this option can be turned off. 

I like that in this model, the process report belongs to the author. The author can review it, choose to print it to PDF and share it with their comments. Students could do this to defend themselves against a charge of AI misuse. Grammarly Authorship doesn’t sync with learning management systems or provide ways to manage a whole class set of process reports. Still, teachers could require submission of PDF authorship reports along with essays. 

There are some ways that Grammarly Authorship as it currently works could lead to unjustified teacher suspicions. There are a few ways their reports will likely mislabel text. For now, voice typing shows up as “unverified” rather than “human authored” which might cast suspicion on students who routinely use voice recognition to write. The “AI generated” label suggests that the system can actually identify what is AI generated, but that might instill some false confidence. The label seems only to be applied when text is copied directly from ChatGPT. If copied from Claude or another AI app, it gets labeled “unverified.”  Also, many people use ChatGPT to suggest grammar or style revisions to text they have come up with themselves. Then they copy the resulting edited version back into their document. I tried that, and the Grammarly Authorship report identified my edited text as “AI generated.” Potentially a student could share a transcript of their chat session to defend themself in such a case, but they would still undergo the stress of being accused. I’d like to see text labeled more precisely as “Pasted from ChatGPT” or “Pasted from an AI app” rather than “AI generated.” 

I am hoping that the final version will distinguish between text edited by Grammarly and text wholly generated by Grammarly and acknowledge that the generated text comes from AI. So far in my experiments, anything edited or generated by Grammarly is labeled as a known source, but it doesn’t identify this source as Grammarly or label the generated text as “AI generated.” I hope they’ll change that! In my one meeting with Grammarly product marketing manager Cliff Archey, he seemed receptive to input.

Other writing process tracking options

Grammarly Authorship is one new option among many for writing process tracking. I have focused on it here because it seems likely to gain wide adoption since it is free and so many already use Grammarly. 

However, one I’d like to spend more time looking into is Integrito by PlagiarismCheck.org. I’m not clear on how the pricing works; there may be free options. Chrome extensions like Revision History and Draftback are definitely free, but these require students to share edit access to their Google Doc.

If the colleges where I teach could find funding, I would be glad to use

Antecedent Writing, PowerNotes Insight, or Rumi instead. They seem like the best designed to allow teachers to more easily track a whole class worth of process reporting. They are FERPA compliant and don’t include statistical AI detection. (Txtreplay might be another option.)

I’m going to participate in Turnitin’s beta testing of its own writing process tracking system this fall. Turnitin’s classroom management or learning management system integration would make it administratively easier to review student process, but I’m concerned they may not default to giving students first view of and control over their writing process reports.

Encouraging student reflection on the writing process 

In addition to the pedagogical approaches mentioned above, I think we will continue to need software that makes AI misuse less tempting and makes it easier for students to do their own writing outside of class. Process tracking seems to me like the best path forward for this, but process tracking also has other uses that haven’t been explored.  What I want to see ultimately is an app that invites students to reflect back on their process as they comment on it. The teacher could then respond to the comments in ways that support the student’s future writing strategies.

I’m going to try out using Grammarly’s Authorship report in this way by asking students to print it to PDF, upload it to Google Drive, annotate it there, and share it with commenting privileges. Here’s a sample where I’ve commented on my own Authorship report to highlight some of the pitfalls I mentioned above.

I’m looking forward to process tracking software that helps the student reflect on their process of authoring the document and also helps student and teacher distinguish between the student’s choices and any AI influence. That may be the ed tech we need to support writing for thinking and learning in an age of AI.

AI use statement

I reflected on Claude 3.5 feedback on drafts of this post but did not include any AI-generated text.

Your thoughts?

I know this is a tricky issue, and I hope to see much more discussion of process tracking as educators try to figure out whether it’s needed, how it compares to alternative approaches to academic integrity, and if it’s needed, how it should be done. Please feel free to comment here!

Update note September 4, 2024: I changed the references to Grammarly Pro because Carol Bailey pointed out on X that even the free versions of Grammarly include text generation and revision, though they limit users to 100 prompts per month. Thanks, Carol!

Share

Leave a comment

Subscribe now

This is an adaptation by Anna Mills of an article by Liza Long, originally part of the Writing and Artificial Intelligence section of the textbook Write What Matters, licensed CC BY NC 4.0..

Background and request for comments

This is a draft document addressed to students that will eventually become part of the open textbook How Arguments Work: A Guide to Writing and Analyzing Texts in College. A grant from the California Community Colleges Chancellor’s Office Zero Textbook Cost Program via Cañada College funds this work in Summer 2024 under the guidance of Sarah Harmon. 

You can find this same document on LibreTexts or in Google Docs if you prefer to read and comment there.

I would appreciate feedback from both teachers and students on this draft! (I’m inspired by the open peer review process for Remi Kalir’s book Annotation, as described by Troy Hicks.) Please join in by leaving public comments or by emailing me at info@howargumentswork.org to suggest a way you'd like to contribute. 

Overview: A Checklist for Acknowledging and Citing Generative A.I. Tools

You've probably heard about the importance of citing your sources in academic work. But in cases where you are asked or allowed to use an AI chatbot such as ChatGPT for some purpose in the writing process, how do you acknowledge and cite it? These five steps can be helpful if you are considering whether or not to use and cite generative artificial intelligence in your academic work:

1. Check with your instructor to make sure you have permission to use these tools and to find out what the limits are on these uses.

2. Reflect on how and why you want to use generative artificial intelligence in your work. What skills are you supposed to practice through this assignment? Will using generative artificial intelligence in the way you are planning help you build those skills, or will it get in way of practice that you need?

3. Check the accuracy of any information provided by a generative artificial intelligence tool against a trusted source. Can you verify it? Is anything made up? Would it suit your purpose to cite the trusted source instead of the AI system?

4. Check if the AI output is based on a source that should be cited. Chatbots sometimes give answers that reflect a particular source’s text or ideas without citing that source. So humans need to do searches on the topic to see not just whether the output is accurate but whether it should be attributed.

5. If you decide to use generative artificial intelligence, acknowledge your use, either in an end note, an appendix, or a cover letter (see below for examples).

6. Cite any ideas or word sequences that come from generative artificial intelligence both by mentioning the source in the body of the essay and by citing it on a References/Works Cited page according to the style format your teacher specifies, such as MLA or APA.

Whew. Let’s just admit that’s a long list of steps to take before incorporating generative AI in your essay. It’s worth pausing to ask if it’s worth it. As teachers and students alike are discovering, generative AI is not always a time saver especially if we’re trying to use it in an intelligent, strategic, and ethical way. 

It’s not always easy to check the accuracy of outputs, and it can be even harder to figure out whether the text or ideas that a chatbot spits out really come from a specific source that needs to be cited. 

If these steps seem too difficult, don’t be shy about turning away from generative AI if another kind of source will suit your purpose. If you do need to use AI, consider asking your teacher, a tutor, or a librarian for help checking, acknowledging, and citing.   

Acknowledging Use of AI Chatbots

An acknowledgment statement is meant to give detail about how AI was used so that readers can understand how it may have influenced the text and the author's thinking.

Monash University provides helpful recommendations for how to acknowledge when and how you've used generated material as part of an assignment or project. They suggest that acknowledgment statements include the following:

• Specify which technology was used.

• Include explicit descriptions of how the information was generated.

• Identify the prompts used.

• Explain how the output was used in your work.

The format Monash University provides is also helpful. Students may include this information either in a cover letter or in an appendix to the submitted work. Academic style guides such as APA already include guidelines for including appendices after essays and reports. Review the Purdue OWL's entry on Footnotes and Appendices for help.

Sample Acknowledgment Template

I acknowledge the use of [insert AI system(s) and link] to [specific use of generative artificial intelligence]. The prompts used include [list of prompts]. The output from these prompts was used to [explain use].

Using chat session transcripts to document AI use

It's a great help to have a full transcript of our interaction with a chatbot when we're trying to acknowledge what we did. ChatGPT and Gemini allow users to share links to specific chats (see the directions for sharing ChatGPT chats and directions for sharing Gemini chats). Here's an example of a shared chat link in ChatGPT. When you click on the link, you'll be able to see both the author's prompts and the chatbot's responses. This give you automatic documentation and details to include in your acknowledgment.

Claude does not currently have this built-in functionality, though AIArchives.org offers a browser extension that will create links to chat transcripts of sessions with Claude. For other tools like the Google Docs Writing Assistant, links are not yet available. One option in that case is to take screenshots of the chat session and share those in an appendix.

Citing AI Chatbots

Academic essays require us to tell readers where we got ideas and quotes, both in the body of the essay and in a list at the end. That way readers understand where the ideas and words came from and can look up the sources and judge for themselves.

Now let's be clear: the chatbot output is not a "source" in the usual sense. Sources for academic essays are generally expected to have some person or group of people who wrote them and who stand behind them. Chatbots can't stand behind their output. If you put the same prompt into the chatbot a second time, you will most likely get a different result. So one could argue that we can't really "cite" a chatbot in the usual sense of academic citation, which assumes that the words were put together with intention and understanding.

Yet even if chatbots aren't traditional sources, it's still important to cite them if we reference text or ideas from them. Readers still need information about where the text and ideas are coming from. The American Psychological Association (APA) and the Modern Language Association (MLA), two of the most frequently used style guides for college writing, have both provided guidelines for how to do this.

Note: Even though APA and MLA do not require it, many consider it a best practice to include links to chat session transcripts in your citation as described above.

APA Style

According to the American Psychological Association (APA) as of February 23, 2024, AI chatbots should be cited like this:

Sample APA Citation

In the body of the essay:

When prompted for a list of "public-service jobs for men," ChatGPT listed two stereotypically masculine jobs first: "Firefighter" and "Police Officer." (OpenAI, 2024).

In the reference list:

OpenAI. (2024). ChatGPT (Mar 14 version) [Large language model].

https://chat.openai.com/chat

The APA blog post by Timothy McAdoo acknowledges that this format may evolve, and indeed the comments on the post include many suggestions and requests. Readers note that the version history date is no longer listed at the bottom of ChatGPT sessions as of Spring 2024. Some are concerned that mentioning the company in the parenthetical citation suggests that the company is the source of the idea or text when in fact the chatbot's output is shaped by all the text from the Internet that it was trained on. Others have suggested that the link to the chat session transcript should be included instead of the general link to the chatbot. You may want to check with your teacher for more guidance.

Here is one possible alternative format that follows APA style with adjustments to provide more information.

Alternate-Style Reference List Citation

OpenAI. (Prompted July 3, 2024). ChatGPT 4o [Large language model]. https://chatgpt.com/share/78c36969-aba3-4803-8bdb-0085332182e8

MLA Style

The Modern Language Association (MLA) provides the following guidelines:

• Cite a generative AI tool whenever you paraphrase, quote, or incorporate into your own work any content (whether text, image, data, or other) that was created by it.

• Acknowledge all functional uses of the tool (like editing your prose or translating words) in a note, your text, or another suitable location.

• Take care to vet the secondary sources it cites. (MLA)

Below are some examples of how to use and cite generative AI with MLA style. For more details, see How do I cite generative AI in MLA style?, MLA In-Text Citations, and MLA Works Cited Pages.

The MLA invites comments about its guidance and even says "the MLA template of core elements is meant to provide flexibility in citation. So if you find a rationale to modify these recommendations in your own citations, we encourage you to do so." I have modified the Works Cited entry by listing the link to the chat session transcript rather than the general link to the tool and by adding a reference to the specific model used within ChatGPT (in this case, GPT-4o) to provide more detail.

Sample MLA Citation

In-text citation of quoted text where the prompt is described in the sentence:

When prompted to list public-service jobs for men, ChatGPT listed two stereotypically masculine jobs first: "Firefighter" and "Police Officer." (2024).

In-text citation of paraphrased text where the prompt is referenced in parentheses:

ChatGPT has been known to output stereotypical answers when asked to list jobs for a certain gender. ("Public-service jobs for men" 2024).

In the Works Cited list:

"Give me a list of ten public-service jobs for men" prompt. ChatGPT 4o, version unknown, OpenAI, 3 Jul. 2024, https://chatgpt.com/share/78c36969-aba3-4803-8bdb-0085332182e8.

Further Resources

The Research Process chapter in How Arguments Work: A Guide to Writing and Analyzing Texts in College

References

McAdoo, T. (2023, April 7). How to cite ChatGPT. APA Style Blog. https://apastyle.apa.org/blog/how-to-cite-chatgpt

Modern Language Association. (2023, March 17). How do I cite generative AI in MLA style? https://style.mla.org/citing-generative-ai/

Monash University. (n.d.). Acknowledging the use of generative artificial intelligence. https://www.monash.edu/learnhq/build-digital-capabilities/create-online/acknowledging-the-use-of-generative-artificial-intelligence

Attribution: This article is an adaptation by Anna Mills of an article byLiza Long, originally part of theWriting and Artificial Intelligence section of the textbookWrite What Matters, licensed CC BY NC 4.0.

First, let me clarify that this is not a grading app. I am still responding to and evaluating my students’ writing. Students still do peer review, visit human tutors, and get plenty of comments from me. AI is a secondary supplement. The idea is that AI feedback may help students clarify what they want to communicate to human audiences.

As I see it, MyEssayFeedback is aimed at using AI to

• Stimulate thinking and writing rather than replacing them

• Build critical AI literacy

• Support students’ confidence in their own judgment.

Unlike other feedback apps, MyEssayFeedback allows teachers to choose, adapt, or write instructions for exactly what kind of feedback the system should give on a student draft. With the open library of feedback prompts, teachers can build on what others have done and adapt prompts to their own philosophy and assignment specifics.

MyEssayFeedback allows students to use AI with guardrails and instructor guidance. Students can ask teachers’ opinions on the AI feedback, and teachers can weigh in anytime we see something we don’t agree with. We can assign reflections or revision plans in response to the feedback and automatically give students credit for their engagement.

MyEssayFeedback is simple for teachers and students to use and syncs easily to common LMSes.

It's healthy to be suspicious of ed tech sales pitches. This is not a project aimed at making a lot of money; we are committed to keeping it very low cost. Thanks to a grant of $25,000 Azure credits from Microsoft for Startups Founders Hub, new instructors will be able to test out MyEssayFeedback.ai for free through at least Fall 2024 (limit 3 teachers per school). We are already FERPA compliant and have verified accessibility. Future essay analysis will leverage Microsoft's GPT4 model, backed by a stringent privacy policy.

We are not promoting AI as a bigger deal for writers than the typewriter. But we do think AI writing feedback can be useful when approached skeptically. Students will be using AI in the workplace, and an educational app with guardrails can help them practice recognizing places where AI misses something as well as where it’s useful. (See these suggested phrases for reflecting on AI feedback)

In addition, by showing students an appropriate way that AI can support the writing process, we hope to discourage overreliance and academic dishonesty.

We would like to see more research into feedback quality and student perceptions of AI feedback. So far, student comments have been largely positive. Here are a few, shared with permission:

"By only editing previously written content instead of generating new content, MEF eliminates the fears of plagiarism that usually accompany the use of AI systems, while also promoting the benefits of targeted essay feedback."

-Andrew

"I think there was a nice balance when MEF gave students help because it was not just giving answers, it was providing help that engages and provokes stronger writing."

-Grace

"I really appreciated being able to have my work reviewed before I showed it to another person, as that alleviated some of my stress before there was another pair of eyes on it. This tool was effective in multiple ways, including strengthening my writing and teaching me how to interact with A.I. in an efficient and advantageous manner. My favorite part of using the A.I. was the ability to ask a follow-up question, and for the most part, receive an intelligent and beneficial answer."

-Peyton

"MEF did not rewrite anything for me, so while it gave suggestions on things to improve and how to do so, the work was still mine. MEF taught me how to use A.I. in a way that doesn't allow me to cheat or have something else do my homework for me, yet it truly improved my writing."

-Michael

Thanks for reading all this! If you're curious about integrating AI into writing assignments in a responsible way, please feel free to reach out to us. You can also watch our recent webinar or browse the slides.

Let me come right out and say that TextGenEd: Teaching with Text Generation Technologies has shaped my sense of what’s possible and what I want to try when it comes to AI and the teaching of writing. This is above all a useful book, one I have returned to over and over since its publication and referred to in myriad presentations on AI in Fall 2023. It is useful first because it describes many compelling teaching practices. I wish I could try most of the assignments described! Even though the reflections are about classroom activities that took place in Fall 2022 into early 2023, they have not lost relevance. Adapting them to the updated language models of late 2023 is for the most part trivial.

Second, the collection is useful because it frames discussion of the place of AI in the higher ed classroom. It models both critique and curiosity around AI as well as supplying historical and technical context.  It invites us to our own explorations without pressuring us to be for or against AI or to resolve all the uncertainties about the future of writing in a time of rapid technological change.

The Introduction

I have listened to synthesized voices reading the collection’s  introduction to me at least three times now in addition to reading and annotating it. It stands on its own as an introduction to AI and writing in higher education. What's really striking is the way it bridges divides in the discourse. Over the last year as higher education has grappled with AI since the release of ChatGPT, there has been a lot of pressure to either “embrace” or resist student use of AI text generators or large language models (LLMs). We might expect a collection titled “Teaching with Text Generation Technologies” to fall in the “embrace” camp, but in fact the collection isn’t promoting the technology or recommending indiscriminate use. Rather, it promotes a hybrid response of critical engagement. The authors “believe that these tools are likely to be adopted rapidly in certain sectors of the writing economy in the coming months and years, and fostering student understanding of them is important.”  They bring home this sense of a coming sea change when they describe the prototype of text generation in Google Docs where the user “ clicks a "Help me write" button and has the tone of their paragraph changed.” “To the first generation of AI users, it might feel like magic. To the second generation, it might feel ordinary.” According to them, the kind of “instructional experimentation” showcased in the collection “will collectively put us in a much better position to determine, to the extent that we are able, how these tools should be adopted, and how we might resist them when necessary.”

Many teachers have not yet learned too much about text generation AI, and the editors hope the collection “offers something for teachers with all levels of comfort with technologies.” The introduction provides just enough historical context and technical information to jumpstart explorations.  The editors give a brief, readable history of automated writing that belies the current sense that ChatGPT is sui generis. Without getting sidetracked by detail, they explain how the advances that preceded ChatGPT’s release made automated writing into a new realm of sophistication. 

If teachers incorporate text-generation technologies into class activities, which ones should we use? How should we vet these systems around student privacy and use of student data?  The authors draw attention to open-source models like those from Hugging Face as the only option to fully protect student privacy. However, they suggest that using systems like ChatGPT while mitigating their harms is also a defensible approach. In the near future, they suggest, corporate models with privacy protections may well be available through educational institutions. In the meantime, they do “recommend that instructors provide alternate assignments if a student objects to using a commercial application.”

If we are not wholly embracing AI and not banning it, how do we determine how much and in what ways it should be used in pedagogy? The authors acknowledge that some limits will likely be needed, asking, “[w]hich parts of the writing process can we cede to AI while retaining what we value about writing? We will soon learn if it is tenable to allow students to use AI for some parts of the writing process (e.g., brainstorming and grammar/style checkers) but not for others (e.g., text generation). We may want to embed constraints in our assignments so as not to offload too much of students’ cognitive work to AI.” Yet will we be able to maintain such constraints? And even if we can, will the tension between AI constraint in students’ academic writing and AI use practices in other writing contexts be too great? 

The authors gently warn writing teachers that it may become increasingly difficult for us to establish norms around writing in the classroom that promote our vision of writing for inquiry, given how differently writing is seen and practiced outside the academy and given the difficulty of constraining student use of AI.  They highlight the divides between norms around student authorship, norms around authorship in academic journals, and norms in the workplace, and ask if AI will “expose the artificiality of writing practices in the academy.” I was a little surprised that the warning wasn’t paired with any more hopeful speculation. The first impulse of many writing teachers, like me, will be toward the hopeful and evangelistic. Many of us, surely, are already preaching that teaching about text generation is an opportunity for us to make a yet more compelling case to our students that writing helps us think, both in the classroom and beyond. Professional writing tasks can be both process and product; writing a memo can be drudgery, but it can also help us figure out the nuances of how to handle a situation in a way in line with our values. But I am glad that instead of preaching to the choir, the authors have alerted us to a way in which our pedagogy may be fundamentally challenged by AI. I am left wondering if my optimism about the writing-as-thinking gospel  is more reflexive than considered.

The Assignments

Many educators feel teaching about or with AI must involve introducing new material into courses that are already overstuffed. Yet the assignments in this collection show that in many cases, teaching about or with AI can support existing learning goals. For example, Kyle Booten teaches student awareness of genre conventions by inviting students to prompt an LLM to produce an example of podcast dialogue. In the process of iteratively nudging the LLM to improve its outputs, the students become more aware of the genre conventions and also learn prompt writing techniques. Booten takes care to point out that such near-instant responsive revisions of a writing sample are not possible without an LLM. The AI-enabled exercise is used to prepare students to write their own example in the genre. 

Those of us relatively new to thinking about AI might assume that it would be preferable to use the most advanced language model for every pedagogical application. However, several contributors to the collection, such as Kyle Booten, Nick Montfort, and Alan Knowles, emphasize the benefits of asking students to work with a clunkier model because it is older or smaller, such as the open-source models available on HuggingFace. These less capable models can make the predictive nature of the systems and the way they break text into tokens more apparent. Their reliable failures are opportunities for students to articulate what is wrong and what is needed.

AI Literacy

The AI Literacy collection of assignments shows a spectrum of ways that teachers can “help students understand how the machines work,” all of which are likely to remain relevant even as AI models and interfaces evolve. Many of these assignments support other goals like reading comprehension and summary writing as well. Some are immediately useful in any course that assigns writing. These invoke little to no jargon and require little to no technical expertise on the part of the educator, yet they do teach students about the sophistication as well as the inaccuracies of autogenerated text. For example, Daniel Hutchinson and Erin Jensen show students how to prompt a language model to explain a difficult reading by inventing an example in terms of a popular culture reference of the student’s choice. The students then check whether the example really reflects the meaning of the passage and discuss what any inaccuracies reveal about AI. Nathan Murray and Elisa Tersigni suggest a blinded exercise where students critique three instructor-chosen texts. Later, these are revealed to be one accurate, authoritative human-written text; one authoritative-sounding, inaccurate human-written text; and one authoritative-sounding, inaccurate AI-generated text. Alan Knowles’ activity has students use a set of social media posts to prime a language model to generate new posts in the same style (he chooses posts about the January 6 attack on the U.S. Capitol). Then students analyze the rhetorical features of the generated posts and their implications about language models’ training data and societal impacts.

Other, more technical assignments in this section are designed to give students experience under the hood. Students work with Markov chains, find gender bias in word embeddings, and even build their own neural networks. The assignment descriptions educate teachers who are not familiar with these building blocks of large language models. They build our AI literacy even if we do not end up assigning the activities in our classes. 

Creative Explorations

The collection’s introduction celebrates creative experiments with text generation, including a history of “ a small number of artists and programmers” who “were going against the grain of what computation was generally designed to be used for.” It affirms that “ The spirit of early creative computational writing… is still very much alive.” The “Creative Explorations” assignments show us what that looks like through a variety of playful experiments that teach students how to use AI and encourage them to reflect on its limitations. Brandee Easter asks students to use generated text and images to create a children’s book, a process that spurs discussion of biases in the models. Mark Marino’s “Grand Exhibition of the Prompts” frames prompting image generators as a genre in itself, one that demands concentrated, evocative language. Dana LeTriece Calhoun has students get ChatGPT to generate spells and then compare these with their own spells inspired by the Black Southern American practice of conjuring. 

Other activities include more technical exploration. Nick Montfort asks students to use a rules-based AI system to generate multiple versions of a story and analyze the differences. Mark Sample asks students to be the ones to define rules for automatic content generation in a specific genre. Jason Boyd invites students to explore cyborg authorship using a range of lightweight, specialized text-generation programs, and kathy wu invites students to use Markov chains to create found writing inspired by Dadaism. Pretty much all of these activities sound fun and intriguing.. They will likely inspire teachers to try inviting students to prompt AI in creative genres even in classes without a creative focus.   

Ethical Considerations

I was relieved to find that TextGenEd addresses the ethics of large language models without attempting to be comprehensive.  The ethical concerns around AI text generation are legion; they demand a book of their own, like Leon Furze’s Teaching AI Ethics. The ethics-related assignments we do find in TextGenEd are variou and creatively designed. For example, Christopher D. Jimenez invites students to share information about their interests with a language model, then ask it to guess their race, class, gender, and sexual orientation, and then reflect on the bias revealed in the process.

Several assignments ask students to develop their own informed opinions about ethical AI use in college. Marc Watkins develops an academic integrity policy around AI with students. Jentery Sayers prompts students to consider whether and how a university should use AI to evaluate essays. Paul Fyfe invites students to test the instructor’s ability to distinguish an essay they have generated from one they have written. Perhaps future iterations of TextGenEd will include assignments that deal with a broader range of such as intellectual property, privacy, energy use, labor, and linguistic justice. 

Professional Writing

The assignments from teachers of professional writing introduce students to the prompting and critical evaluation skills needed to work with text generation in professional settings.  As the authors point out, these teachers may consider AI in a different light in part because “unlike academic discourse, professional writing is not grounded in an ethos of truth-seeking and critical inquiry; it tends to be grounded in an ethos of efficacy as well as constraints of legality and workplace ethics. “ Douglas Eyman’s assignment, for example, asks students to review AI summarizing software tools and reflect on how they might choose to use such apps in the workplace. Nupoor Ranade has students prompt text generators and then manually edit their outputs. In her class, this led to heated debate on whether professional human editors will still have an important role in the future.

I was intrigued to see that many activities in this section seem suited for a composition classroom as well. For example, Heidi McKee has students reflect on a hybrid process where they write a summary of a text and then with an AI-generated one. Many of her students noted “the need, always, for human decision-making and human agency in the writing process.” Tim Laquintano asks students to translate a policy document for different age audiences and compare results against LLM translations. 

Rhetorical Engagements

As the authors put it, the assignments in this section “help students build out the new rhetorical competencies enabled by LLMs and also” reflect “the possibility of using them to enhance more traditional literacies.” John J Silvestro invites students to define their own writing styles by comparing their paragraphs to stylistically different versions generated by a language model. Justin Lewis and Ted Wayland ask students to consider AI-generated counterarguments early in the research paper process so they are more likely to make fundamental improvements to their theses. Anthony Byrd has students compare human peer reviews with language model feedback on their writing and reflect critically on the AI suggestions. He frames the activity in part as a way to discourage overreliance on AI. Juan Pablo Pardo-Guerra offers students substandard auto-generated responses to his essay prompts and asks students to annotate and improve them. He sees his assignment as a kind of answer to academic integrity concerns around AI, writing “The solution to this puzzling situation is not avoiding LLMs but accepting them as extensions of our analytical and pedagogic toolkits.” 

TextGenEd as a Living Collection

“Continuing Experiments” is the last section of TextGenEd, and I was delighted to discover that it consists of a call for submissions and a promise to publish at least two new collections. The authors write, “The work of teaching writing is a constant process of experimentation, revision, and collaboration, particularly when teaching with technologies that shift and evolve quickly.” Even as we anticipate future additions to TextGenEd, they invite us to participate on our own, treating TextGenEd as “a living collection, adapting the assignments to local conditions and new technologies as they evolve.“ 

Teachers and the Future of Text Generation

In the acknowledgments for TextGenEd, there’s a comment I’ve been puzzling over. The authors write, “A backhanded thank you to OpenAI for releasing ChatGPT and instigating an AI arms race with little understanding of how LLMs will be exploited by malicious actors or weaponized against the poor. Keep on believing in that future of techno utopianism! You made this collection necessary. “ 

“Necessary” suggests that the collection is remediating a problem, yet we know the teaching experiments here do much more than that. The complaint is exaggerated, perhaps out of playfulness and anger. But on the other hand, there's an exhortation to the techies to keep going. That reminds us of the excitement, curiosity, and playfulness found throughout the introduction and the assignment descriptions. Teachers can and do enjoy exploring this terrain, engaging with these powerful language systems made on others’ terms. 

Like many commentators, the authors are angry about the ethical failings of large language models. But they are also outraged at the exclusion of writing teachers and historians of writing technologies from the responsibilities and joys of working on systems that forever change the writing environment. They write, “ “As Big Tech rushes ahead in its AI arms race with the intention of having large language models (LLMs) mediate most of our written communication, ​​writers and teachers are forced to consider…” the workings of this technology. They are indignant about the passive role into which writers and teachers have been “forced.”

Are we not needed to help shape the technology? We have had glimpses of the bafflement of the tech world as it attempts to steer text-based AI. Soon after the release of ChatGPT, OpenAI president Greg Brockman tweeted, “Big takeaway from the GPT paradigm is that the world of text is a far more complete description of the human experience than almost anyone anticipated.” He seems unaware of those who have found the world of text rich enough to dedicate our careers to it.  

In “Now The Humanities Can Disrupt “AI” Lauren Goodlad and Samuel Baker have called on educators to “claim a seat at the table where tech entrepreneurs are already making their pitch for the future.” Will teachers continue to be angry, excited, and largely reactive as others develop text generation AI further? Or will we find more active roles in the regulation and creation of LLMs? Vee, Laquintano, and Schnitzler are not quite so assertive as Goodlad and Baker, but even In this practical, pedagogy-oriented collection, they too send a message about teachers’ possible role in LLM development. They write, “writing teachers are poised to help steer the discourse and paths of generative AI technology. This collection serves to orient writing teachers in that essential work.” In addition to expanded explorations of classroom practice, I would love to see a TextGenEd collection of ways teachers are participating to help shape the future of text generation. 

However we feel about large language models, writing teachers may find this moment in the history of writing to be energizing. Immediate pedagogical questions and societal conversations about writing and technology are so closely linked. Though TextGenEd’s emphasis is critical engagement, it also feels like a joyful collection. It encourages us to turn toward other educators and celebrate our reflections on teaching as timely and essential. Those of us who lean in to “the world of text,” as Greg Brockman puts it, have something to say to shape an era where both humans and machines put words together. 

Photo by Dylan Hunter on Unsplash

Works Cited

Furze, L. (2023, September). Teaching AI Ethics. https://leonfurze.com/2023/01/26/teaching-ai-ethics/.

Brockman, G. [@GDB]. (2023, January 6). Big takeaway from the GPT paradigm is that the world of text is a far more complete description of the human experience than almost anyone anticipated [Tweet]. X. https://x.com/gdb/status/1611429677218004992?s=20.

Goodlad, L. M. E., & Baker, S. (2023, February 20). Now the Humanities Can Disrupt 'AI'. Public Books. http://www.publicbooks.org/now-the-humanities-can-disrupt-ai/.

Vee, A., Laquintano, T., & Schnitzler, C. (Eds.) (2023). TextGenEd: Teaching with Text Generation Technologies. The WAC Clearinghouse. https://doi.org/10.37514/TWR-J.2023.1.1.02

Subscribe now