Six weeks to make AI part of how you actually work

AI for Media is The Media Copilot's live six-week certification for journalists, PR pros, and communications teams who want working AI systems instead of a chatbot habit. Fully revamped for the fall cohort.

• Build custom assistants for pitches, beats, briefings and drafts that keep your voice instead of flattening it.

• Run grounded research with NotebookLM and deep-research tools across filings, transcripts and large coverage troves.

• Vibe code interactive pages and live dashboards for your reporting or campaigns, with no dev team required.

• Deploy agents that handle media monitoring, sentiment tracking and the repetitive parts of your week.

You finish with a capstone tied to your own beat, accounts or organization, plus two 30-minute one-on-one coaching sessions with me.

📅 Tuesdays, Sept. 22 to Oct. 27, 12:00 to 1:30 p.m. ET
💳 $1,500, or 25% off with code EARLYBIRD25 through Sept. 11
📦 Includes prompts, recordings, course materials and a certificate

Enroll in AI for Media

Publishers are fighting the wrong battle against AI

The bots are winning.

That’s the clearest takeaway from the first section of TollBit’s most recent State of the Bots report. TollBit builds payment rails between publishers and AI crawlers, and the report includes commentary from Jonathan Roberts, chief innovation officer at People Inc., who lays out the company’s approach to AI bots scraping content across its media properties: block unauthorized bots aggressively, allow access to legitimate crawlers, usually under some kind of licensing agreement.

It sounds good on paper. Then Roberts admits the part that matters: the bad actors have gotten too good to fully stop. Bad bots masquerade as legitimate ones like Google’s crawler, rotate IP addresses the moment they get blocked, and in some cases route traffic through networks of home devices to look like ordinary human visitors. People Inc. has real resources and deep expertise here, and it still can’t block everything.

That’s worth taking in if you run content, audience, or comms. If a company with People Inc.’s scale and know-how can’t fully lock down its content, what chance does a newsroom or brand with less resources have? Which means the fight over bot access can’t be the place where publishers define their future in the AI era. Build a strategy around winning that fight and you’ve already lost.

Read the rest at mediacopilot.ai

The Chatbox

All the AI news that matters to media*

Op-ed editors refuse to follow the Journal

None of them wanted to be next. After the Wall Street Journal defended running Stanley Druckenmiller’s AI-assisted column (see “Of course I used AI,” below), Nieman Lab’s Joshua Benton asked opinion editors at 24 prominent U.S. newspapers whether they would take the same position. Not one who responded said yes. The Los Angeles Times allows AI for research and background but not for producing submitted copy.

Underneath the policy sits a practical problem: Detection is unreliable and the familiar tells are fading. New York Focus ran 43 recent op-eds by New York politicians through Pangram and most came back flagged for at least some machine-generated text, which proves less than it appears to. There is also a gray area editors have lived with for decades: Communications staff have been ghostwriting executive op-eds all along. (AI-assisted)

The Times lets the machine write

The New York Times has started showing AI-generated summaries to a small share of readers who use its on-site search, Semafor’s Max Tani reported. The new search page returns excerpts, links and several-sentence summaries built only from Times journalism. What makes it notable is the part that’s missing. Semafor says this is the first AI-written text the Times has put in front of readers without a journalist or editor reviewing it first. The Times Guild, told during July contract bargaining that the test was coming, wants human oversight written into the contract. Unit chair Jim Luttrell said generated summaries can get things wrong and cost readers their trust. For a newsroom organized around layers of reporting, editing and corrections, that is a meaningful line to step over. (AI-assisted)

Of course I used AI, says the author

“Of course I used AI.” Stanley Druckenmiller did not hedge when NOTUS asked how his Wall Street Journal opinion column criticizing Treasury Secretary Scott Bessent came together. Economist Claudia Sahm had run the piece through the detector Pangram, which flagged the text as entirely machine-written. Journal editorial page editor Paul Gigot defended publishing it, saying the test for an outside contributor is whether the argument is the author’s own and whether the author is qualified to make it. Chris Roberts, who teaches journalism ethics at the University of Alabama, said AI use still raises questions about how much time and thought went into a piece, and expects more cases like this one. Druckenmiller was blunt: “My name is on the piece. It’s my message.” (AI-assisted)

Axios says stop betting on SEO

“SEO is not the answer anymore.” That’s Axios publisher Nicholas Johnston, who told Press Gazette that any newsroom still building audience strategy on search is working from a broken assumption. His first piece of advice is diagnostic rather than editorial. Find out where you already stand inside ChatGPT, Claude and Perplexity before rewriting a single headline. Johnston also pushed publishers to decide what they actually want from AI companies. A nickel every time a model uses your work, or your name spoken aloud in the answer. For a subscription business, he argued, attribution may be worth more than the check. Axios has been living both sides of that question since it signed a three-year deal with OpenAI in January 2025. (AI-assisted)

Chatbots caught propaganda search missed

Chatbots debunked false foreign propaganda narratives about three-quarters of the time. Search engines did worse. NPR and NewsGuard built 30 questions from disinformation pushed by China, Iran and Russia between December 2025 and July 2026, then ran them past six chatbots and four search engines, according to NPR. That produced 180 chatbot responses and 120 pages of first-page results.

Google disputed the methodology, saying answers counted as failures often carried useful context and links, and that the queries were rare rather than typical. Digital literacy researcher Mike Caulfield read the numbers differently. If three-quarters of students got these right using a search engine, he said, you would be thrilled. One caveat: A Washington University study found roughly 11% of claims in Google AI Overviews were unsupported by the pages cited beside them. (AI-assisted)

Reuters checks hundreds of images daily

Hundreds of photos and videos arrive from social media every day. Reuters publishes about a dozen. That ratio anchors an explainer the agency published on how its visual verification team decides whether user-generated imagery shows what it claims to show. The process is unglamorous: Reverse image searches, weather records, shadow angles, timestamps, plus several AI-detection tools the agency is candid about not fully trusting.

The material is also getting harder to catch. Extra fingers and garbled signage are mostly gone, and anyone can now feed real photos of real places into a model to produce altered versions of events that actually happened. Reuters points to content credentials from the C2PA standard as one durable check. For newsrooms writing a verification policy from scratch, it’s a useful template. (AI-assisted)

Reuters puts 39 years of news up for sale

Reuters is listing journalism dating back to 1987 on Snowflake Marketplace, giving enterprises a licensed route to pull news archives into AI systems, the agency announced. The collection spans five languages and is pitched at applications that track developing events, analyze companies over time and surface market or geopolitical signals.

Strategic business development director Erika Young framed the point as control, saying the integration lets Reuters participate in the AI ecosystem under terms that uphold its intellectual property. Pricing, revenue sharing and customer commitments were not disclosed. The listing follows the Model Context Protocol server Reuters launched in July, which lets subscribers’ AI agents search and retrieve content directly. Two products, one underlying bet. Distribution on your own terms beats scraping on nobody’s. (AI-assisted)

A frontier model for $40 million

What does it cost to build your own model when you already own the data? Thomson Reuters says about $40 million. That is what it spent on talent and computing power to build Thomson, a large language model announced Monday that started from an open-source foundation. Training material came from Westlaw, Practical Law, Checkpoint and Reuters, with hundreds of subject-matter experts involved from setting objectives through final evaluation.

The claim publishers should notice is the one about specialization. Thomson Reuters says its early results challenge the idea that a general-purpose model reaches expert-level performance simply by getting access to good content. Expertise, in its telling, is a separate input from the archive. (AI-assisted)

Advertising learns to count the agents

Ads are increasingly reaching machines. The industry still has no agreed way to count what happens next. The Interactive Advertising Bureau plans to publish a framework on Nov. 12 for measuring and crediting conversions influenced by AI, Digiday reported. Caroline Giegerich, the IAB’s vice president of AI, is drafting it with a working group of tech companies, publishers, agencies, measurement vendors and brands whose members she declined to name.

Jaime Schultheis of Bombora said a standard could make the publisher-platform relationship more reciprocal after years of technology companies building audiences on publisher content. OpenAds co-founder Michael Bishop sees the harder problem. If AI platforms control the signals showing how they influenced a purchase, outside measurement firms end up dependent on platform-supplied data, the way third parties were with Facebook. (AI-assisted)

*AI-assisted news items are drafted with AI and then carefully edited by Media Copilot editors.

For creators, that means you can turn your expertise into something others can subscribe to rather than something scraped out from under you. I push them on authorship, privacy, pricing, and who gets paid when human knowledge finally scales.

The AI course built for reporters and PR pros

AI for Media is The Media Copilot’s live six-week certification for journalists, PR pros, and communications teams who want working AI systems instead of a chatbot habit. Fully revamped for the fall cohort.

• Build custom assistants for pitches, beats, briefings and drafts that keep your voice instead of flattening it.

• Run grounded research with NotebookLM and deep-research tools across filings, transcripts and large coverage troves.

• Vibe code interactive pages and live dashboards for your reporting or campaigns, with no dev team required.

• Deploy agents that handle media monitoring, sentiment tracking and the repetitive parts of your week.

You finish with a capstone tied to your own beat, accounts or organization, plus two 30-minute one-on-one coaching sessions with me.

📅 Tuesdays, Sept. 22 to Oct. 27, 12:00 to 1:30 p.m. ET
💳 $1,500, or 25% off with code EARLYBIRD25 through Sept. 11
📦 Includes prompts, recordings, course materials and a certificate

Enroll in AI for Media

Building the iTunes moment for human knowledge

Today, an expert’s work can become part of an AI-generated answer without identifying the source, compensating the creator, or accurately reflecting that person’s point of view.

Onix wants to build a different model.

On this episode of The Media Copilot podcast, Pete Pachal speaks with Onix co-founders David Bennahum and Nicholas Nadeau about their vision for expert-owned AI: specialized models built around an individual expert’s knowledge, voice, judgment and perspective rather than a massive model trained to know a little about everything.

Listen or watch:

• YouTube

• Spotify

• Apple Podcasts

Bennahum and Nadeau argue that expertise is more than factual information. It includes the judgment and perspective an expert develops over a lifetime, including knowledge that may never have appeared online. Onix combines publicly available work with what the company calls “dark data,” including private notes, manuscripts and other material that can capture how an expert actually thinks.

Expertise is factual knowledge plus a point of view, a perspective and judgment.

— Nicholas Nadeau, Ph.D., P.Eng., Co-Founder & CTO, Onix

The conversation goes well beyond the technology. Pete digs into authorship and intellectual property in the AI era, data sovereignty and privacy, the economics of expert knowledge, small language models versus frontier models, hallucinations, and whether creators can build sustainable businesses around AI versions of their own expertise.

Bennahum describes the opportunity as creating an “authorized version” of an expert in AI, rather than relying on what he calls the “bootleg” version assembled by general-purpose models from information found across the internet.

Take back control of your genius.

— David S. Bennahum, Co-Founder & CEO, Onix

The implications could extend far beyond Onix’s initial focus on health and wellness. The company envisions expanding into personal finance, self-improvement, relationships and lifestyle expertise, with a model that could eventually let more creators build, privately share and monetize their own Onix.

What we cover

• Why Onix is betting on small, domain-specific AI models

• The difference between factual knowledge and true expert judgment

• Why AI’s current approach to authorship and intellectual property may be unsustainable

• Giving experts ownership and control over their AI models

• The role of private or unpublished “dark data” in creating higher-fidelity AI

• Why privacy and data sovereignty are central to the Onix model

• Whether specialized models can reduce hallucinations

• Turning an expert’s knowledge into a subscription business

• The potential for journalists, creators and other professionals to build their own AI products

• What a healthier “author economy” could look like five years from now

Why this matters

Pete also asks the larger question hanging over the creator economy. If AI is going to generate enormous value from human knowledge, who should own that value?

For Onix, success would mean the economic benefits of AI are distributed among the experts whose knowledge makes these systems valuable, rather than concentrated among a handful of technology companies.

It’s a conversation about AI, but ultimately it’s about authorship, ownership and what human expertise is worth in a world where knowledge can suddenly scale.

Onix is currently in early access. Media Copilot listeners can visit Onix.life and use the access code COPILOT to try the platform.

About the guests

• David S. Bennahum, Co-Founder & CEO, Onix — LinkedIn

• Nicholas Nadeau, Ph.D., P.Eng., Co-Founder & CTO, Onix — LinkedIn

• Company: Onix.life

Produced by Pete Pachal and Executive Producer Michele Musso. Edited by the Musso Media team. Read the full post on mediacopilot.ai.

The clicks are real. They’re just landing somewhere other than news. Today’s column looks at what changed inside ChatGPT that day, why the timing lines up so neatly with OpenAI’s advertising plans, and what happens to everyone in media once in-line links become the thing brands compete for. Further down, in The Chatbox, three years of Press Gazette data show how much ground the biggest US news sites gave up while all of this was going on.

Six weeks to make AI part of how you actually work

AI for Media is The Media Copilot's live six-week certification for journalists, PR pros, and communications teams who want working AI systems instead of a chatbot habit. Fully revamped for the fall cohort.

• Build custom assistants for pitches, beats, briefings and drafts that keep your voice instead of flattening it.

• Run grounded research with NotebookLM and deep-research tools across filings, transcripts and large coverage troves.

• Vibe code interactive pages and live dashboards for your reporting or campaigns, with no dev team required.

• Deploy agents that handle media monitoring, sentiment tracking and the repetitive parts of your week.

You finish with a capstone tied to your own beat, accounts or organization, plus two 30-minute one-on-one coaching sessions with me.

📅 Tuesdays, Sept. 22 to Oct. 27, 12:00 to 1:30 p.m. ET
💳 $1,500, or 25% off with code EARLYBIRD25 through Sept. 11
📦 Includes prompts, recordings, course materials and a certificate

Enroll in AI for Media

Brands are the biggest winners of ChatGPT's traffic surge

OpenAI made an important change to how ChatGPT does links back in the spring, and hardly anyone noticed. Starting May 7, ChatGPT began sending considerably more traffic to websites: referrals jumped 157% overnight and have held at that level since, according to new data. That same day, OpenAI also began rolling out advertising inside AI answers, a project that had been in the works for months.

The two changes are connected, and the proof is in where those links are pointing to: homepages. Since the change, when a ChatGPT answer recommends a brand, it'll often drop an in-line link straight to that company’s homepage rather than the small footnote citations or link cards that sit at the bottom of the response. In-line links reliably pull more clicks than footnotes, because people want the extra context right when they’re reading, not after.

So why would OpenAI make links more prominent in its answers at the exact moment it starts selling ads? The answer matters for advertisers and publishers alike, and it complicates the assumption that AI is quietly killing off the open web. The zero-click future was always an exaggeration, but the shape of the web inside an AI-driven economy is shifting, and anyone in media or marketing should understand what it’s shifting into.

Where answers become clicks

The prevailing AI narrative goes something like this: people ask chatbots for answers, they never leave the chat window, and the web slowly turns from a place people visit into a pile of data that machines scrape. But a recent Similarweb report complicates that story. It shows the share of ChatGPT referral visits landing on brand homepages jumped from roughly 25% to 60% after May 7, and other data puts ChatGPT at 63% of all AI-driven referral traffic already.

The good news: people still click on links when they need something specific. The bad news, if you work in news, is where those clicks are going. Ask ChatGPT about the latest developments in the war in Iran, or which movies topped the box office this week, and you'll see plenty of links, but they'll be at the bottom, not in-line. Brands get the prime real estate. Newsrooms get the footnotes.

Read the rest at mediacopilot.ai

Graduate from AI tools to AI agents

AI Quick Start is The Media Copilot’s 1-hour class for journalists, marketers, and communications professionals who want to move past chatbot basics and start building real AI workflows into daily work.

• Use the latest AI apps that don’t just treat AI like supercharged Google, but actually get things done.

• Understand how to use deep research for professional-grade analysis of your beat or client data.

• Build reliable AI assistants that accelerate tedious work so you can concentrate on what matters.

You’ll also get an introduction to agents—using AI to take on bigger tasks by operating apps just like you can. Once you go agent, you never go back.

📅 Next class: Friday, Sept. 11 at 1 p.m. ET
💳 Price: $49
📦 Includes: Prompts, deck, recording, and more

It works as a standalone skills boost or as an on-ramp to AI for Media, and your seat is just $49.

Enroll in AI Quick Start

The Chatbox

All the AI news that matters to media*

CNN lost 61% of its search traffic

CNN's organic search traffic fell from 124.4 million monthly visits in July 2024 to 48.4 million this July, the steepest drop among the ten largest U.S. news sites, according to a three-year Press Gazette analysis of Similarweb data. Combined traffic across the top 50 sites peaked at 4.9 billion visits in July 2024, shortly after Google began rolling out AI Overviews to all U.S. users, and now sits at 3.1 billion. The timing is suggestive, though the data stop short of proving cause. A few outlets held their ground: The New York Times was essentially flat, People rose 10%, and traffic to Substack URLs jumped 179%. What the numbers do not show is a replacement channel: Generative AI accounted for an average of 0.4% of total traffic across those 50 sites, and 0.2% at the Times. (AI-assisted)

Licensing deals buy ChatGPT citations

Publishers with OpenAI licensing deals averaged 10.2 ChatGPT citations per news page in June. Publishers without them averaged 6.9. That 48% gap is the headline finding in a joint Press Ranger and OtterlyAI study that matched 129.3 million citations across seven AI platforms against confirmed licensing agreements. Outlets signed only with OpenAI did better still, pulling 112% more citations per page than unlicensed rivals. The obvious caveat is that none of this establishes causation, and both companies behind the study sell products tied to AI visibility. One pattern is harder to wave off: OpenAI-licensed publishers drew 57.9% of their AI citations from ChatGPT, a far more concentrated mix than everyone else. Service journalism dominated what got cited, with best-of lists, buying guides and product reviews accounting for 46.9% of citations to licensed publishers. (AI-assisted)

Blocking crawlers is winning licensing deals

Blocking AI crawlers has turned into negotiating leverage. People Inc. CEO Neil Vogel told investors on the company’s Aug. 4 earnings call that its Cloudflare blocking is what brings AI companies to the table, Press Gazette reports. “People are understanding the value of our content because we are able to restrict almost everybody from using our content,” Vogel said. “And when people have to pay for it, it seems they’re really coming to the table.” Cloudflare chief strategy officer Stephanie Cohen calls the tactic “reliable scarcity,” and says the Financial Times, The Atlantic, Ziff Davis, Condé Nast and the Associated Press are all working with the company. Next comes pay-per-use, which would pay publishers when their material actually appears in an AI answer rather than when a bot merely fetches the page. (AI-assisted)

OpenAI is quietly funding local newsrooms

Robert Sanchez writes a newsletter covering Arapahoe County, Colorado, on Mondays and Wednesdays, then does the same for neighboring Douglas County on Tuesdays and Thursdays. Every local byline in both belongs to him. His job exists because OpenAI is paying for it, part of a three-year partnership funding startup costs for 13 Axios Local newsletters in exchange for training rights on Axios’s freely available journalism, as Issie Lapowsky reported for CJR. Axios CEO Jim VandeHei frames it as a test of whether AI can automate everything a local market needs except the journalist. The economics remain unsettled. Axios Local has more than 2 million subscribers and is not profitable, which the company says was expected. VandeHei wants each market profitable within five years, expects to operate in 43 cities by year end, and eventually wants 1,000. (AI-assisted)

GB News automates away a third of staff

GB News told staff in July it was proposing to cut as many as 90 jobs, roughly a third of its workforce, as it introduced automated production workflows. A spokesperson said the broadcaster “has led the way in introducing groundbreaking workflows in UK media,” and stopped short of attributing the cuts to AI alone. The proposal is one line in Press Gazette’s rolling tracker of 2026 journalism job cuts, which logged at least 3,434 cuts in 2025 and 3,875 in 2024. Automation is only part of the picture. Pink News proposed cutting up to 11 staff after Snapchat retired formats it depended on, BuzzFeed shed about 180 employees following Byron Allen’s acquisition, and the Chattanooga Times Free Press laid off more than 50 around the close of its sale to Firecrown Media. (AI-assisted)

A TV news channel for $50,000

Gaurav Misra built and ran a 24-hour television news operation for about $50,000, roughly $30,000 of which went to Claude Code credits, and hired no journalists to do it. The Mirage News Network aired Aug. 11 with four AI-generated anchors styled after 1980s broadcasters, reading Reuters-licensed stories that were already several days old, Variety reported. It drew about 50,000 viewers on X, though Mirage paid several accounts to promote the stream. The guests were partly synthetic too, with more than a dozen people, many of them venture capital professionals, licensing their likenesses to AI doubles reading approved scripts. Misra told Variety he instructed the system to follow “general journalistic standards” rather than AP style or the SPJ code of ethics. On-screen graphics did disclose that the anchors, guests and ads were artificial. (AI-assisted)

What actually breaks an AI watermark

How much editing does it take to strip a watermark off AI-written text? Anthropic has now put some boundaries around the answer, detailing in a TechCrunch post how its Claude watermarking plan holds up under revision. The system, built on Google DeepMind’s SynthID Text, nudges word choices into a detectable pattern, picking “overcast” over “grey” in ways no reader would notice. Light editing probably will not remove it. A complete rewrite that replaces every word will. And if someone hands Claude a human-written draft for minor edits, there is “very little (if anything) for the watermark to attach to.” For newsrooms writing disclosure policy, the practical read is that watermarks flag heavy machine authorship, not the light-touch assistance most journalists actually use. (AI-assisted)

Fifty minutes of verification becomes three

A six-person fact-checking team at India Today spends its days sorting AI-generated archival photos, flood footage stripped of context and deepfakes of public figures. It is now testing Backstory, an experimental Google DeepMind tool that pulls several verification steps into one interface, as Nieman Lab’s Andrew Deck reported. A reporter uploads an image, asks a question, and Backstory decides which checks to run, from SynthID watermark lookups to C2PA content credentials to tracing where an image has surfaced before. Digital literacy researcher Mike Caulfield estimates it can compress a 50-minute first pass into roughly three. The tool is free but limited to Google’s Trusted Testers Program, and it still requires human review. That caveat matters, given a July Princeton report finding most authentication tools are built with too little newsroom input. (AI-assisted)

UK editors want AI fluency on day one

Basic familiarity with AI no longer clears the bar for entry-level journalism jobs in Britain. Fourteen UK news organisations told the National Council for the Training of Journalists that graduates need critical prompting, fact-checking, context awareness and ethical judgement, Journalism.co.uk reports, with particular weight on identifying bias, verifying AI-generated material and catching hallucinations. That carries force because the NCTJ is still the main door into British newsrooms. Its Journalists at Work 2024 survey found 81% of UK journalists hold a journalism qualification, and 83% of those came through NCTJ programmes. The body is folding AI literacy into the existing Diploma in Journalism rather than bolting on a standalone module, and it will now appear in exams and portfolio work. Editors hiring in the U.S. should expect the same baseline to arrive shortly. (AI-assisted)

*AI-assisted news items are drafted with AI and then carefully edited by Media Copilot editors.

We also want to highlight AI for Media, our upcoming live course for journalists, PR pros, and comms teams who wan…

Read more

First, though, two pieces of news from our side of the desk:

We're hiring. If you've been visiting The Media Copilot lately, you've probably noticed there's more of it. More…

Read more

Also in this issue, a loaded Chatbox: a federal judge keeps Reddit's lawsuit against Perplexity alive, the EU makes AI content labels mandatory, and Google quietly rewires how its To…

Read more

Read more

Also in this issue: The AI effect on the media ecosystem is getting real. G…

Read more

Read more

First, a quick word about an event I’m hosting later this month.

The AI search conversation for media leaders

The GEO Dinne…

Read more

Extremely related: The Media Copilot’s GEO Dinner Series continues this month with an event for media lea…

Read more

Read more

Graduate from AI tools to AI agents

AI Quick Start is a one-hour live workshop for journalists,…

Read more

Read more

Read more

Read more

Reminder that AI for Media—our 6-week AI training program for journalists, PR professionals, and communicators st…

Read more

Before that, a quick reminder: Enrollment is open for AI for Media, a six-week deep dive into AI…

Read more

Also, my n…

Read more