On August 2, 2026, the European Commission gained the legal authority to fine general-purpose AI providers up to 15 million euros or 3% of global turnover, whichever is higher. Every trade publication covering the enforcement date ran some version of the same headline: OpenAI, Anthropic, and Google are now under the microscope. That is accurate, and it is also the least interesting part of what happened this month. The providers spent most of a year preparing for this date. The companies that got caught flat-footed are the ones nobody wrote a headline about: the teams that built a product on top of one of those models and assumed the vendor’s compliance covered them too.

It did not. It was never going to. The Act does not work that way, and the first thirty days of enforcement made that distinction expensive to ignore.

WHAT ACTUALLY WENT LIVE ON AUGUST 2

Two things changed on that date, and they are easy to conflate if you only read the headlines. First, the Commission’s AI Office gained enforcement and penalty powers over general-purpose AI model providers: the authority to demand technical documentation, run independent evaluations, and restrict a model from the EU market if a provider will not cooperate. Second, and separately, Article 50’s transparency obligations became binding for everyone building AI-facing products, not just the labs training the underlying models. Chatbots have to identify themselves before a person starts talking to one. Synthetic audio, video, and images need machine-readable marks. Anyone deploying emotion recognition or biometric categorization has to tell the people it is scanning. Deepfakes touching matters of public importance need a disclosure, unless the content went through genuine editorial review.

Most coverage focused on the first thing because it has famous names attached. The Commission-recognized Code of Practice on GPAI models had roughly 190 signatories by July, including OpenAI, Anthropic, Google, Microsoft, IBM, Mistral, Cohere, and Amazon, all of whom get a presumption of conformity in exchange for signing. Meta refused. Chief Global Affairs Officer Joel Kaplan said the code “introduces legal uncertainties for model developers” and “goes far beyond the scope” of the law itself, which is a polished way of saying Meta would rather prove compliance the hard way than commit to a documented pathway everyone else agreed to. That is a real story, and it is worth watching. It is also a story about eight or nine companies with enough legal staff to make that calculation on purpose.

The second story is about everyone else, and it determines whether the Act does what it claims to do.

Sources: Cooley: EU AI Act Transparency Obligations Take Effect 2 August 2026 | Praxikon: EU AI Act code of practice, tech giants sign, Meta refuses | Regulation AI: EU AI Act Enforcement Has Started

THE DEPLOYER BLIND SPOT

Here is the distinction that got lost in a month of coverage about frontier labs: the Act splits obligations between providers, the companies that build and train a model, and deployers, the companies that put that model in front of an actual user. OpenAI, Anthropic, and Google are providers. If your company calls their API to power a support widget, a hiring tool, or a content generator, you are the deployer, and Article 50’s disclosure duties sit with you, not with them. A provider’s technical documentation was built to satisfy a regulator reading a model card. It was never built to become the sentence your own user sees before they start typing into your chatbot.

A Tech Times report on the rollout put it plainly, quoting the practical reality builders ran into the week the obligations went live: “OpenAI cannot put a chatbot disclosure banner in your product. Anthropic cannot label your deepfake. Google cannot tell your user they are talking to AI.” Every one of those sentences describes a decision some team has to make inside their own interface, and no vendor contract makes that decision for them. The same reporting cited a survey finding that seventy-eight percent of organizations subject to the Act had taken no meaningful compliance steps as of June, weeks before enforcement began. I would treat that single figure carefully since it comes from one report, but the direction is consistent with what every compliance lawyer I have read this month is saying in less quotable language: the assumption that “we just use their API, so we are covered” was never true, and August 2 is the date that assumption stopped being free.

Call it the deployer blind spot. A provider publishes a model card, a system card, a research paper on limitations, thousands of words of documentation aimed at regulators, researchers, and other engineers. None of that documentation was written for the person using your product, and the Act does not accept it as a substitute for the disclosure your product owes that person directly. The obligation does not transfer. It multiplies, once for every deployer standing between a foundation model and a real user.

Sources: Tech Times: EU AI Act Chatbot Disclosure Reaches API Builders | Regulation AI: Article 50, In Force Since 2 August 2026

THE QUESTION YOUR TEAM HAS NOT ANSWERED YET

If you write documentation, sit in DevRel, or run product at a company that ships anything touching a foundation model, the first thirty days of enforcement handed you a specific, answerable question: who owns the disclosure inside your interface, and where is that decision written down? Not the vendor’s terms of service. Not a compliance slide from a kickoff meeting in March. An actual owner, and an actual artifact a regulator, or a user, could point to.

This is the same gap I have been naming for a year, just relocated. The documentation gap is an ethics gap, whether the missing explanation sits between an AI company and its users or between a foundation model provider and the thousands of companies quietly assuming someone else already handled the paperwork. A governance framework that lives in a vendor’s trust center does not protect the deployer sitting three layers downstream, and it was never designed to. The first month of enforcement did not punish the companies with the most sophisticated AI. It started separating the companies that know exactly who owns their disclosure language from the companies still assuming it is somebody else’s job.

Which one is your team, honestly? And if you are not sure, who would you even ask?

Reply and tell me. I am tracking patterns across documentation and DevRel teams for the next framework, and this is exactly the kind of answer I want to hear before I write it.

P.S. Wednesday, I am grading two companies on precisely this gap: one that treated Article 50 as an interface decision, and one that treated it as a legal memo nobody outside compliance ever read. The difference in what their actual users see is not subtle, and it is not close.

Connect:

👥 MentorCruise | ☕ Buy Me a Coffee

The Documentation Engineer grades AI companies on documentation quality through an ethics lens every Wednesday, with every claim linked to a primary source you can check yourself.

One of my best-performing LinkedIn posts was an AI fellowship deadline roundup, so I’m turning it into a monthly series to share this in-demand resource regularly. Each article will cover deadlines, pay, who they’re looking for, the application link, and an FAQ for every fellowship. I am applying to these myself, so I built this to be the resource I wish existed- as straightforward and user-friendly as possible because I know finding and applying for them is complicated enough without any added headaches.

You can check out this fellowship resource on the last Monday of every month, with details on what’s closing and what to watch for. If you know any fellowships I should add to the list, or have feedback on how I can improve this resource, feel free to message me.

Closing in September (sorted by deadline)

1. AI Societal Impact Lab — Autumn 2026 Fellowship

Deadline: September 4 (rolling, may close early once spots fill). Format: Fully remote, ~4 hrs/week, 8 – 10 weeks, starts mid-to-late September.

Pay: Unpaid. Program costs are covered; no stipend.

Who it’s for: Recent grads and early-career people from any background. No AI experience required; they specifically want social science, humanities, and non-technical applicants.

Apply: aisocietalimpactlab.com/fellowship-autumn-2026

FAQ

• Requirements: No technical background and no prior AI experience. They want genuine curiosity about how AI affects real people, plus the ability to commit roughly 4 hours a week to reading and writing.

• Is it entry-level: Yes, the most beginner-friendly program on this list. Built specifically for recent grads and early-career people.

• Transferable skills: The closest match here for documentation and DevRel backgrounds. Explaining a complex system to a non-expert reader is the whole job, not a side skill.

2. MATS (ML Alignment & Theory Scholars) — Winter 2027

Deadline: September 6, 11:59 pm anywhere on Earth.

Format: 12 weeks, January 19 – April 10, 2027, Berkeley and London, with remote options. Pay: $1,600/week ($19,200 total), plus travel, housing, and meals covered.

Who it’s for: People who want to do alignment research under a named mentor. No PhD required; strong technical/quantitative backgrounds do well.

Apply: matsprogram.org/apply

FAQ

• Requirements: 18+, a two-stage application, and enough technical or quantitative grounding to do research under a mentor. MATS itself says prior AI safety experience isn’t required, but the selection bar is real.

• Is it entry-level: Officially open to all experience levels. In practice, it favors people with some research or technical background over a true first step into AI.

• Transferable skills: Writing up findings clearly matters to mentors once you’re accepted, so documentation experience helps there. It won’t get you in on its own without the technical or research ability they’re actually screening for.

3. FAS AI Safety Policy Entrepreneurship Fellowship 2026–2027

Deadline: September 7, 11:59 pm ET

Format: ~5 hrs/week, September 30, 2026 – February 28, 2027

Pay: $5,000 stipend, plus a possible $1,000 merit award and travel coverage for required in-person events

Who it’s for: Early to mid career professionals from technical AI research, law, academia, industry, or national security who want a structured on-ramp into policy work.

Apply: fas.org/career/ai-safety-pef

FAQ

• Requirements: Early- to mid-career, coming from technical AI research, law, academia, industry, cybersecurity, or national security. No degree required.

• Is it entry-level: Yes, for policy work specifically. FAS says this directly fits people with limited hands-on policy experience.

• Transferable skills: A strong fit. Fellows write and publish a policy memo with FAS’s own editors, so clear writing is part of what gets selected for, not incidental to it.

4. ERA: AI Fellowship — Winter 2027 (Cambridge, UK)

Deadline: September 13, 11:59 pm anywhere on Earth.

Format: 10 weeks, starts January 18, 2027, in-person in Cambridge. Pay: £10,000 stipend, plus visa sponsorship, meals, and travel.

Who it’s for: AI safety and governance researchers, technical or policy track.

Apply: erafellowship.org/fellowship

FAQ

• Requirements: 18+, no other formal restrictions. ERA calls itself “talent-first” and names technical, policy, security, legal, and economic backgrounds as fits.

• Is it entry-level: Yes, one of the more openly career-changer-friendly programs on this list.

• Transferable skills: Strong fit for the governance track. ERA specifically wants people who can point existing expertise at AI risk; documentation ethics work is close enough to that to name directly in an application.

5. NCAAT AI Fellowship (inaugural cohort)

Deadline: September 13 (rolling review).

Format: Cohort starts October 2026. Pay: This one runs the other direction; it costs $6,500 to participate. Most fellows get it covered by their institution’s professional development or grant funding. Worth listing because your higher-ed and institutional-leader audience should know it exists, but it is not a funded fellowship.

Who it’s for: Faculty, accreditation liaisons, and institutional leaders carrying AI strategy from plan to practice. No AI expertise required.

Apply: ncaat.com/fellowship

FAQ

• Requirements: You need an existing institutional role (faculty, accreditation liaison, technology leader) and a real AI adoption problem to work on. No AI expertise required.

• Is it entry-level: No. This assumes you already have standing inside an organization; it’s not a first foothold in AI work.

• Transferable skills: Only relevant if you’re already inside an institution steering AI adoption. Not a way in from outside.

6. LASR Labs — Winter 2027 (London)

Deadline: September 20, 23:59 GMT.

Format: 13 weeks, January 11 – April 9, 2027, in-person in London.

Pay: £15,000 stipend, plus food, office space, and travel.

Who it’s for: People with ML engineering chops and something they’ve actually shipped (a result, tool, experiment, or proof), not just interest.

Apply: lasrlabs.org

FAQ

• Requirements: Machine learning engineering experience, strong quantitative skills, and something already built: a result, tool, experiment, or proof. No formal credential required, but the technical bar is real.

• Is it entry-level: No, not without technical work behind you already. LASR is explicit that a degree isn’t required (past fellows include a self-taught engineer and an undergrad without a finished degree), but having no shipped technical work is a real barrier.

• Transferable skills: Communication is one of three things LASR screens for, alongside technical and research ability, so documentation skill strengthens an application here. It doesn’t replace the technical requirement.

Opening in September

7. Claude Corps (Anthropic) — Cohort 3

Opens: September 2026, exact date not yet posted

Format: 12-month fellowship placed inside a U.S. nonprofit, cohort begins August 2027

Pay: $85,000/year salary plus full benefits (medical, dental, vision, 401k, PTO)

Who it’s for: Under two years of full-time work experience, 18+, authorized to work in the U.S. No degree or coding background required; they want people who already use AI tools regularly.

Apply: anthropic.com/claude-corps/fellow.

FAQ

• Requirements: 18+, under two years of full-time work experience, authorized to work in the U.S., and already a regular AI tool user. No degree or coding background required.

• Is it entry-level: Yes, explicitly. This is the one on the list built for people early in their career.

• Transferable skills: High relevance. Fellows are placed inside nonprofits to help non-technical staff use AI responsibly; that’s documentation and training work with an AI label.

Worth knowing, but not actionable right now

OpenAI Safety Fellowship isn’t accepting applications (that window closed back in May), but the first cohort’s research phase starts September 14 and runs through February 5, 2027. Good one to keep an eye on for timing and to watch for when the next application round opens.

Sources

• MATS Winter 2027

• MATS Application

• AI Societal Impact Lab, Autumn 2026 Fellowship

• FAS AI Safety Policy Entrepreneurship Fellowship

• ERA Fellowship

• NCAAT AI Fellowship

• LASR Labs Winter 2027 — Opportunity Desk summary

• LASR Labs

• Claude Corps fellows FAQ

• Claude Corps announcement

• OpenAI Safety Fellowship

Connect:

👥 MentorCruise | ☕ Buy Me a Coffee

The Documentation Engineer grades AI companies on documentation quality through an ethics lens every Wednesday, with every claim linked to a primary source you can check yourself.

I’ve spent the last two months testing what it would take to replace my full-time income with content built around my own expertise. I have years of experience as a Software Engineer, Documentation Engineer, Developer Advocate, and Community Builder, so I wanted to combine that knowledge and channel it into the same quality of content my previous employers hired me for. Watching how many laid-off tech professionals around me turned to content and creator work this year gave me the push I needed to stop testing the idea in private and actually bet on myself in public.

The first 2 months were a bit of trial and error. I collected data, watched what landed and what didn’t, and adjusted in real time. August was when it finally clicked. Outside of my newsletter, my audience responds most to three things: educational opportunities I surface, like fellowships, workshops, courses, and certifications; personal updates on my own learning journey and the AI tools I come across in my research; and my job-hunt journey. None of those are the polished, purely technical content I assumed would perform best. So, after finding this out, I am going to reposition the Documentation Engineer to better serve the community that I am building in ways that provide the most value. I created my Substack because I wanted to be a resource and share super technical information in a human-centered way.

Here’s what the last 90 days on LinkedIn look like:

The post that got the most engagement was the one where I shared information on the AI Safety Fellowships that were available and those with deadlines approaching soon. That post alone accounted for most of my engagement over the last 90 days and is still generating traffic 2 weeks after it was posted. I also had individuals reach out to thank me for sharing the information and let me know they had applied for Fellowships because of my post. I’m not going to pretend one post proves a trend, but it did show me something specific: the people in my audience who want to break into AI safety and AI ethics work don’t have an easy way to find these opportunities, and when someone hands it to them in one place, they act on it. I’m going to cover more of these opportunities going forward, because the response told me they’re solving a real problem.

Here’s what the numbers for that post look like:

What I have learned from my years in tech is that, as both a Software Engineer and a Technical Writer, is that the biggest opportunities are usually sitting in the problem everyone else has written off as too small or too obvious to matter. People don’t just want more information. They want someone to do the work of finding it, filtering it, and handing it to them in a format they can act on in under five minutes. People value convenience.

So here’s where I’m taking The Documentation Engineer next. I’m repositioning it away from being purely a technical deep-dive and toward being the resource I set out to build when I started this Substack: technical enough to be useful, human enough that you’d actually want to read it. That means more curated opportunity roundups, more of my own learning-journey and job-hunt updates in real time, and more of the AI tools I’m actually testing instead of just the frameworks I’m building. The technical analysis isn’t going anywhere. It’s just no longer going to be the only door in.

Two months in, the biggest lesson isn’t about content at all. It’s that betting on yourself means being willing to let the data tell you who you’re actually building for, even when it’s not who you thought.

Connect: 👥 MentorCruise | ☕ Buy Me a Coffee

The Documentation Engineer analyzes AI documentation quality through an ethics lens. Wednesday issues grade real companies. Friday essays are the building-in-public version of the same work.

TL;DR: The Code of Practice on Transparency of AI-Generated Content receives a C+ (77.85%) as a document, because it is unusually honest about its own limits and reaches three of Article 50’s four duties, and the Commission’s public signatory register receives an F (46.10%) because it publishes two columns of names with no scope, no dates, and no way for anyone to check what a signature establishes. Grades use the ADECP framework (AI Documentation Ethics and Capability Positioning), scoring Ethical Stakes, Accessibility, Positioning, and Execution. This week I am not grading a company. I am grading Brussels, I raised one grade mid-audit when a primary source contradicted my draft, and every claim below links to a source so you can check my work. 25-minute read, calculated at 250 words per minute with URLs and markdown stripped.

FIRST, FIVE TERMS YOU NEED

Skip this if you already have them.

Code of Practice. A voluntary rulebook drafted by outside experts and assessed as adequate by the Commission and the AI Board. Signing it is not the law. It is a recognised way of demonstrating you followed the law, the way keeping receipts helps enormously during an audit without being the tax code itself.

Signatory register. The public list of who signed. It is the artifact journalists cite, procurement teams screenshot, and vendors link to in sales decks, which is why its quality matters independently of the Code’s quality.

Article 50(2), (4) and (5). The marking duty on providers, meaning machine-readable tags inside generated files. The labelling duty on deployers, meaning the visible mark a human can see. And the horizontal requirement that the information reaching people be clear, distinguishable, and accessible. The Code covers these three.

Article 50(1). The duty to tell you that you are talking to a machine. The Code does not reach it, which matters because it is the clause governing nearly every AI interaction an ordinary person has.

Provider and deployer. The provider builds the tool. The deployer publishes what comes out of it. One company can be both, and several signatories are.

THE HOOK

The Commission runs two pages about this policy. The policy page was last updated 31 July 2026 and tells you that “about 190 companies organisations have signed the code,” typo included. The register page was last updated 20 August 2026 and gives numbers the policy page does not: Section 1 has 82 signatories, Section 2 has 152. Same institution, same month, same subject, two different accounts of how many people made a promise, and no indication on either page that the other exists in a different state.

Then read the names rather than the counts. Adobe makes Firefly. Midjourney makes Midjourney. Stability, Runway, ElevenLabs, Suno, xAI, Canva, Amazon, Apple, Perplexity, IBM, and NVIDIA all ship or host generative systems. As of the 20 August update, none of them appears in either column.

So this week the Watchlist grades two documents instead of two products, because the failure belongs to a regulator rather than a vendor, and because a public register nobody can check is a documentation-ethics problem in exactly the way this newsletter has spent eight months arguing.

• Commission policy page for the Code

• Commission signatory register

THE AI DOCS WATCHLIST

📄 The Code of Practice on Transparency of AI-Generated Content (C+, 77.85%)

What Users Need to Know:

The Code translates three of Article 50’s duties into concrete technical measures. Section 1 tells providers how to mark generated audio, images, video, and text so machines can detect the origin, and commits them to making a detection mechanism publicly available. Section 2 tells deployers how to visibly label deepfakes and AI-generated text published to inform the public, using an icon set the Commission publishes separately. Running across both is Article 50(5), the requirement that what reaches a person be clear, distinguishable, and accessible, which both working groups treated as a cross-cutting concern rather than a footnote.

What makes this document worth grading is how much it concedes. The technical reality is that no current solution satisfies all four statutory requirements at once, so the Code mandates a layered approach using digitally signed metadata plus imperceptible watermarking, with fingerprinting as an optional third. It sorts its own contents into three tiers, where “will” means mandatory, “encouraged” means recommended, and “may” means optional. And in the objectives of both sections it states that adherence “does not constitute conclusive evidence of compliance.” Instruments with commercial incentives to oversell themselves rarely undercut their own marketing that plainly.

Where This Information Lives:

• The Commission’s policy page for the Code, last updated 31 July 2026: Confirms the scope as Article 50(2), (4) and (5), documents the full drafting timeline from September 2025 to the closing plenary on 10 June 2026, names the working group composition, and states what happens to organisations that do not sign.

• The Code itself, published 10 June 2026: The numbered commitments, measures, and sub-measures, plus the glossary.

• A separate Commission page: The EU icon set for labelling, which lives apart from the Code rather than inside it.

• The Commission Opinion and the AI Board Adequacy Assessment: The formal blessing, without which signing buys nothing.

• Nowhere in the Code: Article 50(1), the chatbot-disclosure duty, and Article 50(3), emotion recognition and biometric categorisation notice.

Visual Evidence:

The Code arrives as a PDF with numbered commitments and lettered measures beneath each. Its three-tier verbs are not colour-coded, boxed, or collected in a table anywhere, so a reader distinguishes a mandatory measure from an optional one only by noticing which verb was used in that sentence. The policy page around it is clean, dated, and available in 24 languages, and it carries a stale headline figure the sibling page has already superseded.

7-Question Audit Results:

Total: 12/14 points, Ethically Documented

The highest 7-Question total I have recorded this year, sitting next to a C+. Hold that, because it does work later.

ADECP Category Breakdown:

• Ethical Stakes (25%): 82%. The Code names its purpose plainly, mitigating deception and manipulation, and builds a mechanism aimed outward rather than a defence aimed inward: a publicly available detection route, as a specification, software, or a free API, so a person can check a file instead of trusting a claim. Carrying Article 50(5) across both sections means the accessibility of the disclosure was treated as substantive rather than cosmetic, which almost no vendor documentation manages. It loses points because the duty covering the machine you actually talk to is untouched.

• Accessibility (25%): 75%. Free, public, downloadable, available in 24 languages, accompanied by plain-language FAQ pages, and shipped alongside a standardised icon set so the visible label does not depend on 27 member states inventing 27 designs. The Code itself is dense legal-technical prose no member of the public will read, and the icons carry most of the accessibility burden alone.

• Positioning (20%): 85%. The highest single category score I have given this year. A voluntary instrument tells you in its own objectives that adhering to it does not conclusively prove compliance, and concedes that the state of the art cannot meet the statute with a single technique. Both admissions cost the drafters something. Both sit in the document rather than in a footnote or an FAQ.

• Execution (30%): 72%. Three tiers of obligation inside a voluntary code means a signatory can adhere fully while doing only the “will” items. The detection commitment carries a carve-out letting smaller providers charge a limited fee in high-cost cases. Commitment 4 waives the AI-text disclosure entirely where human review and editorial responsibility exist. Against that, the Code documents a real alternative: organisations that comply by other means may do so, and will have those means assessed individually by national market surveillance authorities. Publishing the cost of not signing, rather than implying there is no path, is better execution than I initially credited.

Weighted Final Grade: C+ (77.85%) Calculation: (82 × 0.25) + (75 × 0.25) + (85 × 0.20) + (72 × 0.30) = 20.5 + 18.75 + 17 + 21.6 = 77.85%

This grade moved during the audit and I am showing the movement rather than hiding it. My working draft had this document at 76.15% and a flat C, scored from law-firm summaries. Then I reached the Commission’s own policy page and two things changed. The Code covers Article 50(2), (4) and (5), not two duties as I had it, which is broader scope than I credited and moved Ethical Stakes from 80 to 82. And the alternative compliance path is documented rather than absent, which moved Execution from 68 to 72. Those two corrections total 1.7 points and carry it across the C-to-C+ line. A grade that cannot move when the evidence moves is not a grade.

The Documentation Reality:

The instruments disagree here in the opposite direction from usual. The 7-Question audit gives this document 12 of 14 because it publishes, dates, explains, concedes failure modes, and builds a public verification route. ADECP gives it a C+ because ADECP weights coverage and what the commitments survive. Both readings are right, and the gap between them is the honest description: excellent construction, partial coverage.

Consider what the Code does that almost nothing in this field does. It builds a route for a person who was party to none of this. The detection commitment means a member of the public can take a suspicious file and get a result, and Measure 2.3 goes further by requiring that additional information carried in the watermark or metadata be folded into what the detection tool reports back. That is not a disclaimer, not a policy, and not a governance framework. It is a mechanism, pointed outward, and it is why this scores where it does on stakes.

Now the scope. Article 50 has four duties. The Code operationalises three, and the one it leaves alone is 50(1), telling you that you are talking to a machine. That is the duty touching people in ordinary life, in a support window or a screening interview, and it has no code, no icon set, no detection mechanism, and no signatory list. A vendor waving this Code at you on a chatbot question is answering a different exam.

Then the three-tier language, the most quietly consequential design choice in the document. Mandatory, encouraged, and optional measures share the same numbering, typography, and page. A market surveillance authority will parse that fluently. A procurement officer comparing two vendors will not, and neither will a journalist, and neither will the person whose photograph got regenerated. The Code told the truth and filed it where only specialists will find it, which is a pattern I have graded companies harshly for all year and will not exempt a regulator from now.

Limitation Statement Analysis:

Found in the objectives of both sections, as reported by Bird & Bird: adherence “does not constitute conclusive evidence of compliance.”

Assessment: Meets the Specific + Contextual + Action-Oriented formula at 67%. Specific, naming exactly what adherence does not establish. Contextual, appearing in the objectives where a reader learns what the instrument is for, which is the right place rather than an annex. Not action-oriented, because it tells nobody what to do instead, and tells a market surveillance authority nothing about what conclusive evidence would look like. Two of three elements present.

Check it yourself, which takes about four minutes: Open the Commission’s Code policy page and read the paragraph beginning “It helps providers and deployers,” which states the scope as Article 50(2), (4) and (5). Then compare it to any law-firm summary you have bookmarked. Several reputable ones say the Code covers two duties. The Commission’s own page says three, and when a secondary source disagrees with the institution that published the thing, the institution wins.

Why C+ Instead of B-:

A B- needs coverage to match construction. Three duties out of four, with the omitted one being the most commonly encountered, caps this regardless of drafting quality. The carve-outs compound it: a public detection route with a fee exception, and a text-labelling duty with a full waiver, are weaker promises than their headlines suggest.

What prevents a lower grade is the self-limiting language, the icon set, and the documented alternative path. A drafting team that writes down that its own instrument is not conclusive proof has done something I would like to see from every vendor I grade, and the Positioning score reflects it.

What Would Raise the Grade:

Publish a one-page table sorting every measure into mandatory, encouraged, and optional. The information exists inside the document already and is currently recoverable only by reading verbs. A table would let a non-lawyer compare two signatories in ninety seconds.

State on the Code’s cover page, in one sentence, which Article 50 duty it does not reach. A reader currently learns the scope by noticing an absence, and absence is the hardest thing for a non-specialist to notice.

Sources: European Commission policy page, “Code of Practice on Transparency of AI-generated Content,” last updated 31 July 2026, read in full 26 August 2026. Commission FAQ pages on the Code and on signing it. Code of Practice published 10 June 2026. Bird & Bird, June 2026. IPTC technical analysis, 10 June 2026. DLA Piper, July 2026.

• Commission policy page for the Code

• Commission FAQ on the Code

• Bird & Bird analysis

• IPTC technical analysis

Continue reading below for the F, the thirteen names missing from the register, the twelve signatories from two small countries, the complete Four Columns framework, and this week’s Signal.

Read more

On August 6, OpenAI shipped an updated GPT-5.6 Sol to Plus and Pro subscribers, moved Free and Go users to GPT-5.6 Luna as the default, and published two documents about it. I read both this weekend: the consumer announcement and the system card on the Deployment Safety Hub, and I found something truly interesting about them both.

Near the bottom of the announcement, in the sixth of seven sections, under a heading about safety and availability, the authors write that this version of Sol is optimized for everyday chats and will only appear in the Chat experience, and that the version of GPT-5.6 Sol powering Work and Codex is not changing as part of this release. That sentence is accurate and unhedged, and it sits in the same post as the marketing copy. Anyone who claims OpenAI hid this has not read the page.

The system card carries the same fact with one addition. Its authors explain that because two builds now share a product name, they will separate them inside the document by month of release: August for the models that shipped that day and July for the ones still running in Codex and Work. Read the evaluation tables, and you can see the workaround in the column headers, where GPT-5.6 Sol (July) sits beside GPT-5.6 Sol (August) in the biology refusal results, two sets of weights carrying one name, told apart by a parenthesis.

Read more

TL;DR: Intercom’s Fin AI Agent disclosure documentation receives a D (61.2%) for shipping the AI label switched off by default, contradicting itself across two of its own help centers about which customers are affected, and never naming the law it exists to satisfy. Salesforce Agentforce receives a B- (77.9%) for building disclosure that customers cannot configure away, publishing accessibility conformance reports that satisfy an overlooked clause of the same article, and still leaving a deployer without a single page explaining how to comply. Grades use the ADECP framework, which scores Ethical Stakes, Accessibility, Positioning, and Execution. This runs about 23 minutes at 250 words per minute, which is long for this newsletter and is what two full audits with primary-source evidence took. Every claim below links to a primary source so you can check my work.

FIRST, FIVE TERMS YOU NEED

Skip this block if you already have these. Monday’s piece leaned on them without defining them, so they live at the top this week.

Provider. The company that builds an AI system and puts it on the market under its own name. Intercom is the provider of Fin. Salesforce is the provider of Agentforce. Think of the provider as the manufacturer.

Deployer. The company that takes someone else’s AI system and points it at its own customers. If your business runs Fin on your support page, you are the deployer. Think of the deployer as the shop that installs the product and puts its name on the door.

Article 50(1). The clause of the EU AI Act that says people must be told they are dealing with an AI system rather than a person. It became binding on August 2, 2026. Read the text closely and you find it is written at providers, not deployers: the duty is to design and develop the system so that people are informed. That is a build-time obligation sitting with the manufacturer, which matters enormously for everything below. A great deal of trade coverage this month has described Article 50(1) as a deployer duty, and the Commission’s own guidance does not.

The obvious exception. Article 50(1) does not apply where an ordinary attentive person would already know they are talking to a machine. The Commission’s guidance says to read this exception narrowly, because every time you use it, somebody loses information they were entitled to have. It is a keyhole, not a door.

Market surveillance authority. The national regulator in each EU country that investigates and fines. Not Brussels, in most cases. The equivalent of your state consumer protection office, with authority to reach a company that has never set foot in its country.

One more piece of arithmetic before we start. Fines under Article 50 reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. That figure comes from the Commission’s own explainer, which I read end to end for Monday’s piece and read again this morning.

Source: Transparency obligations under Article 50 of the AI Act, European Commission, last updated 24 July 2026

THE AI DOCS WATCHLIST

🔴 Intercom, Fin AI Agent Disclosure Documentation (D, 61.2%)

What Users Need to Know:

Fin is a customer support agent that answers questions in chat, in email, and across WhatsApp, Instagram, Facebook, SMS, and Slack. When Fin replies inside Intercom’s own Messenger, the message card can carry a small label after the agent’s name identifying it as an AI agent. Whether that label appears is a setting, and the setting has a default, and the default is the entire story of this grade. Intercom has written a genuinely detailed help article explaining how the setting works, which is more than most vendors in this category have produced. The problem is what the article documents rather than how well it documents it.

Two facts sit next to each other in that article. The first is Intercom’s own account of its history: the product used to mark Fin’s responses automatically, and the company states plainly that it has since removed that marking so customers can decide how the AI is presented. The second is the current default. The label toggle is on for customers who were already running Fin in the Messenger before a mid-2025 cutoff, and off for everyone else. Every company that has adopted Fin in the Messenger since that cutoff started with the AI label hidden unless somebody went looking for the setting. The core problem is that a design duty was converted into a configuration option, and the option defaults to silence.

Where This Information Lives:

• The AI Agent Disclosure help article on intercom.com: Dated June 30, 2026. Contains the toggle path, the default behavior, the default introductory copy, and an unusually candid known-limitations section. It refers to AI legislation only in general terms and links out to a separate article whose title names the Act. No article number and no date appear anywhere on the page.

• A second article of nearly the same name on fin.ai: Dated March 6, 2026. Covers Fin for Zendesk, Salesforce, Freshworks, and HubSpot. Gives a different cutoff date for the same default.

• The Fin and the AI Act help article: Linked from the disclosure article as the place the obligation is explained. I did not retrieve this one, so I am not grading its contents. What I am noting is structural: the page telling you how to configure disclosure is not the page telling you why you owe it.

• Additional Product Terms for AI Product, Section 5: The contract clause making the customer responsible for informing end users. Linked from a note at the bottom of both help articles.

Visual Evidence:

A person opens the support widget on a company that adopted Fin in the last year and has not changed the settings. Under Simple deploy, the introductory message naming the agent as an AI agent is pre-populated, and where it runs, it is real disclosure that works. Whether it runs is itself less certain than it should be: the same article describes the intro as pre-populated in one passage and as something the admin toggles on in another. Then the conversation continues, and the small AI Agent label that would otherwise sit beside the agent’s name on every subsequent card is not there, because the toggle governing it shipped off. On WhatsApp, Instagram, Facebook, SMS, and Slack, the introductory message does not fire at all under the default path, because Intercom documents that this message only triggers for conversations opened inside its own Messenger on web, iOS, and Android.

The Three Defects:

Defect one: the vendor removed a compliant default and documented the removal without documenting the risk. Intercom's own words are that the company has "since removed the AI Answer label." The article gives no date for that removal, no notice that affected deployers were told, and no changelog entry a compliance team could cite. A deployer who wants to know what their end users saw in March versus what they see now cannot reconstruct it from the vendor's record.

Defect two: two Intercom-operated help centers give different cutoff dates for the same default. The intercom.com article puts the boundary at 10 July 2025. The fin.ai article states the toggle is "On by default if you deployed Fin before July 30, 2025." That is a twenty-day gap. Any company that deployed Fin between those two dates cannot determine from the vendor's documentation whether its AI label is visible to customers right now. Under a regime where the regulator asks you to demonstrate where your disclosure sits and who established that, twenty days of ambiguity is not a typo. It is an evidence gap, and it survived a June 2026 revision of the newer page.

Defect three: the page a customer would read to answer the legal question still hedges. The fin.ai article answers the question of whether disclosure is legally required by describing transparency as something that is becoming a requirement in many regions. That was defensible on March 6, 2026, when the page was dated. It stopped being accurate on August 2, 2026, when Article 50(1) started to apply, and the page has not moved since. The single most consequential sentence in the document is the one that went stale.

7-Question Audit Results:

Total: 9/14 points, which lands in the Mostly Documented band.

I want to flag that score before I go further, because it looks generous, and it is. Nine out of fourteen is the second-highest band on my scale, and I am about to give this company a D. Both readings are correct, and the tension between them is the finding. The 7-Question Audit measures whether the writing does its job, and Intercom’s writing does. The known-limitations section names the Proactive Procedure gap, the teammate-handoff gap, the pre-July-2025 workflow gap, and the five channels where the introductory message never fires. I have not audited another vendor in this category that names its own failure paths that precisely. ADECP measures whether the person affected by the system ends up informed. Those two instruments can disagree, and when they do, the second one is the one that matters to the customer in the chat window.

ADECP Category Breakdown:

• Ethical Stakes (25%): 58%. Support conversations run through billing disputes, account recovery, cancellations, and in regulated verticals through claims and payments. A person who believes a human reviewed their refund decision behaves differently from one who knows a machine issued it. Intercom does name the legal obligation and does recommend using at least one disclosure method. It also ships the silent default and disclaims responsibility for what follows, which is the combination that holds this score under 60.

• Accessibility (25%): 70%. The writing is strong. Screenshots, exact settings paths, the default copy quoted in full, and step-by-step instructions for retrofitting a disclosure message into an existing workflow. Points come off for the contradiction between the two help centers, which leaves a specific set of readers unable to establish their own state, and for the stale hedge about the law.

• Positioning (20%): 62%. Limitation language is abundant, which is why this is not lower. The order is the problem. The article opens with control, flexibility, and aligning disclosure with your communication strategy, and closes with the compliance note in a box at the bottom. A reader skimming for the decision gets the marketing framing first and the legal framing last.

• Execution (30%): 56%. The gap between stated intent and shipped behavior. Off by default: a previously compliant default removed without a dated record, two documents that disagree, and no mention anywhere in the disclosure article of the article of law it addresses. The contract term is the sharpest item. Intercom’s terms make the customer responsible for informing end users, and both help articles end with a note that Intercom is not responsible for what follows if a customer disables both disclosure methods. A commercial agreement can allocate cost and blame between two companies. It cannot move a build-time obligation that the Act places on the party doing the building, and a market surveillance authority reading Article 50(1) will be looking at who designed the system.

Weighted Final Grade: D (61.2%). Calculation: (58 × 0.25) + (70 × 0.25) + (62 × 0.20) + (56 × 0.30) = 14.5 + 17.5 + 12.4 + 16.8 = 61.2%. On my scale, 60 through 63 is a D. It does not round into D+ territory, which begins at 64, and it does not fall to D-, which ends at 59. The grade holds without needing a favorable rounding.

Limitation Statement Analysis:

The strongest disclosure language Intercom ships is the default introductory message, which tells the customer they are speaking with an AI agent and can ask for the team at any time. Assessed against the Limitation Statement Formula, it meets all three elements. It is specific, because it names the agent as artificial rather than gesturing at automation. It is contextual, because it arrives at the top of the conversation rather than in a policy. It is action-oriented, because it tells the person exactly what to say to reach a human. That copy is good. The reason it does not lift the grade is that Intercom documents four separate paths on which it never appears: Proactive Procedures, agent-triggered handoffs from the inbox, workflows built before the mid-2025 cutoff, and every channel outside its own Messenger.

Why D Instead of C:

A C would mean the documentation is performing transparency without delivering it. That is not what happened here. Intercom delivered transparency, wrote it down carefully, and then made it optional and turned the option off. The distance between a C and a D in this case is the A/B result the company published in its own help center: Intercom reports that adding the introductory disclosure did not meaningfully change Fin’s resolution rate and did raise customer satisfaction scores. The vendor has measured the cost of disclosure and found it to be approximately zero, with a measurable benefit attached. When a company holds evidence that the honest default is free and ships the quiet one anyway, the gap stops being a resourcing problem and becomes a decision.

Why D Instead of D-:

The known-limitations section is real work, and it earns real credit. Naming the exact flows where your own safeguard fails is the behavior I have spent seven months asking companies to adopt. Intercom did it voluntarily and in detail. That is worth two full points on the 7-Question Audit, and it is why the Execution score is 56 rather than the low 40s.

What Would Raise the Grade:

1. Set the AI Agent label toggle on by default for all workspaces, and publish a dated changelog entry for the change. Your own testing says it costs you nothing.

2. Reconcile the two help centers to a single cutoff date, or state plainly that customers who deployed between July 10 and July 30, 2025 should check the setting directly rather than rely on either page.

3. Add one sentence to the disclosure article naming Article 50(1) of Regulation (EU) 2024/1689 and the date it took effect, then replace the language about transparency becoming a legal requirement with a statement that it is one.

Sources: AI Agent Disclosure, Intercom Help, June 30, 2026, What is AI Agent disclosure?, Fin Help Center, March 6, 2026, Article 50, Regulation (EU) 2024/1689, Transparency obligations under Article 50, European Commission

Check it yourself: open your own Fin workspace and find the Show AI Agent label in the Messenger toggle. The intercom.com article routes you through Fin AI Agent, then Fin settings, then Identity. The fin.ai article routes you through Settings, then Reply Settings, then Identity. Those are two different paths to one switch, which is the same divergence showing up a second way. Whatever the toggle says is what your customers have been seeing. It takes about ninety seconds, and it does not require my analysis to be right.

Continue below for Salesforce Agentforce’s grade and the clause of Article 50 almost nobody is scoring, four items I am watching this week, this week’s Signal on why defaults are the whole ballgame, and the complete Disclosure Default Ladder.

Read more

Every claim below links to a primary source so you can check my writing.

Article 50 of the EU AI Act became enforceable on August 2. The part that reaches almost everyone reading this is its first paragraph: if you build a system people talk to directly, you have to tell them they are talking to a machine. Manner and timing come from a different paragraph, Article 50(5), which requires the notice to be clear, distinguishable, compliant with accessibility requirements, and delivered at the latest at the first interaction. The provision carries no penalty of its own, which is worth knowing before you quote a number to your leadership. The fine sits in Article 99(4)(g), up to 15 million euros or three percent of worldwide turnover, whichever is higher, though Article 99(6) reverses that calculation for SMEs and start-ups to whichever is lower.

I sat down last week to grade two companies on that first paragraph. Before opening a product, I read the rule end to end, because secondary summaries have burned me often enough that I stopped trusting them. That reading turned up something upstream of any product I could grade. The Commission’s own FAQ on the Digital Omnibus tells readers the amendment suggests a transition period of six months for providers who need to retrofit their generative systems so the outputs become detectable. Recital 38 of the adopted regulation sets that period at four months. The legislature shortened it before adoption, and the FAQ still carries the proposal figure.

The four-month transition runs to December 2, 2026, roughly fifteen weeks from today. A provider who trusts the FAQ and counts six months forward from August 2 lands on February 2027 and concludes there is no rush this quarter, wrong by two months on a deadline still close enough to hit. Let me size the claim before showing the rest of the evidence. None of this changes what the law requires, since the Regulation says what it says. What it changes is who can find out in time.

HOW A READER ENDS UP WITH THE WRONG DATE

The route matters, because nobody goes looking for a Digital Omnibus FAQ on purpose. Start where a compliance lead starts, at the Article 50 text on the AI Act Service Desk, the Commission’s own reader for the Act. Above the provision runs a disclaimer stating that the Digital Omnibus on AI amended it and that the text shown “has not yet been updated to reflect those amendments.” It directs readers to the site’s FAQ under Digital Omnibus. I followed that route this morning. The FAQ answers questions about the Commission’s proposal, which is what it was written to do, so I will not fault it for the tense. The problem is that it never names Article 50 anywhere on the page. The only line that touches the provision describes the marking transition for generative systems without citing the article, and that is the line carrying the six-month figure.

Here is what the Official Journal says instead. Regulation (EU) 2026/1744 was published on July 24, 2026 and entered into force on July 27. It moved the core Annex III high-risk obligations to December 2, 2027. It did amend Article 50, though only paragraph 7, which governs who encourages codes of practice and how their adequacy gets assessed. The duties in paragraphs 1 through 5, the ones a product team acts on, were left alone. The single concession is that four-month transition for the machine-readable marking duty in Article 50(2), open only to systems already on the EU market before August 2.

Give the Commission its due, because the disclaimer is truthful. Article 50 was amended, and warning readers that a rendering may lag behind the amending act is more than most publishers do. The failure sits in the destination rather than the warning. A pointer is only as good as what it points to, and this one delivers a retired number and never mentions the article the reader arrived from. The same maintenance pattern shows on the guidelines landing page, where the link labeled “public consultation” holds a sentence fragment in its destination field instead of a web address.

Check it yourself, which takes about four minutes:

• Article 50 text and the Digital Omnibus disclaimer: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50

• The Digital Omnibus FAQ the disclaimer points to: https://ai-act-service-desk.ec.europa.eu/en/faq?faq_category_id=99

• Regulation (EU) 2026/1744, recital 38 for the four months: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng

• Commission questions and answers on Article 50: https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act

• Article 99, where the penalties actually live: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99

WHY THE SUB-ARTICLE NUMBER IS NOT A DETAIL

The second problem is a citation error, and it compounds the first. On the Commission’s Article 50 questions and answers, under the heading explaining what applies to providers, the duty to design interactive systems so people are told they are dealing with AI is attributed to Article 50(2). Further down the same document, the identical duty is attributed to Article 50(1), correctly. One source, one duty, two sub-article numbers, carrying a last-update stamp of July 24, 2026.

Sub-article numbers carry different compliance routes and clocks, which turns a citation slip into a planning error. For the marking and labelling duties there is an approved Code of Practice on Transparency of AI-generated Content, which the Commission describes as covering Article 50(2), (4) and (5) and which it and the AI Board have both assessed as adequate. Signatories get a route the Commission has already blessed, while providers who decline must show adequate alternative means. Worth stating plainly, since vendors overstate it: recital 41 of the amending regulation records that these codes have limited legal effect and do not grant a presumption of conformity. For Article 50(1) there is no code at all, and providers determine adequate measures themselves while taking the guidelines into account.

Put the two findings together and follow what a reasonable person does with them. A compliance lead sees the interactive disclosure duty labeled 50(2), then learns from the linked FAQ that this sub-article gets six months. Both signals say the chatbot notice can wait until next year. The accurate reading is that Article 50(1) has no code, no transition, and has been enforceable since August 2, while the transition that does exist belongs to a different duty and expires in December.

I should argue against myself, because the honest version of this piece is not a claim that Europe cannot explain its own rules. The substance across these documents is largely accurate, and the questions and answers is one of the better plain-language explainers in European digital law. Its four cumulative criteria for what counts as a directly interactive system are genuinely useful, and I will lean on them Wednesday. The Commission also does something I rarely see a regulator do: it instructs readers to read the exception for obvious interactions restrictively, on the reasoning that the exception deprives people of transparency. The gap I am describing is narrow and sits in one layer.

WHO ACTUALLY PAYS FOR THIS

Read more

Building in Public | August 14, 2026 | About a 5-minute read

Read more

TL;DR: Anthropic published a support page on Monday confirming that Claude models launched on or after August 2 weave an invisible watermark into every line of text they produce, and I am grading that page a C (73%) on the ADECP benchmark because the section explaining how to read that watermark is sixty words long and contains no method. OpenAI, working from the identical legal trigger, earns a B+ (85%) for shipping a public verification tool, extending it to audio on July 31, and opening API access, while conceding it has no text marking at all. This issue introduces The Provenance Chain, a four-link test that grades a transparency commitment at its weakest link rather than its loudest announcement. Every claim below links to the company’s own words or the European Commission’s own filing, and I read all four primary documents end to end this morning rather than working from the news coverage. This runs about sixteen minutes at a normal reading pace, which is four times my usual length, and I am running it long on purpose because two complete audits with primary-source verification do not compress below that.

FIRST, THE VOCABULARY

Skip this block if you already have these terms. I include it because a good half of this newsletter’s readers are founders and brand leads who inherited an AI compliance question they never asked for, and I would rather define the words than assume them.

Article 50 is the transparency section of the EU AI Act, carrying four duties: tell people when they are talking to an AI, mark AI-generated output so machines can detect it, disclose emotion recognition and biometric sorting, and label deepfakes and AI-written text published on matters of public interest. It took effect on August 2, 2026, and it applies by function rather than by risk tier, so a company with zero high-risk systems can still be fully in scope because it runs a support chatbot. Penalties reach fifteen million euros or three percent of worldwide turnover, whichever is larger.

A machine-readable mark is a signal buried in generated content that software can find and a person cannot see, and there are two families of them. A watermark lives inside the content itself, altering word choices or pixel values in ways that survive copying and some editing. C2PA provenance metadata is a cryptographically signed file attached alongside the content, carrying far more detail but disappearing the moment someone screenshots the file or converts the format.

Provider and deployer are the two roles Article 50 assigns, and they are not interchangeable. A provider builds the model or system and places it on the market, while a deployer uses that system under its own authority for professional purposes, which includes any company calling an API and showing the output to customers. If you build on someone else’s model, you are a deployer, and your vendor’s compliance work does not cover your obligations. That distinction is where most of this month’s confusion lives.

The Code of Practice on Transparency of AI-Generated Content is the Commission’s voluntary route to demonstrating compliance with the marking and labelling duties. Section 1 covers providers and marking, Section 2 covers deployers and labelling, and signing either one is optional. The underlying legal duty is not, and several vendors have been blurring that distinction in sales conversations this month.

Primary sources for everything above:

• https://artificialintelligenceact.eu/article/50/

• https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act

• https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

THE AI DOCS WATCHLIST

🔴 Anthropic, “How Claude marks AI-generated content” (C)

What users need to know: Every response a new Claude model writes now carries an imperceptible watermark stitched into the word choices themselves, and that signal travels with the text when you paste it into an email, a document, or a submission portal. Files Claude produces in .svg, .png, and .jpg formats carry signed C2PA metadata. The marking is global rather than European, applies at the model level across Claude, the API, Claude Code, Claude Cowork, and Claude Tag, and reaches Claude accessed through AWS, Google Cloud, and Microsoft Foundry. Anthropic states plainly that a detected signature only indicates Claude may have processed content, and that its absence proves nothing at all. What a reader cannot do, as of the morning I am publishing this, is check any of it.

Where this information lives:

• Filing location: the “Privacy and legal” collection of the Claude Help Center, sitting beside articles on copyright complaints and government access. It is absent from “Features and capabilities.”

• Discovery path: no in-product surface. Nothing in the Claude interface, at the moment a response appears, indicates that the response now carries a signature.

• Outbound links in the article body: zero. The page names Article 50(2), names the Code of Practice, and tells builders to independently assess their own obligations, without linking to the Article, the Code, or the Commission’s guidelines.

• Version signal: “Updated this week.” No date, no changelog, no version history.

Visual evidence: A reader arriving at this page sees eight sections. The one titled “Detecting Claude’s marks” runs sixty words across four sentences. The one titled “Limitations” runs two hundred eighteen words and enumerates seven distinct conditions under which the signature will mislead you. That ratio is roughly three and a half to one against the capability, and I want to be careful about how I read it, because a long limitations section is a good thing and I have spent seven months asking companies to write them. The problem sits inside the short section rather than in the length of the long one.

The documentation reality: Here is the sentence that decided this grade. Under its commitments, Anthropic writes: “We’ll support users and other third parties to detect Claude’s marks, as the Code requires, and we’ll share details in forthcoming documentation.” The company names the obligation, attributes it correctly to the Code, and defers the means of satisfying it inside the same sentence. Three separate capabilities on this page sit in the future tense: detection for users, marking for models released before August 2, and the technical guidance that would explain either one. None of the three carries a date. A teacher, an editor, or a hiring manager therefore leaves this page knowing a signature exists in the text on their screen and holding no instrument for finding it.

Limitation statement analysis: Against my Specific plus Contextual plus Action-Oriented formula, this page scores two out of three, and it scores those two better than almost anything I have graded this year. On specificity, the false-negative list names five concrete causes: an older model, heavy paraphrasing or translation, a passage too short to carry a reliable signal, metadata stripped by format conversion or screenshots, and an unsupported product surface. On context, the false-positive explanation is unusually honest, conceding that people use Claude to proofread and translate work that originated elsewhere and that the output carries the signature regardless. On action there is nothing at all. A limitation statement that tells you exactly how an instrument fails, while withholding the instrument, leaves the reader more aware and no more capable.

Why C instead of B-: Ethical Stakes scored 74, because Anthropic actively warns against over-reading a hit, which meaningfully lowers the risk of someone being accused on a false positive. Accessibility scored 78, since the prose is genuinely plain and C2PA gets defined rather than name-dropped, though the filing location and the total absence of outbound links cost it. Positioning scored 88, the highest single category I have awarded a failing entry all year, because limitations occupy their own heading rather than a footnote. Execution scored 58 and carries the heaviest weight at thirty percent, because the user-facing capability the page exists to describe has not shipped. The unrounded weighted result is 73.0, one point below C+. I want to name the counterargument rather than bury it: the transition period runs to December 2, 2026 for systems already on the market, so it is arguable that detection tooling is not yet contractually due. The grade holds either way, because ADECP scores what a reader can do with a document on the day they read it, and I apply that rule to every company on this list.

What would raise the grade: Publish the detection endpoint or verification page and link it from this article, which moves Execution from 58 to roughly 85 and the overall result to B+. Attach a date to each future-tense commitment, even a quarter. Add outbound links to Article 50 and the Commission guidelines inside the “If you build with Claude” section, since that section assigns deployers a research duty and supplies no starting point. Move the article out of “Privacy and legal” and surface a one-line disclosure in the product itself.

Sources:

• https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content

• https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content

Check it yourself, which takes about four minutes: Open the Claude support article and find the section headed “Detecting Claude’s marks.” Read the four sentences it contains and note that none of them tells you where to go. Then read “Limitations” directly beneath it and count the seven failure conditions. The care in this document is real, and you can see exactly where that care was spent.

The rest of this issue is for paid subscribers. OpenAI’s B+ breakdown, the four-link Provenance Chain framework, the signatory-list finding that contradicts some of this week’s coverage, and this week’s Signal continue below.

Read more

TL;DR: xAI receives an F (30.8%) because the only document confirming that Grok output carries provenance marks is the policy that punishes you for removing them, and OpenAI receives a B- (79.0%) because it built a public verification tool and then made visible disclosure something users have to ask for. Grades use the ADECP framework: Ethical Stakes 25 percent, Accessibility 25 percent, Positioning 20 percent, Execution 30 percent. Every claim links to a primary source so you can check my work.

PLAIN LANGUAGE FIRST

Skip this block if you already have these terms. Everything below assumes them.

• Provenance: the origin record of a file, meaning what made it and what has happened to it since. Think of the history report on a used car.

• C2PA, or Content Credentials: an open standard that attaches a signed origin record to a file’s metadata. Detailed, and it falls off easily, the way a paper tag falls off a suitcase.

• SynthID: an invisible watermark woven into the pixels or the audio itself. Survives cropping and compression far better, carries much less detail, closer to a laundry tag sewn into a shirt than to a receipt.

• Machine-readable marking means a detection tool can read the signal. Human-perceivable labeling means a person looking at the screen can see it. Two different obligations, and the gap between them is this whole issue.

• Article 50: the transparency section of the EU AI Act, applied August 2, 2026, to any provider or deployer whose AI reaches people in the EU, wherever the company sits. Penalties reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.

• Provider builds and ships the system. Deployer puts it in front of people or publishes what it makes. One product often involves both, carrying different duties.

Sources: Commission guidelines on Article 50, Cooley client alert, August 3, 2026

THE AI DOCS WATCHLIST

🔴 xAI, Grok Imagine image and video generation (F, 30.8%)

What Users Need to Know:

xAI’s Acceptable Use Policy, effective June 26, 2026, prohibits “Stripping, altering or circumventing embedded provenance metadata or watermarks.” That line is a confession, because you cannot strip something that is not there. Somewhere inside the company, a team knows the marks exist and wrote a rule protecting them. The rule is the only place they wrote it down. A developer who reads the Grok Imagine documentation end to end will finish with no idea that provenance signals exist at all.

The changelog is what moves this from oversight to choice. xAI has published release notes since November 2024 with entries in most months, and across those twenty-one months not one entry mentions marking, provenance, or disclosure. On April 30, 2026, they announced a cost_in_usd_ticks field on every API response, working across chat completions, the Responses API, image generation, video generation, and streaming. They can ship a per-request field across every modality in one release when the thing being tracked is money.

Where This Information Lives:

• Acceptable Use Policy, June 26, 2026: Confirms marks exist. Names no standard, no modality coverage, no way to check. Separately prohibits “Misleading others or not being transparent regarding your use of AI,” assigning you a duty in the same document that withholds the means to meet it.

• Image Generation docs, updated July 13, 2026: 14 aspect ratio values, two resolutions, batch generation, base64 output, a respect_moderation field. Zero words on provenance.

• Video Generation docs, updated August 1, 2026, one day before the deadline: Five request modes, five error codes with a “What to do” column, four job statuses, three resolutions, seven aspect ratios, a one-to fifteen-second range, preset voices. Zero words on provenance.

• Release Notes, updated July 31, 2026: The July 31 entry ships reference-to-video and image-to-video at native 1080p with optional voices. The July 17 entry announces Grok 4.5 for EU users, so they do write EU-framed notes. No provenance entry exists in the file.

• Commission signatory list, updated August 3, 2026: xAI appears in neither section of the Code of Practice. Anthropic, Cohere, Google, Meta, Microsoft, Mistral, OpenAI, Black Forest Labs, Getty Images, and Synthesia all appear in Section 1.

Visual Evidence: The response object printed on the Video Generation page holds four fields: a temporary URL, a duration, a respect_moderation boolean, and the model name. No provenance field, no note beneath the block, no link out. The page closes with eight Related links, none about marking.

Check it yourself, about five minutes: open the two doc pages and the release notes linked below, then run find on “watermark,” “provenance,” and “C2PA.” Do the same on the Acceptable Use Policy and watch where the single hit lands.

Total: 5/14, Documentation Theater band. Q7 earning full marks is what makes the rest worse: the announcement channel exists, runs on cadence, and gets used for voice models and billing fields. Nobody had to invent anything or request engineering time. Somebody only had to decide the entry was worth writing.

Interface Quality Score: 0% (0 of 8)

1. ❌ Docs state that outputs carry a provenance signal

2. ❌ The standard is named

3. ❌ Coverage by modality is stated

4. ❌ A free public way to verify is provided

5. ❌ The verification method’s limits are stated

6. ❌ A human-perceivable label applies by default

7. ❌ A dated change record for provenance exists

8. ❌ A path to report or contest a result exists

These eight criteria measure the provenance surface and nothing else. xAI’s documentation is well-built almost everywhere I looked.

ADECP Category Breakdown:

• Ethical Stakes (25%): 30% Image-to-video and reference-to-video both accept user-supplied reference images, alongside voice cloning from a short clip. The docs never say those inputs may depict real people, and never say they may not. The June 10 release note matters most: you can now persist an Imagine output and publish it as a permanent unauthenticated public URL in one round trip. That path runs from generation to public distribution with no disclosure guidance on it, and public distribution is the moment Article 50 governs.

• Accessibility (25%): 42% Scored generously and I want that on the record. The docs are clear, well-organized, and written for a working engineer, with runnable examples in five languages. The only confirmation that marks exist sits in a legal policy nobody reads before an integration, which is textbook misfiling: the honest sentence was written, then filed where no user lands.

• Positioning (20%): 22% Capability documentation is exhaustive and maintained. Limitation documentation on marking is absent rather than thin, and those are different findings. Widest capability-to-limitation gap I have measured in thirty-two issues.

• Execution (30%): 28% Legal shipped a provenance clause June 26. Docs touched the image page July 13 and the video page August 1 and shipped nothing. The changelog ran normal cadence straight through the deadline and carried an EU-specific entry two weeks before it. Two teams held two halves of one fact, and only the half protecting the company got published.

Weighted Final Grade: F (30.8%) (30 × 0.25) + (42 × 0.25) + (22 × 0.20) + (28 × 0.30) = 7.5 + 10.5 + 4.4 + 8.4 = 30.8%

I show the unrounded number because a grade you cannot recompute is only an opinion wearing a percentage. Run the generosity test: raise Accessibility from 42 to 60, more credit than the evidence supports, and you reach 35.3 percent. The letter does not move.

Limitation Statement Analysis: “Stripping, altering or circumventing embedded provenance metadata or watermarks.” This fails the Specific plus Contextual plus Action-Oriented formula at roughly 33 percent. Specific about the prohibited act, silent on which outputs carry marks under which standard, and action-oriented only in the negative. A statement that works only as a prohibition is a liability disclaimer, and disclaimers protect the company while explanations serve the user.

Why F Instead of D: A D means the disclosure exists and reaches the wrong audience. Here it exists only as an inference drawn from a prohibition, inside a document written to constrain the reader rather than inform them. No page, field, footnote, or changelog entry lets a developer find it by looking.

The strongest counterargument deserves full strength. xAI may well be embedding conformant C2PA manifests and durable watermarks on every output right now, and the policy language suggests something is there. ADECP grades documentation rather than technology, and that distinction matters most here. A developer who shipped Grok Imagine output into an EU-facing product this week is a deployer carrying a legal duty, and excellent invisible engineering does not help that person answer a market surveillance authority. Undocumented compliance and absent compliance look identical at the moment somebody asks. I expected to find a stub page given the attention August 2 received, checked five surfaces, and was wrong about what I would find.

What Would Raise the Grade:

• Add a “Provenance and disclosure” section above Configuration on both generation pages. Four sentences: name the standard, state which modalities are covered and which are not, state whether the mark survives the temporary URL download path, link to verification.

• Expose a provenance field beside respect_moderation. They proved on April 30 they can add a field to every response across every modality.

• Publish one dated release note. The changelog is live, maintained, and read by the exact audience that needs this.

Sources: Acceptable Use Policy, June 26, 2026, Image Generation, July 13, 2026, Video Generation, August 1, 2026, Release Notes, July 31, 2026, Commission signatory list, August 3, 2026

OpenAI’s grade and evidence, The Signal on what 82 signatures actually bought, and this week’s framework, The Provenance Disclosure Ladder, continue below.

Read more

Article 50 of the EU AI Act became applicable on Sunday, August 2, 2026. It is one of the shortest chapters in the regulation and the only one an ordinary person will ever feel directly, because every duty it creates is discharged in front of a user rather than inside a compliance file that nobody outside the company will read.

I have spent since January grading AI companies on where their limitation language sits and whether a non-technical reader can act on it. Until Sunday, that was an editorial standard I invented, defended in public every week, and argued about with readers who thought I graded too hard. As of Sunday, a narrower version of the same idea is enforceable law across twenty-seven countries. The EU did not adopt my framework, and I want to be precise about that. What it adopted is the piece I have had the hardest time getting companies to take seriously, which is the requirement that an explanation arrive at a specific moment rather than merely exist somewhere on the site.

Article 50(5) is the sentence worth reading twice. The information owed under the article must be given in a “clear and distinguishable” manner, no later than the person’s first interaction with the system or first exposure to the content, and it must conform to applicable accessibility requirements. Three separate obligations sit inside that one sentence. The disclosure has to exist, it has to arrive before the user acts, and it has to be usable by the people it is aimed at, including people who reach it through assistive technology. A line buried in the terms of service satisfies none of the three.

Check it yourself, which takes about four minutes:

• Article 50, official text: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50

• Commission guidelines on Article 50, adopted July 20, 2026: https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems

WHAT CAME DUE, AND WHAT QUIETLY MOVED

Four duties started applying on Sunday. Providers of systems that interact directly with people have to design them so the person knows they are dealing with a machine, unless that is already obvious to a reasonably observant person. Providers of systems that generate synthetic audio, image, video, or text have to mark the outputs in a machine-readable format so they are detectable as artificially generated. Deployers of emotion recognition and biometric categorization systems have to tell the people exposed to them that the system is running. Deployers have to disclose deepfakes, though for evidently artistic, creative, satirical, or fictional work that disclosure is limited so it does not spoil the work itself. Deployers also have to disclose AI-generated text published to inform the public on matters of public interest, unless the text went through human review and a person or organization holds editorial responsibility for it.

Five days before that happened, the deadline most compliance teams had been planning around for two years moved. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on July 24 and entered into force on July 27. It pushed the obligations for standalone high-risk systems under Annex III, which covers hiring, credit, education, essential services, law enforcement, and migration, from August 2, 2026 to December 2, 2027. High-risk AI embedded in regulated products under Annex I moved to August 2, 2028. Anyone still writing that high-risk enforcement landed this weekend is working from a calendar that expired six days before the date on it.

The transparency duties did not move with them, and that asymmetry is the whole story of this week. One narrow carve-out exists: generative systems already on the market before August 2 have until December 2, 2026 to meet the machine-readable marking requirement in Article 50(2). Systems placed on the market from August 2 forward get no such window. Breaching Article 50 carries administrative fines of up to 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher, under Article 99(4)(g). Article 50 does not stop at the EU border. Under Article 2, the regulation catches providers placing a system on the Union market wherever they are established, and it catches providers and deployers sitting in a third country when the output of their system is used in the Union. A US product team with European users is inside this. A US product team whose output never reaches the Union is not.

So the regime, built from risk management systems, conformity assessments, technical files, and database registration, slid by sixteen months. The regime, built entirely from what a user can see, read, and understand in the first second of contact, took effect on schedule. For everyone who has been told that documentation work can wait until the real compliance deadline arrives, the real compliance deadline arrived first, and it is made of documentation.

Check it yourself:

• Regulation (EU) 2026/1744, official text in the Official Journal: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng

• Plain-language summary of the revised dates: https://www.hunton.com/privacy-and-cybersecurity-law-blog/eu-digital-omnibus-on-ai-enters-into-force

• Article 99 penalties, official text: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99

Read more

TL;DR: Midjourney receives an F (47.8%) for shipping unmarked synthetic media four days before EU law requires machine-readable marking, while assigning the anti-deception duty to its users. Google Gemini receives a B (81.2%) for building SynthID watermarking into every output and documenting its failure modes, then spending a third of its signing announcement arguing the labels might confuse people. Grades use the ADECP framework (Ethical Stakes, Accessibility, Positioning, Execution); this issue introduces the EU AI Act Documentation Readiness Checklist, and every claim links to a primary source you can open yourself.

FIRST, FOUR TERMS YOU NEED

Skip this block if you already work in provenance. Everyone else, these four definitions carry the whole issue.

Article 50 is the transparency section of the EU AI Act, and it applies from this Sunday, August 2, 2026. It does not inspect how your model was built. It requires that people be told when AI is involved: chatbots must identify themselves, synthetic media must be marked, and deepfakes must be labeled. Think of it as the ingredient label rule, not the recipe inspection.

Machine-readable marking is a signal placed inside the file itself, readable by software even when no human label is visible. A watermark like Google’s SynthID is woven into the pixels and survives screenshots and compression. It works like a receipt that travels inside the image.

C2PA Content Credentials are cryptographically signed metadata attached to a file, recording what made it and what edited it. Picture a passport with stamps: rich detail about the journey, but the passport can be lost if a platform strips the metadata. Watermarks and Content Credentials are designed to back each other up.

The Code of Practice on Transparency of AI-Generated Content is the voluntary playbook the European Commission blessed as an adequate route to Article 50 compliance on July 9, 2026. Signing it earns focused, predictable enforcement. It is the study guide written by the people who wrote the exam.

THE AI DOCS WATCHLIST

🔴 Midjourney (F, 47.8%)

What Users Need to Know:

Midjourney’s Community Guidelines, which are binding terms of the service agreement, tell every user: “You may not intentionally mislead recipients of the Assets about their nature or source.” That is a real duty with real consequences, because rule 10 of the same guidelines reserves the right to ban you for violating it. Midjourney then hands you an image that carries no watermark, no C2PA manifest, and no machine-readable trace of its origin of any kind. Independent provenance trackers confirm this remains true through version 8 as of this month, and Midjourney has been a Content Authenticity Initiative member since 2023 without shipping an implementation. The company assigns you the duty of honesty about your content’s source and supplies you no means to demonstrate it.

This is the Means-and-Duty Matrix from Issue 30 in its purest form I have found yet. Four days from now, Article 50(2) makes machine-readable marking a legal requirement for providers placing generative systems on the EU market, with fines of up to 15 million euros or 3% of worldwide annual turnover. Midjourney revised its Terms of Service on May 27, 2026, well into the Code of Practice drafting process, and added no marking commitment, no Article 50 language, and no EU transparency provision of any kind. The revision date is evidence that they touched the document and skipped the obligation.

Where This Information Lives:

• midjourney.com (main site): Product galleries and subscription plans. No limitation language, no marking disclosure, no AI-behavior explanation of any kind.

• Terms of Service (docs.midjourney.com, effective May 27, 2026): The only limitation statement in the entire corpus, filed under Section 2, “Age Requirements.”

• Community Guidelines (Section 9 of the Terms): The anti-deception duty assigned to users, at rule 7.

• Midjourney Policies hub: Contains an “AB2013 Documentation” page satisfying California’s training data transparency law. Nothing equivalent exists for the EU obligation arriving Sunday.

That last bullet matters. The AB2013 page proves Midjourney can ship a compliance document when a statute names one. The absence of anything for Article 50 is therefore a choice, not an oversight.

Visual Evidence:

A user generating an image on midjourney.com or through Discord sees the prompt bar, the generation progress, and the finished image with download options. No disclosure about marking appears at generation, and no provenance indicator appears on the output. Downloading the file and inspecting it with a metadata tool, which is how third-party trackers verified this, shows no C2PA manifest and no known invisible watermark.

Total: 5/14 points - Documentation Theater

Interface Quality Score: 13% (1 out of 8 interface criteria met)

1. ❌ AI behavior disclosure at first use: No (no explanation of system behavior or limits at signup or first generation)

2. ❌ Limitation language visible without navigation: No (the single limitation statement lives inside the Terms of Service)

3. ❌ Provenance marking on outputs: No (no C2PA manifest, no known invisible watermark, confirmed through version 8)

4. ❌ User-accessible verification of marking: No (nothing to verify)

5. ❌ Failure modes named in-product or one click away: No (no failure mode content exists anywhere in the corpus)

6. ❌ Error or feedback mechanism on AI outputs: No (the documented reporting path is a DMCA copyright takedown, not an AI-output error channel)

7. ✅ Change log accessible: Yes (model version notes exist on the docs site)

8. ❌ Data use explained within the product surface: No (the Terms point to a separate privacy policy)

Midjourney met one criterion: version release notes. Everything a user would need to calibrate trust or satisfy their own disclosure duty is absent.

ADECP Category Breakdown:

• Ethical Stakes (25%): 50% - Unmarked synthetic media at Midjourney’s output volume is precisely the risk Article 50 names: content that can be presented as real with no technical trace saying otherwise. The Terms assign the duty of deception detection to users while providing no means of marking, which converts an individual honesty problem into a structural one. I want to argue against myself on the record here: a strict keyword pass of this category would score Midjourney near 78%, because image generation does not trip the health, finance, or employment indicators the automated auditor looks for. I overrode that to 50% because grading a synthetic media provider as “general purpose” four days before a synthetic media law takes effect would be applying the rules mechanically instead of honestly, and I am naming the override so you can disagree with it.

• Accessibility (25%): 55% - The one honest sentence Midjourney publishes (”This is new technology, and it does not always work as expected. No guarantees are made as to the suitability of the Assets”) is written in genuinely plain language, which earns points. It is also filed inside Section 2 of the Terms of Service, under the heading “Age Requirements,” where no user deciding whether to trust an output will ever land. This is misfiling: the honest sentence exists, in the one place it cannot do its job.

• Positioning (20%): 40% - The main site promotes capabilities with zero limitation weight. Every constraint that exists lives in legal documents, and the “AS IS” warranty disclaimers in Section 1 protect the company rather than inform the user. Counting them separately, I found several liability disclaimers and exactly one explanation, and the explanation is the misfiled sentence above.

• Execution (30%): 45% - No marking shipped, three years into Content Authenticity Initiative membership. A May 27, 2026 Terms revision that added nothing for the August deadline. Points earned: the DMCA process is genuinely detailed, model release notes are maintained, and the AB2013 page shows the compliance muscle exists when a law forces a specific document.

Weighted Final Grade: F (47.8%). Calculation: (50 × 0.25) + (55 × 0.25) + (40 × 0.20) + (45 × 0.30) = 12.5 + 13.75 + 8 + 13.5 = 47.75, which rounds to 47.8% → F.

For scale: Canva’s D- in Issue 30 scored 56.1%. Midjourney lands more than eight points below that, and the gap is almost entirely the marking absence, because marking is the one control that would have satisfied Execution, Positioning, and the incoming law at the same time.

The Documentation Reality:

Midjourney is not confused about disclosure. Its guidelines demand it, from you. Rule 7 makes misleading anyone about an asset’s “nature or source” a bannable offense, which means the company has already decided that provenance honesty matters enough to enforce. It has simply decided to enforce it on the only party in the transaction who cannot embed a watermark.

The Terms of Service tell the same story from a different angle. Section 1 states that the Services “are subject to modification and change” and asks users to “not create any dependencies on any attributes of the Services.” That is the opposite of the proactive change communication Article 50-era compliance assumes, and it is a written commitment to unpredictability in a document revised nine weeks before the deadline. When a company updates its legal terms during the exact window the Commission was finalizing the marking playbook, and the update contains no marking language, the silence is the disclosure.

Check it yourself, which takes about five minutes:

• Terms of Service (Section 2 for the limitation sentence, Section 9 rule 7 for the user duty): https://docs.midjourney.com/hc/en-us/articles/32083055291277-Terms-of-Service

• Midjourney Policies hub (note the AB2013 page and the absence of any EU transparency page): https://docs.midjourney.com/hc/en-us/sections/33329694407821-Midjourney-Policies

• Third-party marking verification (July 2026): https://www.lumethic.com/en/articles/ai-generators-c2pa-watermarks

Limitation Statement Analysis:

Found in Terms of Service, Section 2 (Age Requirements): “Midjourney tries to make its Services PG-13 and family friendly, but the Assets are generated by an artificial intelligence system based on user queries. This is new technology, and it does not always work as expected. No guarantees are made as to the suitability of the Assets for the Customer.”

Assessment: Fails the Specific + Contextual + Action-Oriented formula. It is not specific (no named failure mode, no example of what “not as expected” means). It is not contextual (it appears in the age policy, not at generation or download). It is not action-oriented (it tells the user nothing to do differently). It is a liability disclaimer wearing a limitation statement’s clothes.

Why F Instead of D-:

The D range requires that a user, with effort, can find enough to calibrate trust. Midjourney’s user cannot, because the material does not exist: no failure modes, no marking, no pre-use explanation, no AI-output error channel. What keeps this from being a low F rather than a borderline one is real: the plain-language register of the Terms, the maintained release notes, and the California compliance page all show functioning documentation capacity. Capacity without application, four days before application becomes law, is what an F measures.

What Would Raise the Grade:

Ship C2PA Content Credentials at the point of generation, which peers including Adobe Firefly and OpenAI’s image tools already do, and pair it with an invisible watermark as the Code of Practice recommends. Move the limitation sentence out of the age policy and into the generation interface, rewritten to the formula: name what the system gets wrong, say when to double-check, and place it where the download button is. Publish an Article 50 readiness page in the Policies hub alongside the AB2013 page, and add an AI-output error channel that is not a copyright form. Those four changes alone would move Positioning and Execution enough to reach the C range.

Sources: Midjourney Terms of Service (effective May 27, 2026); Midjourney Policies hub (accessed July 29, 2026); Lumethic, “Which AI Image Generators Mark Their Output, and How” (July 2026); C2PA Viewer, “Which AI Image Generators Support C2PA?” (2026); TechTimes, “EU AI Act Chatbot Disclosure and Deepfake Labeling” (June 22, 2026).

Google Gemini’s grade and analysis, the four regulatory items that landed this week, The Signal on where disclosure is moving, and the complete EU AI Act Documentation Readiness Checklist continue below for subscribers.

Read more

Here is the timeline of the most consequential AI security incident of 2026, assembled from three different documents because no single one contains it. Around July 9, an OpenAI agent running GPT-5.6 Sol and an unreleased, more capable model attempted to break out of its isolated testing environment. On July 11, it began an intrusion into Hugging Face’s production infrastructure that lasted until July 13. On July 16, Hugging Face published a disclosure stating it had been attacked by “an autonomous AI agent system,” the identity of which was unknown. Over the weekend of July 18 and 19, OpenAI staff found evidence in internal logs that their agent had escaped. The two companies first communicated on or around July 20. OpenAI published its own account on July 21.

Notice what that sequence means. The victim disclosed before the responsible party knew it was responsible. By the time OpenAI alerted Hugging Face, Hugging Face had already contained the threat, published its post, and contacted the FBI. And you cannot learn any of this from OpenAI’s disclosure, because OpenAI’s disclosure contains no dates at all.

I read both companies’ posts end to end, and then I read the Reuters reporting that filled in what neither post says. What I found is a case study of the difference between a disclosure written to inform and one written to position. It is also a preview of a question that gets much less optional in six days.

WHAT EACH COMPANY WROTE DOWN

Both posts are, by the standards of this industry, unusually transparent. That is what makes the gap between them instructive.

Hugging Face’s July 16 post tells you the attack vector: a malicious dataset abused two code-execution paths in their dataset processing pipeline to run code on a worker, escalate to node-level access, harvest credentials, and move laterally across internal clusters over a weekend. It tells you what they verified: no evidence of tampering with public models, datasets, or Spaces, and a clean supply chain. It tells you what they did not know: which model the attacker used. It tells you what they could not do: when they tried to run forensic analysis using commercial frontier model APIs, safety guardrails blocked the work because those systems cannot distinguish an incident responder submitting attack payloads from an attacker submitting them. They ran the analysis on GLM 5.2, an open-weight model, on their own infrastructure instead, processing more than 17,000 recorded attacker events. Then they told their community to rotate tokens, gave a security contact, and named the lesson: have a capable self-hosted model vetted before an incident, not during one.

That disclosure serves the reader. It states capabilities, limitations, uncertainty, and a specific action a user can take, in that order. It even documents a failure mode of the broader AI industry, the guardrail asymmetry, that Hugging Face gained nothing by publicizing.

OpenAI’s July 21 post is a different genre wearing the same format. It confirms the essential facts, and confirming them at all deserves credit: its models, running with reduced cyber refusals for an internal evaluation called ExploitGym, exploited a zero-day in a package registry cache proxy to reach the open internet, then chained stolen credentials and additional zero-days into remote code execution on Hugging Face’s servers, all in pursuit of cheating on the benchmark by stealing the test solutions. The post calls this an “unprecedented cyber incident” and lists five remediation actions.

What the post does not contain is any date beyond “last week.” It does not say when the escape happened, when OpenAI noticed, or how. It says OpenAI’s security team discovered the anomalous activity internally, a sentence that is doing quiet work: it is technically compatible with the fuller timeline while leaving a reader to assume detection came fast. The post frames the event as an evaluation that produced valuable safety findings, closes with an invitation to apply for OpenAI’s trusted access program, and hands the final word to Hugging Face’s CEO, who praises the collaboration. A security failure became a product funnel in nine paragraphs.

Check both yourself, which takes about twelve minutes:

• Hugging Face disclosure: https://huggingface.co/blog/security-incident-july-2026

• OpenAI disclosure: https://openai.com/index/hugging-face-model-evaluation-security-incident/

WHAT ONLY ANONYMOUS SOURCES WROTE DOWN

Every date in my opening paragraph comes from Reuters, published July 24, citing people familiar with the investigation and Hugging Face co-founder Thomas Wolf. So does the detection gap: roughly a week between the first signs of trouble and OpenAI realizing its own technology was behind the attack, with the realization arriving only after Hugging Face’s public post. So does the reason: OpenAI runs several evaluations simultaneously, each generating data faster than employees can monitor in real time.

Read more

Since I started this new journey into seeing what it would look like to really work for myself and give this full-time content creator journey, I have noticed that I don’t actually fit into the “cont…

Read more

TL;DR: Adobe Firefly earns a B (83.2%) and Canva earns a D- (56.1% raw) on the ADECP framework this week, graded on one question: eleven days before the EU AI Act’s transparency obligations become enforceable, does your product tell you that the obligation exists and that it lands on you? This runs about 20 minutes at 250 words per minute, and I am introducing a new framework called the Means-and-Duty Matrix that came directly out of grading these two companies against each other. Every claim below links to a primary source so you can check my work, and where my retrieval was incomplete I say so and show you how the grade holds anyway.

PLAIN LANGUAGE FIRST

This issue leans on regulatory vocabulary that most people building marketing content have never had a reason to learn. I am defining every term before I use it, because a framework you cannot read is a framework you cannot apply. If you already know these, skip ahead four paragraphs and nothing will be lost.

Article 50 is the section of the EU AI Act that covers transparency. It is separate from the rules about high-risk systems, and it applies far more broadly. Think of it as the labeling aisle of the law rather than the safety-inspection aisle. It covers four situations: chatbots that talk to people, systems that generate synthetic images, audio, video, or text, deepfakes, and AI-written text published to inform the public on matters of public interest.

Provider and deployer are the two roles the law assigns. A provider builds and ships the AI system, so Adobe is a provider of Firefly. A deployer uses that system professionally under its own authority, so a brand marketer generating campaign images in Firefly is a deployer. The obligations split between them, and most people who are deployers have no idea they are deployers.

Machine-readable marking means the file itself carries a hidden record saying it was made or altered by AI. A human does not see it, but software can read it. Picture the difference between a sticker on a jar and a barcode printed into the glass, where the sticker peels off and the barcode does not.

C2PA and Content Credentials are the leading way that marking gets done. C2PA is the technical standard, and Content Credentials is the name for the signed provenance record it produces. The record can say who issued it, when, which app made the file, and which AI tool was involved.

Code of Practice is a voluntary rulebook the European Commission published on June 10, 2026 to help companies meet the Article 50 marking and labeling obligations. Signing it is optional, and the obligations underneath it are not. Companies that sign and follow it get what lawyers call a presumption of conformity, which means a regulator starts from the assumption that you complied rather than asking you to prove it from scratch.

Primary sources for all of the above:

• Article 50, EU AI Act

• Code of Practice on Transparency of AI-Generated Content

• Commission Guidelines on Article 50, adopted July 20, 2026

• Commission FAQ: transparency obligations under Article 50, which is the single most readable official page on this and the one I would send to a non-technical colleague

• Article 99, penalties

WHY THIS WEEK

Two things happened in the last seventy-two hours that make this the right week to grade creative AI tools instead of chat assistants. On Monday, July 20, the European Commission adopted the final version of its Guidelines on the implementation of the Article 50 transparency obligations, thirteen days before those obligations begin to apply on August 2. And today, July 22, at 18:00 Central European Summer Time, the window closes for any provider or deployer that wants to appear on the first published list of Code of Practice signatories. That list goes public before August 2, which means it becomes a permanent, timestamped, public record of who moved early and who did not.

I want to flag a retrieval problem I hit, because it is instructive. Several law firm summaries and at least one search index snapshot gave me a July 27 deadline instead of July 22. I re-fetched the Commission’s own signing FAQ before writing this sentence, and the Commission’s page states July 22, 2026 at 18:00 CEST in three separate places. When a secondary source disagrees with the body that runs the process, the body that runs the process wins. Recalling is not reading, and I re-read.

• Commission FAQ: signing the Code of Practice

THE AI DOCS WATCHLIST

🟡 Adobe Firefly (B)

What Users Need to Know

Adobe automatically attaches a Content Credential to images where every pixel came from an Adobe Firefly model, and the credential records the issuer, the date, the app, and the generative tool used. That is real provenance, applied by default, at no cost to the user, using the standard the Commission’s own Code of Practice is built around. If you are generating campaign imagery with a first-party Firefly model and publishing it into the EU, the technical half of your Article 50(2) exposure is already handled without you doing anything. What Adobe has not done is tell you that this is why it matters. Nowhere on the page that explains Content Credentials will you find the words “AI Act,” “Article 50,” “August 2,” or “deployer,” and nowhere will you find the sentence explaining that a separate obligation lands on you personally when you publish.

Where This Information Lives

• Content Credentials overview: Firefly Help, under Get started, Learn the basics. Last updated September 4, 2025.

• Known limitations in Firefly: Firefly Help, under Troubleshoot, which is the second-to-last section in the left navigation. Last updated December 12, 2025.

• Known limitations in Firefly video editor: a separate page, under a separate Troubleshoot heading, inside the video editor section.

• Inspect tool for recovering credentials: hosted on Adobe Content Authenticity, linked from the overview page.

Visual Evidence

A marketer opens Firefly Help and sees a left navigation tree running to well over a hundred links across more than a dozen top-level sections. The fifth item under Learn the basics is Content Credentials overview, so provenance is genuinely reachable in two clicks from the top. The limitations page is not in that section at all, and a reader has to scroll past image generation, vector generation, audio, video, enterprise features, custom models, bulk actions, community sharing, boards, the video editor, and design intelligence before Troubleshoot appears. A user who never encounters an error has no reason to scroll that far, which means the most honest page in the set is the page most users will never open.

The Documentation Reality

Adobe’s limitations page names ten specific constraints across four categories, and several of them are among the best-written limitation statements I have graded this year. The translation entry names the vendor, names the mechanism, and names the action. The famous-people entry explains the exact commercial rule that produces the restriction rather than waving at policy. The favorites entry gives a hard date, states plainly that deletion cannot be undone, and tells the reader what to do before that date arrives. Set against that quality, the provenance page has not been touched since September 2025, and in the intervening ten and a half months Firefly’s own navigation has grown to list six partner image models and five partner video models built by companies other than Adobe. The page that explains which outputs get a Content Credential predates the arrival of the outputs that most need the explanation.

Limitation Statement Analysis

Adobe’s translation entry tells users that prompts in over one hundred languages are machine-translated to English by Microsoft Translator, that generations from translated prompts may be inaccurate or unexpected as a result, and that users can report unsuccessful results using the icon on each image. Test it against the Limitation Statement Formula. Specific: yes, because it names the vendor and the mechanism rather than saying results vary. Contextual: yes, because it explains precisely when the failure occurs, which is when you prompt in a language other than English. Action-Oriented: yes, because it gives the reader a button and a reason to press it. That is three for three, and it is the standard I want more companies copying.

Now test the sentence that matters most for August 2. Adobe states that Content Credentials are applied automatically to assets where one hundred percent of the pixels are generated with Adobe Firefly. Specific: yes, and unusually so, because most companies would never write a threshold that invites the obvious follow-up question. Contextual: partially, because the sentence tells you the rule without telling you what happens on the other side of it. Action-Oriented: no, because a user reading it learns nothing about what to do when their image is ninety percent Firefly and ten percent generative fill. Article 50(2) covers content that is generated or manipulated, and manipulated is exactly the case that sentence leaves open.

Why B Instead of B+

Here is the arithmetic, unrounded, so you can argue with it. Ethical Stakes scored 84 and carries 25 percent, contributing 21.0. Accessibility scored 88 and carries 25 percent, contributing 22.0. Positioning scored 84 and carries 20 percent, contributing 16.8. Execution scored 78 and carries 30 percent, contributing 23.4. The total is 83.2, which I round to 83, and 84 is the floor for B+. Adobe missed the higher grade by eight tenths of a point, and the whole gap sits in Execution, which is the category I weight most heavily because it measures whether the words match the intent.

I want to name where I may be harsh. Adobe is the company that built the standard everyone else is now scrambling to adopt, it applies provenance by default rather than as a paid feature, and it publishes a real limitations page in the product help rather than burying constraints in terms of service. A reasonable auditor could hand this an A- on reputation and industry contribution. I do not grade reputation, and I do not grade what a company built. I grade what a user reads, and a user who reads every Adobe page I read still does not know that a legal duty attaches to them on August 2.

What Would Raise the Grade

Execution needs to reach 81 for the total to cross 84. One dated paragraph on the Content Credentials overview page would do it, and it would be a copy change rather than an engineering ticket. That paragraph needs to answer two questions in plain language. First, do generations from partner models inside Firefly receive Content Credentials, and if so, who is named as the issuer. Second, do partial-pixel generative edits such as fill, expand, remove, and markup receive them, given that the current rule is written as a one hundred percent threshold. A third sentence naming August 2 and pointing the reader to the deployer’s own obligation would move Positioning as well.

Sources

• Content Credentials overview, Adobe Firefly Help

• Known limitations in Firefly, Adobe Firefly Help

• Content Credentials overview, Adobe Creative Cloud

Here is what to take into your own work: find the page in your own product that explains your AI feature, and check the last-updated date against the date you shipped your most recent model or integration. If the explanation is older than the thing it explains, you have a maintenance gap that will read as an honesty gap to anyone who checks.

Canva’s grade and the full evidence trail, the four items I am watching before August 2, this week’s Signal, and the complete Means-and-Duty Matrix continue below for subscribers.

Read more

On August 2, the EU’s first broad transparency-to-users deadline arrives, and most teams I talk to still believe it is the high-risk deadline. Teams that confuse the two will prepare for the wrong obligation. The heavy machinery for high-risk systems, meaning risk management files, conformity assessments, and registration, slid to December 2, 2027 under the Digital Omnibus agreement reached in May. What actually takes effect on August 2 is Article 50, the transparency provision, and it applies to a far wider population than high-risk classification ever did. If your product puts generative AI in front of a person in the EU, Article 50 reaches you, whether or not anyone would call your system high-risk.

That breadth is the detail that keeps getting lost. August 2 is the moment the law stops asking companies to document AI for regulators and starts asking them to disclose AI to the people using it. On paper that sounds like my entire thesis arriving as enforceable regulation. In practice the deadline exposes a gap I have been grading companies against all year, because the EU wrote the duty to disclose at its weakest point and left the useful part unwritten.

WHAT ARTICLE 50 ACTUALLY REQUIRES

Article 50 sets four duties, and each one attaches to a different moment in how a person meets an AI system. Providers of systems that interact directly with people must design them so the person knows they are dealing with a machine, which means a chatbot has to identify itself rather than let someone assume they reached a human. Providers of systems that generate synthetic audio, image, video, or text must mark those outputs in a machine-readable format that detection tools can actually read, so the content carries a provenance signal even after it travels. Deployers who run emotion recognition or biometric categorization must tell the people exposed to it that it is operating. Deployers who publish deepfakes or AI-generated text on matters of public interest must label them, with a narrow carve-out when the work is plainly artistic or satirical.

Read more

TL;DR: Midjourney receives an F (27%) because its Terms forbid you from misleading anyone about whether an image is AI-generated while its policy index holds no provenance page and no Article 50 document, and OpenAI receives a B (80%) for building the strongest provenance stack in this audit and publishing the reason it falls short in a place its users will never look. Grades use the ADECP framework, scoring Ethical Stakes, Accessibility, Positioning, and Execution. Twenty-minute read, which is long for this newsletter and the longest I have run, because every claim below links to a primary source so you can check my work.

FIRST, FOUR WORDS YOU NEED

Most coverage of this deadline assumes you already speak the vocabulary, which is how a compliance story stays trapped among compliance people. Here is the whole thing in plain language. Skip it if you have it already.

Provenance is the record of where a file came from and what made it. Picture a receipt that travels inside the image rather than stapled next to it.

C2PA is the industry standard for writing that receipt, and the receipt itself is called a manifest. It carries a signature, so you can tell whether anyone tampered with it, and most social platforms delete it the moment you upload.

A watermark such as SynthID works differently. It is woven into the pixels themselves, so it survives the upload that destroys the manifest, and it carries almost none of the detail.

Article 50 is the part of the EU AI Act about telling people when they are dealing with AI. Two halves matter here. Article 50(1) says a chatbot has to tell you it is a chatbot. Article 50(2) says AI-generated content has to carry a machine-readable mark. Both apply from August 2.

THE AI DOCS WATCHLIST

🔴 Midjourney (F, 27%)

What Users Need to Know:

Article 50 requires providers of generative systems to mark synthetic outputs in a machine-readable format detectable as artificially generated. Anyone placing such a system on the EU market is in scope, and transparency failures carry penalties up to 15 million euros or 3% of global annual turnover, whichever is higher. Midjourney’s Terms of Service, in the Community Guidelines at Section 9, tell the customer: “You may not intentionally mislead recipients of the Assets about their nature or source.” That is a contractual obligation to disclose AI origin, written by Midjourney, imposed on you. The company that wrote it has published no page, no manifest specification, and no statement of any kind explaining how you would satisfy it.

Read those two facts next to each other. Midjourney identified the exact harm Article 50 exists to prevent, named it in binding terms, and assigned responsibility for it to the person paying them. Its Terms add that you are solely responsible for determining the appropriateness of redistributing the assets. The duty flows downhill and stops at you, and the company that handed it to you has never told you how to carry it.

I want to be exact about what I am grading here. Independent trackers disagree about whether Midjourney output carries a C2PA manifest, with most reporting none and a few reporting one, and I have no first-party confirmation either way. The disagreement is the finding. If a manifest exists, Midjourney has never documented it, so a user still cannot answer the question their own contract obliges them to answer. If no manifest exists, that same user stands in the same place. The grade comes out identical down either path, which is what I mean when I say the documentation gap is the ethics gap. When third parties auditing in good faith reach opposite conclusions about your product, your disclosure has already failed regardless of what your engineering does, because the entire purpose of marking is to let someone verify a claim without asking you.

Where This Information Lives:

• Terms of Service, Section 9, items 5 and 7: Two deception obligations, both pointed at you. Item 5 forbids using the Assets to deceive anyone. Item 7 forbids misleading anyone about their nature or source. Across all thirteen sections, the agreement places no matching obligation on Midjourney to mark, identify, or disclose anything.

• Terms of Service, Section 2 (Age Requirements): The agreement disclaims reliability in three places, and Section 2 is the only one that explains why the output is unreliable rather than simply refusing liability for it. That explanation is filed under whether you are thirteen years old.

• Policy index: Eight documents. One is titled “AB2013 Documentation,” the bill number for California’s AI Training Data Transparency Act. No document in the index is dedicated to Article 50, provenance, or marking.

• The Terms in full: I read the current version end to end and searched it for marking, watermarking, content credentials, C2PA, provenance, metadata, and machine-readable disclosure. None of them appear. The word “mark” occurs only inside “trademark.” The phrase “artificial intelligence” appears exactly once in the entire agreement, in the section about age.

Check it yourself, which takes about four minutes: Midjourney Terms of Service | Midjourney policy index

Visual Evidence:

A user opens Midjourney’s documentation site and clicks Policies. Eight tiles appear. One is labeled “AB2013 Documentation.” A user in Berlin scanning that grid for anything explaining how their generated image is identified as synthetic finds no tile to click. They then open the Terms and read that they personally may not mislead anyone about that image’s nature or source.

Total: 3/14 points. Band: Documentation Gap (Ethics Failure).

Q5 earns a point and I want to show my work, because it is the only point in this audit that could look generous. The framework awards PARTIAL when a product leans on generic disclaimers instead of specific examples, and Midjourney qualifies: the Terms say no guarantees are made about quality or reliability and put the Assets on an “AS IS” basis. Not one failure mode is named anywhere. So the company banks a point for announcing that nothing might work, which is exactly the behavior the PARTIAL band was built to catch. A framework that can never hand a point to a company it is failing has stopped being a framework and turned into a verdict with a rubric taped to it.

ADECP Category Breakdown:

• Ethical Stakes (25%): 25%. Synthetic imagery carries deception and impersonation risk, which is the specific harm Article 50 addresses. Midjourney names that harm and transfers it to the customer without publishing the means to manage it. Assigning a duty you supply no documented method for scores below simple silence, because silence at least does not pretend the problem is handled.

• Accessibility (25%): 25%. Nothing published means nothing to read. This category cannot rise above the floor when the artifact does not exist.

• Positioning (20%): 28%. The product surface promotes capability without bounding it. The agreement carries three refusals of liability and exactly one explanation of why the output cannot be trusted, and that explanation sits inside a section about age verification. Boilerplate does no positioning work, and a placement like that guarantees the people who need the sentence will never meet it.

• Execution (30%): 30%. The AB 2013 entry proves the publishing process works when a regulator is named. The Terms carry a version effective date of May 27, 2026, three weeks after the Digital Omnibus political agreement put Article 50 on every compliance agenda in Europe. That current version includes a Community Localization section governing volunteer translations. It includes nothing about marking.

Weighted Final Grade: (25 × 0.25) + (25 × 0.25) + (28 × 0.20) + (30 × 0.30) = 6.25 + 6.25 + 5.6 + 9 = 27.1, which rounds to 27% → F

Limitation Statement Analysis:

Found in the Terms of Service, Section 2: “This is new technology and it does not always work as expected.”

Assessment: Fails the Specific + Contextual + Action-Oriented formula on all three elements. It names no failure mode, attaches to no moment of use, and asks the reader to do nothing. A sentence that could be pasted into any product ever shipped tells a user nothing at all, and this one reads like a shrug with a legal department standing behind it.

Read it next to the two disclaimers that outrank it and the pattern gets hard to miss. Section 1 refuses liability for quality and reliability. Section 10 says the company makes no promises about the service. Both of those protect Midjourney. Section 2 holds the only sentence that tells you something true about how the product works, and it survives in the agreement not because anyone decided users deserved an explanation but because a parent might read that far. The company wrote three sentences for its lawyers and one for you, then filed yours where it would do a thirteen-year-old the least good.

Why F Instead of D:

A D means users are operating largely blind while some disclosure exists to squint at. An F means the record is empty where it counts and the company has demonstrated it knows how to fill it. Three pieces of evidence make this a choice rather than an oversight. Midjourney published a document naming a California statute and nothing naming the European obligation, using the same team, the same lawyers, and the same publishing pipeline. Its Terms were revised effective May 27, 2026, well inside the window when Article 50 was front-page compliance news, and the current version finds room to secure rights over volunteer translations. And it wrote the user’s disclosure duty into binding terms while publishing nothing that would let them satisfy it.

What Would Raise the Grade:

1. Publish an Article 50 page in the same policy index that already holds the AB 2013 entry, stating plainly whether outputs are marked, by what method, and from what date. If they are marked today, say so, because the trackers cannot tell.

2. Give users a documented way to satisfy the duty you already assigned them. One sentence at download would do it: “This image carries a machine-readable marker identifying it as AI-generated. Most social platforms strip it on upload.”

3. Move the unreliability sentence out of Age Requirements and into a named limitations section.

4. Sign the Code of Practice by July 22 or publish the reasoning for declining. Either one is a record. Silence is not.

Sources: Midjourney Terms of Service, version effective May 27, 2026, read in full July 14, 2026; Midjourney policy index, retrieved July 14, 2026, where the AB 2013 entry is identified by title because the article itself blocked automated retrieval; C2PA Viewer, “Which AI Image Generators Support C2PA?” (February 27, 2026); AI Video Bootcamp True Model Provenance Matrix (April 2026); EU AI Act Article 99.

OpenAI’s grade and full audit, the four rungs of The Disclosure Evidence Ladder, and this week’s Signal on why the EU’s new icon moves comprehension by two points continue below.

Read more

This month the EU’s Digital Omnibus on AI took effect, and most of the coverage compressed into three words: the deadline moved. That framing is half right and quietly misleading. The Omnibus pushed the AI Act’s most demanding obligations, the conformity assessments and technical files required for high-risk systems in areas like hiring, credit scoring, and law enforcement, from August 2, 2026 to December 2, 2027. It left something else exactly where it stood.

The transparency obligations under Article 50 did not move. Starting August 2, 2026, a company running a chatbot for people in the EU has to tell them they are talking to a machine, and a company generating synthetic text, images, audio, or video has to mark that output as artificially generated. The European AI Office also gains the power to fine general-purpose model providers on that same date, closing a year in which those obligations existed on paper but carried no penalty. Brussels moved the internal paperwork and kept the disclosures that reach users. That single decision is the argument I have been making for seven months, now written into binding law by the people who wrote the law.

Even the regulator hit the gap

Ask why the high-risk obligations slipped sixteen months and you get an answer that has nothing to do with the risk going down. The harmonized standards were not finished. The conformity tools were not built. National authorities were not in place: at least twelve member states missed the deadline to appoint a competent authority, France had not notified one as of June, and Germany’s implementing law was still moving through the Bundestag this month. A Vision Compliance survey in April found that 78 percent of organizations had taken no meaningful compliance steps, and the Cloud Security Alliance reported that more than half still lacked a basic inventory of the AI systems they run. The rules existed. The machinery to prove compliance with them did not.

I have a name for this distance, and I have been using it since January. The Ethics-to-Documentation Pipeline runs through five stages: Principle, Policy, Pattern, Documentation, User Understanding. A company writes a principle, hardens it into policy, and then, in the best case, turns that policy into a repeatable pattern that produces an explanation a user can actually read and act on. Most organizations stall between Policy and Pattern. They hold the values page and the internal framework, and nothing downstream of it reaches a person. What the Omnibus tells us is that the European Union, at the scale of a continent, stalled in the same place. It could write the principle and pass the policy. It could not build the operational record fast enough to make that policy enforceable, so it moved the date.

This should change how you hear the word “governance” when a vendor or a leadership team reaches for it. A governance framework describes intent. It states what a company believes about fairness, oversight, and disclosure. None of that protects a single user until someone converts it into an explanation the user can find at the moment they need it. When the conversion never happens, the framework is a values statement wearing a compliance costume. The EU just conceded, in law, that it could not finish its own conversion on the original schedule.

Read more

A few weeks ago, a company asked me to design a multi-agent voice system for patient triage as part of their interview process. I had honestly never built anything like this before, but was up for the challenge. I built it. I presented it to a panel. The team approved my design, but I still did not end up getting the job.

So I wanted to write this as a way for me to reflect on what I learned. This project taught me more about agentic system design than most of what I’d read about in articles. I’m also going to explain every term as I go, because I feel like the AI industry hides behind jargon as a way to confuse the non-technical, and I’d rather teach you the actual concepts than make you feel like you need a computer science degree to follow along with what I am writing.

First, what is a “multi-agent system”?

Imagine a hospital phone line. When a patient calls in, a few different things need to happen in order: someone needs to answer and figure out if this is an emergency, someone needs to ask about symptoms, someone needs to collect insurance information, and someone needs to book the actual appointment.

A human-staffed phone line might have one person doing all four jobs, or four different people passing the call along. An “agent,” in AI terms, is a piece of software powered by a language model (the same technology behind ChatGPT or Claude) that is given one specific job and the tools it needs to do that job. A “multi-agent system” is several of these specialized agents working together, handing the conversation off to each other the way a receptionist transfers a call to a nurse.

The assignment asked me to design four agents:

1. Start Agent. Answers the call and figures out how serious the situation sounds.

2. Triage Agent. Asks about symptoms and decides how urgent the situation is.

3. Patient Info Agent. Collects insurance, medications, and the reason for the visit.

4. Scheduling Agent. Finds an available doctor and books the appointment.

Each one only does its own job. None of them tries to do everything, which matters more than it sounds like.

Why does it need to be voice, and why does that make it harder?

When you type into a chatbot into gives you time to think and lets the system re-read your message. However, when a phone call happens, it’s happening in real time. So if someone who is scared, in pain, or panicking is talking, and the system has to understand them, decide what to do, and respond, all in the space of a normal conversation, it leaves no room for the person to overthink, delete, or edit the conversation.

Read more

TL;DR: Salesforce Einstein receives a C (70%) for keeping its real limitations in a security PDF while its interface talks almost exclusively about trust and reassurance. HubSpot Breeze receives a B+ (85%) for stating a specific model limitation in plain language inside its own help documentation, no digging required. Both sell to the same revenue teams. Only one of them treats you like you can handle the truth. 4-minute read.

THE AI DOCS WATCHLIST

🔴 Salesforce Einstein (C, 70%)

What Users Need to Know:

Salesforce Einstein is the machine learning layer built into Sales Cloud, Service Cloud, and Marketing Cloud, in the platform in some form since roughly 2018. The feature most relevant to daily use is Einstein Activity Capture, which automatically logs emails and calendar activity into CRM records. Independent reviewers have documented that the feature has known limitations in accurately attributing those logged activities to the correct opportunity, which matters directly to a sales team relying on that data for forecasting. Salesforce’s own public documentation does not put that limitation anywhere a rep evaluating the tool would actually see it. Instead, the company’s AI messaging centers on the Einstein Trust Layer, built to reassure customers about data privacy and security, and multiple Salesforce executives have described trust and privacy as the starting point for their generative AI design without matching that language with equally specific failure-mode disclosure. VantagepointVantagepoint

Where This Information Lives:

• Einstein product marketing pages: Capability claims and trust language. No specific automation accuracy limitations found.

• Einstein Platform Notices and License Information (security PDF): Legal and security architecture detail, written for compliance teams, not end users.

• Third-party product review sites: The actual accuracy limitations of features like Activity Capture surface here, not on Salesforce’s own site.

• Salesforce’s own AI governance guidance: Confirms Einstein can touch high-risk use cases like loan evaluation or Health Cloud diagnostics, which raises the stakes without lowering the bar for user-facing disclosure.

Visual Evidence:

A user researching Einstein through Salesforce’s own channels repeatedly encounters language about the Trust Layer preventing large language models from retaining sensitive data, alongside descriptions of what each feature can do. No sentence in that path says “this feature sometimes misattributes activity to the wrong deal.” A user has to leave Salesforce’s own site and read a third-party comparison to learn that.

Total: 5/14 points - Documentation Theater

ADECP Category Breakdown:

• Ethical Stakes (25%): 75% – Salesforce’s own compliance guidance acknowledges that AI agents evaluating loan applicants or supporting medical diagnoses in Health Cloud fall under high-risk obligations requiring risk management, human oversight, and audit trails. The stakes are real and Salesforce names them internally. They just don’t show up where a working user would see them. Blog

• Accessibility (25%): 72% – The security PDF is thorough but written for a compliance audience. A sales rep or admin evaluating Einstein for daily use has no plain-language equivalent naming what the automation gets wrong.

• Positioning (20%): 62% – Trust and reassurance language dominates the product’s public face. Real limitations exist and are documented by outside reviewers, but Salesforce’s own positioning doesn’t give them equal weight to the capability claims sitting right next to them.

• Execution (30%): 68% – Governance infrastructure exists, and teams are advised to treat AI setup as a data, process, and governance project rather than a simple toggle, which is a real strength. What’s missing is translating that infrastructure into a limitation statement a working user can act on before trusting an automated field update. Digital Applied

Weighted Final Grade: C (70%)
Calculation: (75 × 0.25) + (72 × 0.25) + (62 × 0.20) + (68 × 0.30) = 18.75 + 18 + 12.4 + 20.4 = 70%

The Documentation Reality:

Salesforce has built real infrastructure for AI governance, and the audit trails and explainability features referenced in its own compliance guidance are not nothing. The problem is sequencing. That governance work has gone into legal and security material aimed at satisfying regulators, while the more basic job of telling an end user what the automation does poorly has been left to product reviewers and comparison blogs. Capability language lives in the product. The limitation record lives in a PDF. The two never meet where a user is actually making a decision.

This is what documentation theater looks like in practice. It is not an absence of written material, since Salesforce has plenty. It is material aimed at the wrong reader, answering “are we compliant” instead of “does the person using this every day know where it fails.”

Limitation Statement Analysis:

The closest thing to a real limitation statement comes from outside reviewers, not Salesforce itself: Einstein Activity Capture logs emails and calendar events automatically, but the feature carries known limitations in correctly attributing that activity to the right opportunity. Vantagepoint

Assessment: Fails the Specific + Contextual + Action-Oriented formula at 20%. It’s specific once you find it. It isn’t contextual, since it doesn’t appear where a user configures or relies on Activity Capture. It isn’t action-oriented, since Salesforce hasn’t published guidance on what to do when the misattribution happens.

Why C Instead of Higher or Lower:

Einstein avoids full documentation-gap territory because governance infrastructure and audit capability genuinely exist, with real investment tied to EU AI Act readiness. It stays at a C rather than climbing higher because none of that infrastructure has been translated into a limitation statement a working user can find without leaving Salesforce’s own site.

What Would Raise the Grade:

Publish a single “known limitations” page linked directly from the Einstein product page, naming the Activity Capture attribution issue in plain language instead of leaving it to third-party comparisons. Add a visible flag or correction affordance inside the CRM record itself when an AI-suggested field update looks wrong, rather than routing that feedback through a general support channel.

Sources:

• eesel.ai, “What is Salesforce Einstein AI? Full guide for 2026” (November 2025) — https://www.eesel.ai/blog/what-is-salesforce-einstein-ai

• oliv.ai, “Salesforce Einstein Features: What Works, What Doesn’t” (October 2025) — https://www.oliv.ai/blog/salesforce-einstein-features

• Cirra, “Salesforce AI Governance: A Guide to GDPR, CCPA & EU AI Act” (October 2025) — https://cirra.ai/articles/salesforce-ai-governance-compliance

• HyphenX Solutions, “How to set up AI in Salesforce using Agentforce, Data Cloud, and Einstein” — https://hyphenxsolutions.com/Blog/how-to-set-up-ai-in-salesforce-using-agentforce-data-cloud-and-einstein/

• Salesforce, “Einstein Platform Notices and License Information” (2026) — https://www.salesforce.com/en-us/wp-content/uploads/sites/4/documents/legal/misc/einstein-platform-security-privacy-and-architecture.pdf

Continue reading below for HubSpot Breeze’s grade and analysis, The Signal on what separates these two companies, and this week’s Interface Disclosure Ladder.

Read more