Brief Nº4
In English · Svenska: listan över anteckningar
The AI rule that is easy to write down and hard to verify
Since 2 August 2026, anyone running an AI system that talks to people in the EU has had to make sure those people know they are dealing with an AI — not buried in terms and conditions, not two menus deep, but clearly and at the latest at the first interaction (Cooley). That is Article 50 of the AI Act. It is one of the shortest obligations in the regulation, and the one we find most interesting — because of the distance between how simple it sounds and how hard it is to verify.
Our position: Article 50 is not a wording problem, it is a placement problem — and that makes it the rare obligation where the honest answer to "are you covered?" cannot be produced by scanning anything. The counter-view, stated fairly: a disclosure line is an hour of developer time, so this will quietly fix itself as vendors ship it by default. We think that is probably true for the sentence and not true for the evidence that anyone checked.
What it is trying to solve
Not AI deception. Something smaller and more useful: not knowing. When you cannot tell whether the thing answering you is a person, you cannot judge what to do with the answer — whether to repeat yourself, whether a promise will be remembered, whether asking twice will get you a different reply. Article 50 does not ask anyone to stop using AI. It asks that a person be able to tell.
There is a second part with a nearer deadline. Systems already on the market before August have until 2 December 2026 to mark synthetic output in a machine-readable way, so that software downstream can tell too (Baker Botts). That one is a build task, not a wording task.
Three things that make it harder than it reads
The duty attaches to a moment, not to a page. Most compliance questions are about what you said. This one is about where and when you said it. A company can have perfectly good disclosure wording — in a privacy policy, in a footer, on an AI page linked from "about" — and still not meet the standard, because the obligation bites at the first interaction. Commentary is explicit that information buried in terms, or under layers of menus, does not satisfy it (Travers Smith). The sentence can be right and the position wrong.
It cannot be checked from the outside without interacting. We built a screen for this and the useful finding was a negative one. From a public page you can establish two things: that a chat exists, and whether disclosure wording appears anywhere in the page that is served. You cannot establish what a visitor is shown in the first second of a conversation, because that lives inside a widget that loads only when someone opens it. So the part that matters needs a person, or a browser driven like one. Any tool reporting an Article 50 pass or fail from scanning alone is overclaiming on the single point the provision is about.
Detection itself is unreliable in the direction that flatters you. Identifying a chat by its vendor works for sites that load a vendor's script and fails for large organisations that self-host the widget or route it through their own tag manager — which is to say, it fails on the ones with the most users. In our first live run, half the sites we scanned returned "no chat" and were wrong. A screen built this way under-reports, quietly, and under-reporting looks exactly like good news.
The run that taught us the most was the one that failed
Our first attempt reported that a set of large, well-known sites ran no chat at all. Nothing was wrong with those sites. The machine doing the scanning was missing a certificate bundle, so every single request failed verification — and a run in which everything fails looks precisely like a run in which there was nothing to find. The script exited successfully and printed a tidy zero.
We have seen this exact shape before in our own work: a watcher that hit a wall of connection resets and produced an empty report that read like a quiet week. The lesson is not about certificates. It is that a silent failure reports "nothing to see" and is believed. Our checks now carry an expected volume, and a run that comes in under it says "this source may be broken" instead of printing a number. It is also why we will not publish a disclosure rate for Swedish sites today: we do not yet have one we could defend, and a number without its error rate is decoration.
Our own site, checked first
We sell checks, so we run them on ourselves before anyone else. On 17 October 2026 we screened our own pages. Result: no chat widget and no AI system that interacts with visitors, so the first-interaction duty does not currently reach us. That is an easy pass and we are not going to dress it up as one.
The honest finding is elsewhere. The text on this site is drafted by AI agents — we say so on the front page — which engages the part of Article 50 about AI-generated text rather than the part about chat. Reporting describes an exception where content has had human editorial review and a person takes responsibility for publishing it (Cloud Security Alliance). That exception is exactly our operating model: one person reads and approves anything published, and that person is named on the about page. We rely on the review step, not on the drafting being invisible — and if the review stopped, the position would change.
Where we think this lands
- The live obligation is the quiet one. The heavier high-risk requirements moved to late 2027 (Bird & Bird), so most attention sits on what is not yet due while the thing that is due attracts none.
- Enforcement is mid-assembly. In Sweden, authorities were designated on an interim basis while the implementing legislation is settled, and no AI Act fine or investigation is publicly known. A duty that applies before its enforcement exists is a strange interval, and intervals end.
- The ceiling is not small — reported as up to €15 million or 3% of worldwide turnover. That is a number for the end of the interval, not for this month.
- Treat "we added the line" and "we can show we checked" as different assets. The first is an hour. The second is a dated record, and only one of them is worth anything if someone asks a year from now.
Where we sit — honestly
We are a one-person company that checks claims against evidence and prints what it could not check. We do not certify anything and this is not legal advice — we are not lawyers, and on what Article 50 requires in your specific case you should ask one. Every legal statement above comes from published commentary, linked at the point it is used; we have not read the regulation's own text or the Commission's July guidelines, and we would rather say so than imply otherwise.
What we can do is narrower. Take a rule, work out which parts a machine can establish and which parts need a person, build the check, measure how often it is wrong, and write down both numbers. For Article 50 that means an outside view of whether a chat exists and whether disclosure wording is present, an explicit list of what was not examined, and a short manual step for the one test that decides it. If that is useful, write to us — we take one piece of work at a time, under an NDA if you prefer, and we will say so when the answer is no.
Also read: Brief Nº5 — Everything you are told to check is public. None of it is in one place. →