2 months after the EU AI Act's transparency rules bit, 1.7% of AI products say they mark their output
Hlido probed 873 AI products against Article 50 of the EU AI Act. 15 publish any machine-readable marking of AI-generated content. 4 offer a detection tool. Enforcement day changed almost nothing we could observe.
By the Hlido Editor · 2026-09-28
On 2 August 2026, Article 50 of the EU AI Act became enforceable and national market surveillance authorities gained the power to investigate and fine. We had been probing the public surfaces of AI products for weeks before that date, and we kept probing after it.
Here is what we can observe.
The numbers
Of 873 AI products whose public surfaces we probed for Article 50 signals:
- 15 (1.7%) publish any statement that their AI-generated output is marked in a
machine-readable form.
- 4 (0.5%) offer a detection or verification tool for their own output.
- 244 (24%) land in our READY band on transparency readiness overall; 617 (61%) are
NOT-READY.
And across the enforcement date itself, on 709 products whose baseline observation predates 2 August: 0.4% changed anything we could see. 3 products added a marking statement. None dropped one.
That near-zero is the finding, not a disappointing result. A deadline arrived, became enforceable, and the observable market did not move. There is very little drift here because there is very little to drift from — 98.3% of the products we probed were silent on machine-readable marking before 2 August and are silent after it.
What that does and does not mean
It does not mean 858 companies are breaking the law. Article 50(2) requires providers to mark synthetic output in a machine-readable format — and marking lives inside the output, not on the marketing page. A vendor can implement it perfectly and publish nothing about it. Our probe reads public surfaces; it cannot open a model's output pipeline.
What the number does measure is discoverability of the claim. If a company has done this work, it is currently choosing not to say so anywhere a buyer, an auditor, or an automated checker can find it. For an obligation whose entire purpose is transparency, silence is a strange posture.
It also means something concrete for anyone buying AI: you cannot currently verify this from outside, for 98.3% of the market, without asking the vendor and taking their word.
The deadline nobody is talking about
Article 50 arriving on 2 August got the attention. The one that will bite harder is 2 December 2026, when the machine-readable marking duty extends to systems already on the market. Most of what we probed is already deployed. On today's evidence, essentially none of it is publicly ready for that.
The rest of the AI Act's high-risk chapter was deferred sixteen months by the Digital Omnibus, adopted by the Council on 29 June 2026 — standalone Annex III systems to December 2027, embedded systems to August 2028. Article 50 was not deferred. It is the obligation that is live now, and it is the one with the least visible compliance.
Method, and its limits
We probe public surfaces only: the product's own pages, and a small bounded set of conventional transparency paths. We look for an explicit marking statement (C2PA, Content Credentials, provenance metadata, watermarking language) and for a detection tool. Every observation carries the date it was made.
Three things we deliberately do not do:
- We do not say "compliant" or "non-compliant." We are not a notified body, Article 50 has no
conformity-assessment route, and we hold no authority to certify anything. We report what we observed and when.
- We do not score a vendor down for blocking us. A site that refuses our probe is recorded as
not observed, never as an absence of marking. Where a plain fetch is blocked we render the page like a browser; if that also fails, the vendor is excluded from the denominator rather than counted against.
- We take corrections. If your row is wrong, tell us and we will re-probe and re-date it. The
fix list at hlido.eu/eu-ai-act/fix-list/ says exactly what would change a row.
The data
The full register is public and machine-readable at hlido.eu/eu-ai-act/ with per-product rows, and the corpus behind it — 1,011 products, claim-by-claim verdicts, incident history, each observation dated — is a CC-BY-4.0 download at hlido.eu/data/open/. Check any product, including ones we have never reviewed, at hlido.eu/eu-ai-act/transparency-check/.
We take no payment to rank or review, from anyone, ever.
Hlido is an independent evaluation desk based in Europe. Scores and claim verdicts are produced by automated testing of public surfaces; this analysis is model-drafted from that data and published without a person reviewing every row.