what changes on august 2, and what does not
two rules about machine detection of ai content take effect on the same day, on two continents.
the two rules
eu ai act, article 50(2). from 2 august 2026, providers of generative ai systems must mark synthetic audio, image, video and text outputs in a machine-readable format. the obligation asks that marking be effective, interoperable, robust and reliable as far as technically feasible. one enacted exception: generative systems already on the market before 2 august 2026 get until 2 december 2026 to comply; systems placed on the market from 2 august must mark from placement. article 50 was not swept up in the digital omnibus postponement of high-risk rules: chatbot disclosure, deepfake labelling and public-interest-text labelling all apply from 2 august 2026.
california ab 853 (the ai transparency act, as amended). operative the same day: covered generative-ai providers with over 1,000,000 monthly visitors or users must make an ai detection tool available at no cost to the user.
the date table
maintained against the primary sources linked below. last verified 2026-07-30.
| date | what applies | source |
|---|---|---|
| 2 august 2026 | article 50 transparency obligations: chatbot disclosure, deepfake labelling, machine-readable marking for newly placed generative systems | commission article 50 faq |
| 2 august 2026 | california ab 853 free detection tool becomes operative | california legislature, ab 853 text |
| 2 august 2026 | commission enforcement powers over gpai providers (fines up to 3% of turnover or eur 15m) | commission gpai framework pages |
| 2 december 2026 | article 50(2) marking for generative systems already on the market before 2 august 2026 | commission faq; regulation (eu) 2026/1744, recital 38 |
| 2 august 2027 | gpai models placed on the market before 2 august 2025 must be brought into compliance | ai act article 111 |
| 2 december 2027 | high-risk rules for stand-alone (annex iii) systems, as postponed by the digital omnibus | regulation (eu) 2026/1744 |
| 2 august 2028 | high-risk rules for ai embedded in products (annex i), as postponed | regulation (eu) 2026/1744 |
what “detectable and reliable” looks like as evidence
the rules name properties: effective, interoperable, robust, reliable, detectable. none of them name a metric. a claim without a measurement procedure is a hope, so here is what evidence for each property would look like:
- detectable means an independent verifier recovers the mark from real outputs, and every negative states what it means. absence of a mark is only informative when the carrier could have held the mark; that is a per-item self-test, not an assumption.
- reliable is an error rate, measured: how often the mark survives ordinary handling (re-encode, resize, excerpt), with false positive and false negative rates stated at an operating point, with intervals.
- robust is the same measurement after an adversary: the degradation curve, published, not the adjective.
- interoperable means checkable by someone else’s tooling, not only inside the vendor’s own portal.
that is measurement work, and it is the work gaige receipts: evidence statuses instead of verdicts, a per-item carrier self-test before any watermark negative is trusted (gaige verify, shipped in 0.0.3), and calibrated operating points with honest error bars for whatever detector you already run. as of 2026-07-30, our searches found no market verification tool that runs a per-image carrier self-test before trusting a watermark-absence result; we re-verify that quarterly and will correct this page when it changes.
footer
this page is not legal advice; it is a dated reading of public texts with links to the sources. for compliance decisions, read the regulation and talk to counsel. corrections welcome: the sources are linked so you can check us.
the gaige team
sources: commission article 50 faq (digital-strategy.ec.europa.eu) · regulation (eu) 2026/1744, oj l, 24.7.2026 · ai act article 50 · california ab 853
adjacent: guides · the package · custody · faq