What started applying on 2 August
Two things happened on the same date. The European Commission’s AI Office, together with national authorities, began enforcing the AI Act, and the transparency obligations in Article 50 started to apply. Those obligations are specific: interactive systems have to tell people they are dealing with AI rather than a human, deepfakes have to be labelled, and AI-generated or altered content has to carry a machine-readable mark so it can be detected. [1][2]
The duties are split along the value chain. Providers must design systems so a person is informed when they are interacting with one directly, and must add the machine-readable marking to generated or manipulated output. Deployers carry a different set: telling people when they are subject to emotion recognition or biometric categorisation, disclosing deepfakes, and disclosing text published on matters of public interest that had no human review or editorial control. Enforcement runs through national market surveillance authorities, the AI Office for the systems it supervises, and the European Data Protection Supervisor for EU institutions. [2]
Atlas interpretation: The deployer clause about unreviewed text on matters of public interest is the one that reaches beyond model vendors. It puts a disclosure duty on the newsroom, the campaign or the trade body that publishes generated copy without a human in the loop, which is a rule about editorial practice wearing the clothes of a technology regulation. [2]
What did not start on 2 August
The AI Act’s high-risk regime was due to apply from the same date and does not. The AI Omnibus, whose text the Parliament and the Council finalised in July 2026, provides that the rules for stand-alone high-risk systems apply from 2 December 2027 and those for high-risk systems embedded in regulated products from 2 August 2028. That regime is the heavy one: risk management, data quality, human oversight, accuracy, registration, monitoring, record-keeping and incident reporting. [3]
The same package also moved a deadline inside the transparency rules. Providers of generative models released before 2 August 2026 have until 2 December 2026 to mark their outputs. That is why marking arrived in the field attached to new model releases rather than as a retrofit across everything already deployed. [3][4]
Atlas interpretation: So the weight of the date is not a new compliance burden on most systems. It is that the AI Office finally holds enforcement powers, a year after the general-purpose model obligations began binding providers, and against a much-reduced set of things it can enforce this year. A regulator with authority and a narrow current mandate is a different animal from one whose main deadline has been pushed out twice. [1][3]
What "machine-readable mark" turns out to mean
The obligation is technology-neutral, so the interesting question is what providers actually built. Anthropic’s implementation, published twelve days after the date and applied worldwide, is a version of the SynthID-Text approach: when the model is choosing among words that are equally good, the choice is still random, but the randomness comes from a key and the preceding words rather than from an arbitrary generator. Nothing is added to the text, no hidden characters, no extra tokens, and the mark carries no information about the user. [4]
The limits are as specific as the mechanism. The mark needs room: it is sparse in factual passages where only one word is correct, largely absent from code for the same reason, and weak on short samples. It cannot distinguish text Claude wrote from text Claude heavily edited, cannot say anything about text written by a different model, and a complete rewrite removes it. For files rather than text, Anthropic attaches a C2PA content credential in metadata instead, which is a different thing entirely: a signed note that anything can strip. [4]
Atlas interpretation: Put those limits next to what the rule is for, which the Commission describes as reducing deception and manipulation, and the fit is partial by construction. The mark is good at the honest case, where somebody wants to check provenance and nobody is trying to hide it. It is weak in exactly the adversarial case the deepfake and disclosure provisions are aimed at, because removing it is easier than embedding it. That is not a flaw in the Act’s drafting so much as the current state of the art, and it is the reason the same rule also relies on deployer disclosure rather than on detection alone. [4][2]
How compliance gets demonstrated
The Commission published a first list of more than 180 organisations that signed the Code of Practice on Transparency of AI-generated Content, the instrument that operationalises these rules; Anthropic counts around 190 signatories as of July 2026. Signing is not required, but a provider that declines has to show compliance by other, equivalently adequate means. [1][2][4]
Atlas interpretation: The practical reach of that arrangement shows up in one sentence of Anthropic’s explanation: it applied the watermark globally at launch because it did not yet have a durable way to scope it by region. A European rule became a worldwide product behavior for engineering reasons, not policy ones. The same month, Apple Music made AI-content tagging mandatory for submissions, which is the private-contract version of the same idea arriving on its own schedule. [4]
Atlas interpretation: What is missing is the reader’s side. Anthropic’s detection API opened in private preview to regulators, law enforcement, media, fact-checkers, researchers, educational organisations and civil society groups, plus enterprises with their own obligations. That is a sensible list and it is not the public. Until checking a mark is something an ordinary person can do, the transparency in these rules is transparency to institutions, and whether that was the intent is a question the first enforcement actions will answer better than the text does. [4]
Sources
- Commission starts enforcing AI Act rules and new transparency requirements on 2 August
European Commission · Jul 31, 2026
- Guidelines on transparency obligations for providers and deployers of certain AI systems
European Commission
- Law delaying EU’s "high-risk" AI rules finalised
Pinsent Masons · Jul 6, 2026
- How Claude’s text watermark works
Anthropic · Aug 14, 2026