Anthropic sets out how it will mark Claude's text as the new transparency obligations take effect, while experts reject claims that watermarking degrades AI writing.
I like living in EU. I am starting to see AI content labeled and its really such a relief every time.
A bit frightening sometimes too. There was this audio ad playing on radio for months now. Just recnty they added one sentence disclaimer: "Read by an AI lector." I listened to it every other day and had no idea!
Casovan's framing of the shift, from an educated guess to a verified symbol, is the clearest way to describe what actually changed in August. The part I keep circling back to is what the symbol is attached to. Every mechanism in this issue operates on the output. It certifies something about the file and nothing about the person the file depicts.
That distinction gets uncomfortable in the likeness cases specifically. An unmarked synthetic image of a real woman is not really a provenance problem, it is an identity problem, and the transparency stack has no way to reach it. What she would need is a record of her own that predates the image, and that is the piece nobody is building.
I went and counted earlier this year. Across ten states, a federal bill and the European rule, 21 instruments now create a right over a face or a voice. Eight of them let a person file something. Seven of those eight only open after you have died. So the marking obligations are arriving years ahead of any way for a living person to put the real thing on record in the first place. I would be curious whether the transparency Code work ends up reaching that at all, or whether it stays firmly on the generation side.
Provider marking is only half of what applies from 2 August 2026. The Commission's own Quick Facts page on Article 50 puts a separate duty on deployers: disclose deepfakes, and label AI-generated text published to inform the public on matters of public interest where it did not go through human review or editorial control. That reaches ordinary organisations, not only labs. The same page sets the ceiling at 15 million euros or 3 percent of worldwide annual turnover, with proportionality for SMEs. Machine-readable marking by the model provider does not discharge the deployer's labelling obligation.
Risto, an essential analysis on the rollout of the August 2026 transparency rules and Anthropic’s statistical text watermarking.
While setting the statistical watermark threshold at 200 tokens addresses general output quality, regulated enterprise workflows—particularly across financial advisory and wealth tech—face a much sharper structural hurdle: downstream multi-agent tool execution.
As the AI Act Service Desk notes under Articles 3(1) and 3(63), autonomous agents process environment state and execute multi-step tools. In composite agentic pipelines where outputs are re-prompted, parsed through intermediate APIs, or transformed across private models, invisible linguistic watermarks strip out instantly.
For institutional compliance and systemic risk mitigation, enterprise governance cannot rely solely on generative word-choice watermarks; it requires immutable cryptographic provenance at the state-transition and tool-execution layer. Outstanding overview of where regulatory compliance meets frontier engineering reality.
The watermarking date has two deferrals sitting behind it that the coverage rarely names. Article 50(2) applies from 2 August 2026, but the Digital Omnibus pushed marking and detection for generative systems already on the market to 2 December 2026, and watermark-detection interoperability to 2 February 2027. Does the transparency Code say what a deployer is expected to check during that gap?
Watermarking is the pleasant part. Brussels has now sent its first AI Act letters on security, safety and copyright, which is where the transparency talk stops being theoretical.
The watermarking debate is framed as a provider question, which is right for Article 50(2). The gap I keep meeting sits one layer down, with deployers.
Marking attaches to design and development. Disclosure under Article 50(4) attaches to the organisation running the system, and for text it turns on whether the publication informs the public on a matter of public interest and whether the human review and editorial responsibility carve-out is available.
Two practical consequences for bodies that only deploy. A marking capability that ships enabled can be switched off during configuration, and nothing in the provider's compliance story surfaces that. And the deployer's own position rests on a documented review step and a named responsible person, not on the provider's marking.
The question worth putting to suppliers in writing: may the marking capability be disabled, and is any change to it logged.
Software vendors operating in the European healthcare space must design watermarking and audit compliance into their core architectures to avoid substantial non-compliance penalties under the EU AI Act.
The 200-token threshold is the useful number here, since a lot of short-form output sits under it. Worth pairing with it: under the Omnibus, systems already placed on the market before 2 August 2026 aren't covered by the Article 50(2) machine-readable marking duty until 2 December 2026, so for the next few months an unmarked output tells you very little either way. Does the transparency code say anything about that gap?
On the gap itself, the Regulation is the harder constraint and it is narrow in three ways worth keeping separate. Article 111(4) reaches providers only, it covers the Article 50(2) marking duty only, and it applies only to systems placed on the market before 2 August 2026. Paragraphs 1, 3, 4 and 5 are untouched by it.
So for the next few months an unmarked output tells you little about the provider, as you say, but it tells you nothing about the deployer. A body publishing AI-generated text on a matter of public interest sits under Article 50(4) now, and its position turns on a documented review step and a named responsible person rather than on whether the output carries a mark.
I have not read the Code closely enough to say whether it addresses the transitional window directly, so I would not assert either way. What is clear is that adherence is voluntary and cannot move a date set in the Regulation.
I like living in EU. I am starting to see AI content labeled and its really such a relief every time.
A bit frightening sometimes too. There was this audio ad playing on radio for months now. Just recnty they added one sentence disclaimer: "Read by an AI lector." I listened to it every other day and had no idea!
Thanks for sharing your experience, interesting indeed!
Casovan's framing of the shift, from an educated guess to a verified symbol, is the clearest way to describe what actually changed in August. The part I keep circling back to is what the symbol is attached to. Every mechanism in this issue operates on the output. It certifies something about the file and nothing about the person the file depicts.
That distinction gets uncomfortable in the likeness cases specifically. An unmarked synthetic image of a real woman is not really a provenance problem, it is an identity problem, and the transparency stack has no way to reach it. What she would need is a record of her own that predates the image, and that is the piece nobody is building.
I went and counted earlier this year. Across ten states, a federal bill and the European rule, 21 instruments now create a right over a face or a voice. Eight of them let a person file something. Seven of those eight only open after you have died. So the marking obligations are arriving years ahead of any way for a living person to put the real thing on record in the first place. I would be curious whether the transparency Code work ends up reaching that at all, or whether it stays firmly on the generation side.
Very interesting stuff. I've also written about this. Your stack should be a good resource for me in the future.
Provider marking is only half of what applies from 2 August 2026. The Commission's own Quick Facts page on Article 50 puts a separate duty on deployers: disclose deepfakes, and label AI-generated text published to inform the public on matters of public interest where it did not go through human review or editorial control. That reaches ordinary organisations, not only labs. The same page sets the ceiling at 15 million euros or 3 percent of worldwide annual turnover, with proportionality for SMEs. Machine-readable marking by the model provider does not discharge the deployer's labelling obligation.
Risto, an essential analysis on the rollout of the August 2026 transparency rules and Anthropic’s statistical text watermarking.
While setting the statistical watermark threshold at 200 tokens addresses general output quality, regulated enterprise workflows—particularly across financial advisory and wealth tech—face a much sharper structural hurdle: downstream multi-agent tool execution.
As the AI Act Service Desk notes under Articles 3(1) and 3(63), autonomous agents process environment state and execute multi-step tools. In composite agentic pipelines where outputs are re-prompted, parsed through intermediate APIs, or transformed across private models, invisible linguistic watermarks strip out instantly.
For institutional compliance and systemic risk mitigation, enterprise governance cannot rely solely on generative word-choice watermarks; it requires immutable cryptographic provenance at the state-transition and tool-execution layer. Outstanding overview of where regulatory compliance meets frontier engineering reality.
The watermarking date has two deferrals sitting behind it that the coverage rarely names. Article 50(2) applies from 2 August 2026, but the Digital Omnibus pushed marking and detection for generative systems already on the market to 2 December 2026, and watermark-detection interoperability to 2 February 2027. Does the transparency Code say what a deployer is expected to check during that gap?
Starting with August 2, everything is different. Authorities will gave to act and market to comply. I am not sure that both sides are ready for this.
Watermarking is the pleasant part. Brussels has now sent its first AI Act letters on security, safety and copyright, which is where the transparency talk stops being theoretical.
The watermarking debate is framed as a provider question, which is right for Article 50(2). The gap I keep meeting sits one layer down, with deployers.
Marking attaches to design and development. Disclosure under Article 50(4) attaches to the organisation running the system, and for text it turns on whether the publication informs the public on a matter of public interest and whether the human review and editorial responsibility carve-out is available.
Two practical consequences for bodies that only deploy. A marking capability that ships enabled can be switched off during configuration, and nothing in the provider's compliance story surfaces that. And the deployer's own position rests on a documented review step and a named responsible person, not on the provider's marking.
The question worth putting to suppliers in writing: may the marking capability be disabled, and is any change to it logged.
Software vendors operating in the European healthcare space must design watermarking and audit compliance into their core architectures to avoid substantial non-compliance penalties under the EU AI Act.
The 200-token threshold is the useful number here, since a lot of short-form output sits under it. Worth pairing with it: under the Omnibus, systems already placed on the market before 2 August 2026 aren't covered by the Article 50(2) machine-readable marking duty until 2 December 2026, so for the next few months an unmarked output tells you very little either way. Does the transparency code say anything about that gap?
On the gap itself, the Regulation is the harder constraint and it is narrow in three ways worth keeping separate. Article 111(4) reaches providers only, it covers the Article 50(2) marking duty only, and it applies only to systems placed on the market before 2 August 2026. Paragraphs 1, 3, 4 and 5 are untouched by it.
So for the next few months an unmarked output tells you little about the provider, as you say, but it tells you nothing about the deployer. A body publishing AI-generated text on a matter of public interest sits under Article 50(4) now, and its position turns on a documented review step and a named responsible person rather than on whether the output carries a mark.
I have not read the Code closely enough to say whether it addresses the transitional window directly, so I would not assert either way. What is clear is that adherence is voluntary and cannot move a date set in the Regulation.