Under the guise of transparency, a US bill entrusts the state with defining what constitutes AI-generated content, the form of its labeling, and the agency responsible for penalizing those who evade it. It is a framework that raises as many questions as it does reassurances.
One Label, Two Watermarks, One Agency
Senator Brian Schatz reintroduced his AI Labeling Act on June 25, with the support of co-sponsors John Curtis (Utah) and Mark Warner (Virginia). The text, a copy of which was obtained by Reclaim The Net, grants the Federal Trade Commission the power to treat any unlabeled AI image as an unfair or deceptive trade practice.
The proposed framework is not limited to a simple visual warning. Each covered piece of content would have to carry:
- a visible disclosure, readable by a human;
- a hidden machine watermark, recording the generating system, its version, and the time of creation.
Large platforms would then have to propagate this data, flag the content to users, and identify any conversational agent as artificial. Missing a step opens the door to FTC intervention.
A Definition That Stretches at Will
The core of the text rests on the concept of “covered AI-generated content”, a category whose scope remains deliberately vague: it covers any content that a generative system creates or substantially modifies, provided that the meaning is altered and a reasonable person would not assume machine intervention.
A working group convened by the National Institute of Standards and Technology (NIST) would have one year to determine, in technical detail, what this means in practice, which detection tools are authoritative, and what the hidden watermarks must contain. In other words: the state writes the definition, designs the label, and designates the body responsible for enforcing it.
“Turn speakers into government mouthpieces for views they may not hold” - this is how the FIRE (Foundation for Individual Rights and Expression) describes the text’s constitutional flaw.
From Judicial Precedent to Enforcement Weapon
The slide from transparency to compelled speech is not an abstract hypothesis. The ACLU has compared the requirement of a disclaimer on legal synthetic media to forcing a comedian to announce their parody before the punchline.
Courts have already ruled on a similar case: a federal judge blocked California’s law on election deepfakes in Kohls v. Bonta, ruling that its restrictions and disclosure requirements on political AI content violated the First Amendment. Satire, parody, and criticism of officials remain protected, even when the tool producing them is new.
The enforcement of the bill, however, is anything but symbolic:
- a ban on removing, falsifying, or distributing tools designed to bypass disclosures;
- legal action open to the Attorney General, state attorneys general, and private companies;
- statutory damages of up to $2,500 per act of circumvention, or $25,000 per violation, trebled in the event of a repeat offense within three years.
The bill promises no prior restraint on protected speech - but the mechanics deployed around it point in the opposite direction. Only the largest platforms, those with at least 10 million monthly users in the United States or $1.5 billion in revenue, fall into the net - entrusting compliance to the very actors who already decide what the majority of people see online.
The Anthropic Precedent: When the State Already Cuts Off Access
This bill comes in a context where the US executive branch has already demonstrated its ability to intervene brutally in AI systems. In June, Washington invoked national security to force Anthropic to suspend its most powerful models, just three days after their launch.
The injunction, falling under export controls, ordered access to be cut off “to any foreign national, inside or outside the United States.” Anthropic had to “brutally disable” its models for its entire customer base, due to an inability to isolate its users. The company is contesting the move:
“We dispute that the discovery of potential circumvention justifies the recall of a commercial model deployed to hundreds of millions of people” - Anthropic, quoted by Noovo Info.
The episode immediately revived, on the European side, the issue of technological sovereignty. The European Commission, through its spokesperson Thomas Regnier, acknowledged that the incident “further highlights Europe’s need for technological sovereignty,” recalling that Brussels has just unveiled measures aimed at reducing the 27 member states’ dependence on the United States and Asia for key technologies.
Two Continents, the Same Regulatory Drift
Anthropic CEO Dario Amodei has himself advocated for a regime of mandatory audits for the most powerful models, inspired by civil aviation - giving the state the power to block their deployment, provided this takes place within a transparent and fair legal process.
This tension crosses the Atlantic. On one side, a US Congress legislating on the labeling of AI content by entrusting a federal agency with defining the terms; on the other, a European Union which, against the backdrop of technological dependence revealed by the Anthropic affair, is also seeking to build its own control apparatus. In both cases, the same question remains: who decides what must be said, and by what criteria - a technical regulator, or a political authority dressing itself in the clothes of transparency?
The stakes go beyond the mere mechanics of watermarking. It touches on the ability of states to shape, through law, the very conditions under which public speech - human or artificial - can still circulate freely.
