The Disclosure Mandate: How a Bipartisan Senate Bill Would Put Every AI Pixel Under Federal Labeling Law

The Disclosure Mandate: How a Bipartisan Senate Bill Would Put Every AI Pixel Under Federal Labeling Law

S.4915 would require visible and machine-readable labels on AI-generated images, video, and audio — and make stripping them a federal civil violation. Here is exactly what the AI Labeling Act of 2026 demands, who it covers, and where it can still break.


The Bill in One Paragraph

On June 24, 2026, Senators Brian Schatz (D-HI), John Curtis (R-UT), and Mark Warner (D-VA) introduced the AI Labeling Act of 2026, a bipartisan bill that would require providers of generative AI systems to attach both a visible, human-readable disclosure and a machine-readable provenance record to the image, video, and audio content their systems produce, and require major online platforms to preserve and display those disclosures as content travels across the internet. The bill treats violations as unfair or deceptive acts under the Federal Trade Commission Act, meaning the FTC would enforce labeling failures with the same toolkit it already uses against false advertising. The bill was referred to the Senate Commerce Committee, where it awaits action.[1]

Who Has to Label What

The bill imposes duties on three distinct actors, and the division of labor is the core design choice of the legislation. Rather than one sweeping mandate aimed vaguely at “AI,” it assigns each link in the content chain a job matched to what that link can actually control.

  • Providers of generative AI systems must label covered AI-generated content with a clear and conspicuous disclosure embedded in the content itself, and bind a machine-readable disclosure into the file that identifies the content as AI-generated, records the system and version used to create it, and timestamps the creation. The provider must also ensure users and platforms can detect the label “without undue financial burden” using widely available detection tools, or else supply detection tools itself. Content generated solely for internal research and development that is not intended for public release is exempt.[1]
  • Covered online platforms — public-facing forums for user-to-user content sharing, such as social networks, search engines, and content aggregators with at least 10 million U.S. monthly users or more than $1.5 billion in annual gross revenue — must conspicuously identify AI-generated content that carries the machine-readable disclosure, must not strip or tamper with the disclosure when content is shared or transferred to another platform, and must make a good faith effort to combat what the bill calls the “liar’s dividend.”[1]
  • Anyone operating an AI chatbot must include a clear and conspicuous disclosure that the system is an AI chatbot — defined to include systems that publicly post content through an API or integrate with a search engine to provide conversational search.[1]

The logic behind this structure is that provenance information is most reliable at the moment of creation, so the obligation starts with the system that makes the content, and it is most useful at the moment of consumption, so the platforms that display content carry the duty to preserve and surface it. The threshold for platforms — 10 million U.S. users or $1.5 billion in revenue — deliberately concentrates the burden on the handful of services where synthetic content actually reaches mass audiences, rather than scattering compliance costs across every small forum and app on the internet.[1]

The Scope Question: What Counts as “Covered”

The bill’s definition of “covered AI-generated content” does two pieces of work at once. First, it requires that the AI’s use materially alters, adds, or removes the meaning or significance a reasonable person would take from the content. Second — and this is the scope-limiter — it requires that “a reasonable person would believe that the content is not generated using a generative artificial intelligence system.”[1] Read together, these two prongs mean the mandate does not reach content where AI played a merely assistive role, or content whose synthetic nature is obvious on its face. An AI-touched product photo in a listing, or a visibly stylized AI illustration, would generally fall outside the requirement because neither deceives a reasonable viewer about its origins.

Notably, the bill covers digital content defined as “an image, video, or audio content, or any combination thereof” — text is pointedly excluded from the mandatory labeling scheme.[1] The NIST-convened working group is instead directed to develop recommendations for provenance practices for content not covered by the mandate, including text.[1] The exemption of text is a deliberate concession to both practicality and politics: labeling every AI-assisted sentence would be unenforceable at internet scale, and would collide head-on with newsroom and publishing norms — a fight the sponsors chose to defer rather than lose the bill over.

Enforcement: FTC Act Plus Damages

The bill routes enforcement through two separate mechanisms, each aimed at a different kind of violation.

  • FTC enforcement for labeling failures. A violation of the disclosure requirements is treated as a violation of a rule defining an unfair or deceptive act or practice under Section 18(a)(1)(B) of the FTC Act, giving the Commission the same jurisdiction, powers, and duties it holds under the broader consumer protection statute. The FTC may also promulgate regulations specifying exceptions — such as for de minimis content — and may establish safe harbors that deem compliant any system following Commission-specified interoperable standards.[1]
  • Civil damages for tampering. Circumvention is a separate offense under Section 3, which prohibits knowingly falsifying, removing, or adding disclosures with intent to deceive, distributing unlabeled covered content for financial benefit, and manufacturing or trafficking in tools primarily designed to strip or fake labels. The Attorney General may sue under Section 4, state attorneys general may sue under Section 5, and providers and platforms harmed by circumvention may sue under Section 6. Statutory damages run up to $2,500 per act of circumvention, or up to $25,000 per violation, with repeat violations within three years subject to triple damages.[1]

The two-track design reflects a judgment about which failures matter most. A provider that forgets to watermark a model’s output is a compliance problem, well-suited to the FTC’s administrative toolkit and its safe-harbor process. A tool sold specifically to strip labels off deepfakes is a deception business, and the damages provisions — with their triple-damages multiplier for repeat offenders — are aimed at making that business unprofitable. The bill also includes carve-outs: good-faith researchers may circumvent disclosures to test their robustness, and nonprofit libraries, archives, and educational institutions are exempt from the anti-circumvention prohibitions unless they act for commercial advantage or financial gain.[1]

The First Amendment Carve-Out

Section 4(b) instructs courts that in any civil action, they “may not impose a prior restraint on free speech or the press protected under the First Amendment to the Constitution of the United States.”[1] The clause is a structural acknowledgment of the constitutional environment this bill must survive: earlier this year, federal courts struck down state deepfake disclosure laws in Hawaii and California on free speech grounds,[2] and the sponsors know a labeling mandate that reached newsroom workflow would face the same fight. The carve-out limits enforcement to after-the-fact remedies — damages and injunctions — rather than pre-publication censorship, which is where First Amendment doctrine is most ruthless against the government.

But the carve-out also exposes the bill’s central tension. If a provider’s generative system produces unlabeled content that a platform then displays, who is the speaker the government is restraining? The bill answers by aiming at conduct — the act of falsifying provenance — rather than at content, which is the same move the trademark and fraud doctrines make to stay inside the First Amendment. Whether courts accept that framing for a statute that functions, in practice, as a compulsory speech regime for software outputs is the open question that will determine the Act’s durability.

The Liar’s Dividend Provision

The bill’s most unusual feature is that it legislatively defines the “liar’s dividend” — the benefit a bad actor gains “by falsely claiming that non-AI-generated content is covered AI-generated content” — and requires covered platforms to make “a good faith effort to combat” it using strategies recommended by the FTC.[1] The term, coined by law professors Bobby Chesney and Danielle Citron to describe how the mere existence of realistic deepfakes lets political figures dismiss authentic evidence as fake,[3] has entered statute as a named harm the platform layer must actively fight.

The provision matters because labeling regimes create a second-order risk: once “AI-generated” becomes a legally meaningful badge, the badge itself becomes a weapon. A screenshot stripped of its metadata can be declared synthetic; genuine footage can be labeled fake. The bill’s answer is to treat false claims of AI provenance — both directions — as actionable deception, and to push the platforms that display content to maintain the integrity of provenance information as it moves. Whether “good faith effort” is a standard anyone can actually enforce is another question; it is the bill’s softest obligation, and likely the first one tested.

Standards and the Working Group

Within 90 days of enactment, the NIST Director, in coordination with the FTC, must establish an AI-generated content consumer transparency working group whose membership spans federal agencies, AI developers, standards bodies, social media platforms, academics, privacy advocates, media organizations including news publishers and image providers, digital forensics and cryptography experts, UX designers, consumer psychologists, and victims’ groups.[1] The working group must publish technical standards, guidelines, and recommendations within one year of its establishment, and the NIST Director must report to Congress within 180 days after publication. The working group sunsets 60 days after that report.

The breadth of the membership list is the point. The bill’s drafters understand that a labeling standard is only as strong as its adoption — a watermark that only half the industry implements and only some platforms honor is worse than none, because it teaches consumers to trust a signal that is unreliable. The working group is the mechanism for building the interop that the bill itself cannot legislate: standards that survive transfers between platforms, operating systems, and devices. Media organizations and news publishers get explicit seats because the most contested boundary in the entire regime is where AI assistance in news production becomes a disclosure obligation — and because journalism has the strongest legal position to fight a mandate it rejects.

The Global Context: Catching Up Versus Leading

The bill arrives into an already-moving international landscape. Senator Warner framed the stakes directly: “Jurisdictions around the globe are moving to adopt common-sense rules around labeling AI-generated content – it’s time for the U.S. to catch up and in fact lead the world with a disclosure and anti-circumvention model that should be the global standard.”[4]

The benchmark the bill is chasing is the European Union’s AI Act Article 50, whose transparency obligations took effect on August 2, 2026. Article 50 requires providers of systems that interact directly with people — chatbots, virtual assistants, automated phone systems — to design them so users are informed they are interacting with AI, and requires providers of generative systems to mark synthetic audio, image, video, and text outputs in machine-readable formats detectable as AI-generated. It also obligates deployers of deepfakes to disclose that content has been artificially generated, and deployers publishing AI-generated text on matters of public interest to disclose its provenance unless the text has undergone human review and editorial responsibility. Draft guidelines confirm AI agents fall within the chatbot-disclosure rule.[5]

The comparison shows what the Senate bill deliberately leaves out. The EU requires disclosure of AI-generated text published on public-interest matters — the newsroom rule — while S.4915 excludes text from its mandatory scheme and punts provenance for text to the working group. The EU regulates the deployer — the publisher or producer using the system — while S.4915 concentrates duties on the provider and platform layers, leaving deployers untouched except through the general anti-circumvention provisions. Both approaches are disclosure-first; the difference is where each regime draws its perimeter around the content supply chain.

Who Is For It, and Who Is Quiet

The bill is endorsed by a coalition of fifteen organizations spanning creative labor and consumer protection: ASCRL, the Authors Guild, Common Cause, Consumer Action, the Consumer Federation of America, the Department for Professional Employees AFL-CIO, IATSE, Music Creators North America, the National Association of Voice Actors, the National Consumers League, Public Citizen, SAG-AFTRA, the Society of Composers and Lyricists, the Songwriters Guild of America, and the Writers Guild of America East.[4]

The composition of that list is the political story of the bill. It is dominated by unions and professional associations whose members are already being cloned, synthesized, and displaced — voice actors, screenwriters, songwriters, stagehands — alongside long-standing consumer-protection groups. What the endorsement list notably lacks is the major AI developers and the large platforms themselves, which have generally supported labeling in principle while opposing mandates in practice, and whose participation the bill would legally compel through the collaboration and platform duties. Public Citizen’s J.B. Branch captured the coalition’s framing: “In a world where seeing is no longer believing, transparency becomes the prerequisite for trust, accountability, and an informed public.”[4]

Notably, the same week the Senate bill landed, Representative Zoe Lofgren (D-CA) introduced the Voluntary Consumer AI Disclosure Pilot Act, which directs NIST to study when and how private-sector entities should disclose AI interaction — a deliberate alternative that would not mandate labeling at all. Lofgren’s statement made the philosophical divide explicit: “disclosures should be clear, useful, and appropriate for the product or service — not one-size-fits-all labels that individuals ignore or misunderstand.”[6] That the two bills share a party, a chamber topic, and a month — and point in opposite directions — is the clearest signal that mandatory labeling remains genuinely contested within the party that co-sponsors it.

The Enforcement Reality Check

Legal analyses of the bill place it within a broader congressional push: the same month it was introduced, the House Science Committee advanced ten bipartisan AI bills spanning research access, cybersecurity, workforce development, and transparency, and a comprehensive federal framework draft — the Great American AI Act — was circulating with its own transparency mandates. The AI Labeling Act, in this reading, is not an outlier but the leading edge of a congressional move toward mandatory disclosure after three years of voluntary commitments.[7]

But the technical criticisms of watermarking remain unanswered by the bill. Experts studying digital forgeries have long warned that visible watermarks play on people’s unawareness — “if they know they can, they will find a way to break it,” as deepfake researcher Siwei Lyu put it — and that a labeling regime can be exploited in both directions: stripped from real evidence, or stamped on genuine content to discredit it.[8] A senior Senate staffer involved in drafting watermarking legislation described the arms-race concern as “1,000% valid,” comparing it to Olympic doping: “now that I know that you’re looking for this drug, I’ll just take another drug.”[8]

The bill’s answer to the arms race is not to deny it but to structure around it. The anti-circumvention provisions make label-stripping itself the offense — shifting enforcement from detecting every fake to punishing the tools and services of circumvention. The working group is charged with developing tamper-resistant, tamper-evident, interoperable standards precisely because static watermarks lose. And the “liar’s dividend” duty acknowledges that the label’s value decays the moment bad actors learn to weaponize its absence. Whether that structure holds is the bet the Senate is making: that provenance can be engineered to be harder to strip than it is worth stripping, and that the FTC can make examples of the strip-for-hire economy before it industrializes.

What to Watch

  • Committee action. The bill sits in the Senate Commerce Committee, where it was referred on introduction day. No hearings are scheduled. Its predecessor, the AI Labeling Act of 2023 (S.2691), also introduced by Schatz in July 2023, never advanced beyond introduction in the 118th Congress.[9]
  • The House companion question. The 2024 House version, sponsored by Reps. Anna Eshoo (D-CA) and Neal Dunn (R-FL), drew early support from AI developers including Hugging Face, whose chief AI ethics scientist Margaret Mitchell said watermarking would “help the public gain control over the role of generated content in our society.”[10] Whether a Trump-era House takes up the 2026 Senate version is an open question.
  • EU implementation. The EU’s Article 50 regime has been live since August 2, 2026, with its Code of Practice on AI-generated content still being finalized. If the EU model proves workable at scale, it strengthens the case for S.4915; if it produces chaos, opponents of the U.S. bill will cite it as evidence that mandates fail in practice.[5]
  • The text question. The working group’s recommendations on text provenance will determine whether the bill’s deferral on text is a permanent carve-out or the first step toward newsroom disclosure rules. The EU’s text rule — with its human-review and editorial-responsibility exception — is the template advocates will push for.[1]
  • Standards convergence. Whether C2PA, SynthID, or another provenance standard becomes the de facto implementation of the working group’s mandate determines how much of the compliance burden lands on U.S. firms versus being imported from standards bodies they already participate in.[8]

References

[1] Congress.gov — S.4915, AI Labeling Act of 2026, 119th Congress, introduced June 24, 2026. [PRIMARY]

[2] ai-policy.org. “The Enforcement Gap: 33 State Statutes, a Parody War, and an Unpaid $6 Million Fine.” August 31, 2026. [SECONDARY]

[3] Brennan Center for Justice. “Deepfakes, Elections, and Shrinking the Liar’s Dividend.” [SECONDARY]

[4] Schatz, Brian. “Schatz, Curtis, Warner Introduce Bipartisan Legislation To Provide More Transparency On AI-Generated Content.” U.S. Senate Office of Sen. Schatz. June 25, 2026. [PRIMARY]

[5] “The EU AI Act’s Transparency Rules: A Practical Guide to Article 50.” EU Artificial Intelligence Act (artificialintelligenceact.eu). May 14, 2026. [SECONDARY]

[6] Kelley, Alexandra, and Edward Graham. “Tech bills of the week: Labeling AI-generated content; AI standards for private sector companies; and more.” Nextgov/FCW. June 26, 2026. [SECONDARY]

[7] Hecht, Alexander, and Erek L. Barron. “AI: The Washington Report — July 2026 Edition.” Mintz. July 8, 2026. [SECONDARY]

[8] “AI watermarking could be exploited by bad actors to spread misinformation. But experts say the tech still must be adopted quickly.” FedScoop. 2024. [SECONDARY]

[9] Congress.gov — S.2691, AI Labeling Act of 2023, 118th Congress, introduced July 27, 2023. [PRIMARY]

[10] Klepper, David. “New bipartisan bill would require labeling of AI-generated videos and audio.” PBS News / Associated Press. March 21, 2024. [SECONDARY]


AI automated and generated content. Human reviewed.

The Weekly AI Policy Brief

AI policy moves fast. Get the week’s developments, new trackers, and podcast episodes in one short email — free, ad-free, no spam.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *