The Senate Judiciary Committee recently advanced the 2026 version of the NO FAKES Act. Although the bill is aimed at deepfakes, it sweeps in a great deal of lawful speech. The latest versions make only minor changes to earlier drafts, and it remains a drastic and dangerous assault on online expression.
The bill erects a sweeping licensing regime for AI-generated content that will stop malicious or unwanted deepfakes at the cost of preventing Americans from creating, sharing, or accessing a wide range of lawful speech. Several of the politicians championing the bill openly praise it as a tool for stamping out misinformation that embarrasses them — speech that is, in fact, constitutionally protected. And the costs of such an approach could give the government, the powerful, and the overly litigious significant powers to muzzle speech.
A Right to Publicity Approach to AI Likeness
NO FAKES gives each individual an intellectual property right to authorize the use of their voice or visual likeness in digital replicas — realistic, computer-generated audio and visual representations of a person, including content that materially alters real imagery or audio. This aggressively expands the state-law concept known as the right of publicity, a right that generally extends to celebrities, public figures, and commercial speech such as advertising. Under NO FAKES, every individual may license such replicas, and on death the right passes to heirs or other licensees for up to 70 years. The bill sets out detailed terms, from the management of minors’ rights to registration and directory requirements for post-mortem replication rights, along with special provisions for musical digital replicas.
Most importantly, NO FAKES creates liability for anyone who displays, distributes, transmits, or otherwise makes available an unauthorized digital replica — and for anyone who makes available a tool primarily designed to produce them. In plain terms, sharing any computer-generated likeness of a person without authorization, or building a tool capable of making one, becomes illegal.
The bill then carves out several exceptions. Some are entity-specific, covering internet service providers, nonprofit libraries, and educational institutions, but, most notably, NO FAKES creates general exceptions for “fleeting or negligible” uses; “bona fide news, public affairs, or sports broadcast or account”; and uses “consistent with the public interest in bona fide commentary, criticism, scholarship, satire, or parody.” As the bill’s sponsors are fond of pointing out, these are explicit protections for expression, but NO FAKES fails to protect various categories of lawful speech, and while also creating a new moderator’s dilemma, which means even “allowed” speech may find itself unwelcome.
Limiting Lawful Speech and Tools
The effect of NO FAKES’s framing is to narrow the range of lawful AI-generated speech, attempting to shrink the broad protections that the First Amendment and other laws otherwise provide for digital replicas.
While the listed exceptions cover many important areas of speech, they are framed as a finite, closed list, creating significant problems when speech falls outside these limited exceptions. As other scholars have pointed out, this restricted approach to protected speech will likely render illegal — when produced with AI — content that courts have already deemed constitutional.
The terms “bona fide” and “consistent with the public interest” impose further limits found nowhere in the First Amendment. Those who feel wronged by critical speech involving a digital replica of themselves can argue that it serves no public interest or was made in bad faith. Multiple senators have made this argument, with Sen. Peter Welch arguing in the Judiciary Committee markup of the bill that NO FAKES was needed to stop an AI ad attacking Vermont Representative Becca Balint because the ad “was not satire.” Those who think AI content is in bad faith or not in the public interest may well prevail, precisely because NO FAKES leaves determinations of what counts as legitimate criticism, parody, and satire on shaky legal ground.
The bill also explicitly outlaws AI generation tools designed to create unauthorized digital replicas. Any company or user offering a tool classified as “primarily designed” to make unauthorized replicas would immediately face ruinous liability. Imagine a tool that brings photos to life but doesn’t get explicit approval from everyone in the photo. Anyone included in one of these photos can file a notice against the tool and get it shut down.
Open-source projects and smaller companies — least able to absorb the added compliance and liability costs — will be hit hardest. Users, meanwhile, will face a far worse experience as liability-averse companies layer on restrictions and extra authorization steps.
In short, the law tries to turn content and activities that are currently legal and protected by the First Amendment into illicit categories of speech.
Notice, Takedown, and Collateral Censorship
These direct limits on expression, however, are not even the bill’s biggest threat to speech. NO FAKES also establishes a notice-and-takedown regime that invites massive collateral censorship. Crucially, by classifying the new likeness right as a form of “intellectual property,” NO FAKES routes around Section 230, which expressly excludes intellectual property claims from its protections — stripping platforms of the immunity that ordinarily lets them host user speech without bearing liability for it and reviving the very incentive to over-remove that Section 230 was designed to eliminate. A user, or someone acting on their behalf, may notify a service that unauthorized digital replicas of their likeness are circulating on its platform. The notice must identify the material, assert the sender’s right to that likeness, and state that the sender does not believe an exception applies — a low bar that amounts to one person’s word that the content should be removed.
Upon receiving notice, the platform must swiftly remove the identified content and deploy broad content filters to find and remove all other current and future content containing that replica. Failure brings steep statutory damages: $25,000 per work where the platform is deemed to have acted in good faith, rising to as much as $750,000 per work where it is deemed not to have made good-faith efforts. Courts may add other penalties and fees, including attorney’s fees, which are awarded to a prevailing plaintiff but not automatically to a prevailing defendant. As the Computer and Communications Industry Association (CCIA) has noted, because damages accrue on a per-work basis, a single service could face more than $1 billion in statutory liability from as few as roughly 1,300 covered uploads — a scale of exposure wildly disconnected from any plausible real-world harm.
Beyond massive penalties, NO FAKES gives platforms every incentive to remove content. The sheer number of requests that will be received, and the content caught up in filters, will make it impossible for platforms to meaningfully review notices. Just establishing that a piece of content is fake or materially altered will consume significant resources. Add in subjective exceptions where legally trained reviewers or advanced moderation tools will be expected to apply the law yet avoid liability, and every platform, but especially smaller ones, will have no choice but to blindly comply with notices. Little surprise, then, that studies show many companies remove content under existing notice-and-takedown regimes without assessing it at all.
Indeed, we know how such a system can play out and be abused because NO FAKES’ proposed procedure is very similar to the one established under the Digital Millennium Copyright Act (DMCA). The DMCA already contains a narrower notice-and-takedown regime, and it is subject to significant abuse: studies find wide variation in false or improper reports, ranging from 5 percent to 70 percent with many ending up somewhere in the middle. Some abusive takedowns would be almost comical were the assault on protected speech not so troubling. Under Europe’s Right to Be Forgotten regime, Google’s data indicates that around 50 percent of requests are improper. Abuse and improper takedowns will likely be even worse under NO FAKES, given its broader scope and stronger incentives. Like DMCA, NO FAKES also has penalties for bad-faith notices, but clearly they are too ineffective to hold bad actors accountable.
As a result, NO FAKES could chill everything from a Saturday Night Live-style parody skit to an AI reenactment of an event, even if they are technically protected under the law.
Defenders of the bill point out that the newer version adds a counter-notification process, allowing those whose content has been removed to send the platform a counter-notice — signed before a notary — asserting that the content fits one of the exceptions to liability. Even then, a platform that receives a counter-notice may, but is not required to, restore the content; and if it does, the original claimant can force the content down again by bringing legal action, after which the disputed material must stay down until a court rules. But for an ordinary person or small organization, filing a notarized counter-notice for every takedown is both a cost and an inconvenience, with no guarantee that a platform will restore anything.
Attacking Anonymous Speech
NO FAKES also threatens free expression by making it far too easy to unmask, i.e., force platforms to identify anonymous speakers. While there is no uniform standard across the states for unmasking in civil cases, there are several prominent state standards as well as federal standards, and while different, each includes some combination of high evidentiary requirements and/or a free expression balancing test before a judge orders a platform to unmask a speaker.
NO FAKES, however, uses a far less stringent standard, borrowing the DMCA’s unmasking mechanism that was designed for commercial copyright disputes and applies it to a vast amount of speech. Any right holder may subpoena a platform to identify an anonymous user who posted what the right holder believes is an unauthorized digital replica. No evidence is required beyond the notice already provided to the platform: the right holder’s own view that the content violates the law.
A clerk, not a judge, then issues the subpoena with no consideration of the validity of the claim, nor any balancing test to assess the impact that unmasking would have. The result is that whistleblowers, critics of the US and foreign governments, advocates for people in abusive or dangerous situations, and many others can be unmasked with no opportunity to defend themselves and with no consideration for their rights or situation. While platforms may be able to push back against some intrusive subpoenas, the lack of protections will drastically chill anonymous speakers from using AI tools.
Protecting Americans and Their Speech
NO FAKES claims to protect Americans from the harms associated with false AI-generated content, but as I’ve argued previously, there are already ample laws on the books that address those harms. Instead, the bill takes a narrow concept of intellectual property rights drawn from commercial law and applies it broadly to all AI speech involving people, with no meaningful protections, producing a sweeping and unconstitutional assault on expression.
As I have warned in my prior research on AI, such regulation is all but guaranteed to silence Americans, crush innovation and competition, and do little to address any real harms perpetrated with AI.