33 states now regulate election deepfakes, with five different rules

American voters heading into the 2026 midterms are protected against AI-generated political fakery by a set of rules that changes at every state line, and in two states by nothing at all.

As of 5 August 2026, 33 states had enacted laws specifically covering deepfakes in political communications, up from 28 at the start of the year. There’s still no federal baseline.

Disclosure, not prohibition

Most of these laws take the same basic approach, which is to require a label rather than ban the content.

The typical requirement is a disclaimer along the lines of “This ad was generated or substantially altered using artificial intelligence”, displayed on the advertisement itself, according to MultiState’s tracking of the 2026 laws.

The bet behind disclosure law is that voters who know a video is synthetic will discount it. That bet has never been tested at scale.

Some states go further on the detail. Colorado and Utah require disclosure of who created the deepfake, when it was made and how it was edited, and Colorado and Washington also require the label in the content’s metadata rather than only onscreen.

Where election deepfake laws diverge sharply

Approach	Example states	What it means in practice
Disclaimer required	Most of the 33	Label it and you may run it
Metadata label too	Colorado, Washington	The file itself must carry the mark
Banned near an election	Minnesota, Texas	A window before polling day
Banned year-round	Maryland	No synthetic political content at all
Law struck down	California, Hawaii	No election-specific AI rule in force

That table describes a genuine operational problem rather than an abstract one. A campaign buying advertising across state lines has to check the rule for each jurisdiction the ad reaches.

Broadcast and streaming make it worse, because a signal doesn’t respect state boundaries. Wiley’s guidance for advertisers walks through the disclaimer and ban variations that campaigns have to reconcile.

The First Amendment problem

Two of the most aggressive state laws are no longer in force, and the reason matters more than the outcome.

California’s and Hawaii’s deepfake laws were struck down on First Amendment grounds, and both states are now pursuing approaches that don’t specifically regulate election AI content, Axios reported.

Political speech gets the strongest constitutional protection in American law, and a statute regulating what a campaign may say about an opponent runs directly into that. Satire and parody complicate it further, since both are protected and both are synthetic by definition.

Which explains why disclosure became the dominant model. Requiring a label is a lighter touch than banning content, and lighter touches survive judicial review more often.

The pattern is worth noting for anyone reading this as a policy question rather than a compliance one. The strictest laws are the most likely to be struck down, so the surviving regime tends toward the weakest common approach.

Enforcement is the untested part

Thirty-three laws on the books tells you very little about what happens when one is broken.

The 2026 midterms are the first cycle where these statutes face real pressure, as Arizona Capitol Times framed it, and the mechanics look difficult.

Attribution is the first obstacle. A disclosure requirement binds whoever publishes the ad, and anonymous content circulating on social platforms has no publisher to bind.

Timing is the second. A fake released 48 hours before polling day has done its work before any complaint reaches a court, and a penalty imposed weeks later doesn’t unwin an election.

Proof is the third. Establishing that a video was AI-generated rather than merely edited is a technical question, and courts working to an election calendar have days rather than months to resolve it.

Those three together mean the practical effect of most of these statutes may be deterrence among campaigns that were already going to comply, which is not nothing but is considerably less than the headline count implies.

Regulators are visibly working this out in real time, with reporting on the current cycle describing scrambling rather than settled procedure.

Where the transparency approach is going instead

The metadata requirements in Colorado and Washington point at a different and more durable mechanism than onscreen labels.

If provenance information travels with the file, a platform can check it automatically rather than relying on a viewer to read fine print. That’s the same logic Europe applied through the AI Act, whose transparency provisions took effect on 2 August 2026.

The industry response was immediate there. Anthropic began watermarking Claude output worldwide within nine days, and our explainer on what the Act requires covers the scope.

Marking at generation is more robust than marking at publication, because it doesn’t depend on the person distributing the content wanting to comply. It also doesn’t help against tools that decline to mark anything, which is most open-weight ones.

The argument against doing more

It’s worth taking the sceptical case seriously, because it isn’t only a free-speech argument.

Misleading political advertising predates AI by a century, and existing law on defamation and election fraud already reaches the worst conduct regardless of the tool used to produce it.

There’s also a risk the label itself becomes a weapon. If synthetic content carries an official mark, a candidate caught on genuine video gains a ready defence, which is a harm the disclosure regime creates rather than prevents.

The counter is simply scale. Producing a convincing fake video used to require a studio and now requires a prompt, and rules written for an era of expensive forgery do not obviously transfer.

What to watch on election deepfake laws this cycle

The first enforcement action under one of these statutes will tell you more than the count of states that have passed them, and trackers of federal and state deepfake law are the place to follow it.

Watch the constitutional challenges too. If more disclosure laws survive review while bans keep falling, the national picture settles into labelling by default, which is what the current trend suggests.

The contrast with a single binding deadline is instructive, since the EU Transparency Code changed a frontier model in nine days where voluntary labelling had changed very little in years.

And watch whether Congress moves at all. A federal baseline would replace 33 different answers with one, and until it exists the compliance burden falls hardest on smaller campaigns without legal staff, the same asymmetry our piece on the state patchwork describes elsewhere in AI law.