The Fake Is Already Out There
Two videos. Neither is real. One shows Josh Shapiro discussing “deleting incriminating e-mails.” Another features Abdul El-Sayed alongside the words “DEATH TO AMERICA.” Both were generated by artificial intelligence. Both were made to look like something a voter might believe.
What These Videos Are, and Why They Work
The Shapiro clip and the El-Sayed clip are not fringe experiments buried in some dark corner of the internet. They are examples of a category of political content that is growing fast, getting cheaper to produce, and becoming harder to distinguish from authentic footage. The faces move correctly. The voices carry the right cadence. The visual grammar of a legitimate campaign video – direct address, clean framing, confident delivery – is fully intact. What’s missing is the truth.
AI-generated political ads occupy a specific and dangerous position in the media ecosystem. They are not satire, where exaggeration signals intent. They are not parody, where absurdity creates distance. They are designed to pass. The target audience is not someone primed to be skeptical – it is someone scrolling past quickly, half-paying attention, absorbing a clip the way most people absorb most things online: incompletely, in fragments, with a thumb ready to keep moving.
That is the mechanism. Not a sophisticated hack of perception, but an exploitation of ordinary inattention. A clip featuring a sitting governor appearing to discuss destroying evidence, or a public health figure framed alongside a phrase associated with terrorism, lands in the mind as something adjacent to memory – not quite believed, not fully dismissed, somewhere in the soft tissue where impressions form without consent.
The technology required to produce either video is not experimental. It is commercially available, actively marketed, and within reach of campaigns, political action committees, private citizens with grudges, and foreign actors with agendas. The cost of making a convincing deepfake has dropped far enough that the barrier to entry is no longer technical skill – it is only intent.
No Law Is Moving Fast Enough to Matter
There is no federal law in the United States that specifically prohibits AI-generated political deepfakes. A small number of states have passed legislation targeting synthetic media in election contexts, but the coverage is patchy, enforcement is largely untested, and the platforms where this content spreads operate under frameworks built for a different era of misinformation – one in which fabricated content was harder to make and easier to spot.
The major social media platforms have policies that nominally address synthetic media and manipulated video. In practice, those policies depend on detection, reporting, and human review at a scale that does not match the volume of content being uploaded every hour. By the time a flagged video is reviewed, assessed, and removed – if it is removed – it may have completed its work. Shares, screenshots, and secondhand descriptions travel independently of the original post. The video disappears; the impression does not.
This is the structural problem that no content moderation policy has yet solved: virality is faster than review. A clip of Josh Shapiro apparently referencing deleted e-mails does not need to stay live for a week to cause damage. It needs a few hours and an audience that is already inclined to distrust him. The AI does not need to convince everyone. It only needs to deepen an existing suspicion in enough people to shift a margin.
The El-Sayed example carries a different weight. “DEATH TO AMERICA” is not merely a political attack – it is an invocation of a specific fear with a specific racialized history in American politics. Placing those words alongside the image and voice of a Muslim-American public figure is not a neutral act of opposition research or negative advertising. It is an attempt to fuse a person’s identity with a phrase designed to trigger revulsion. The AI makes that fusion look like documentation.
What is particularly striking about both cases is how little production effort the deception requires compared to the potential damage. Traditional smear campaigns required resources – writers, actors, studio time, distribution budgets. A coordinated effort to shape public perception through fabricated video evidence now requires a subscription, a prompt, and an upload. The asymmetry between the cost of creating the content and the cost of correcting it is severe and getting worse.
The Voter in the Middle
The question hanging over all of this is the one the videos themselves are designed to make unanswerable: can voters tell the difference? Research on deepfake detection consistently finds that most people are not reliable at identifying synthetic video without assistance, and that accuracy drops further when the viewer already holds strong feelings – positive or negative – about the person depicted. Motivated reasoning and AI-generated imagery are a functional combination for anyone trying to manufacture belief.
Josh Shapiro is a real governor. Abdul El-Sayed is a real physician and former candidate. Their faces, their voices, and the texture of their public personas are now raw material available to anyone with access to a generative AI platform. They did not consent to appear in these videos. They have no reliable legal mechanism to stop the next ones. And the people who watched those clips – however many that was, wherever they watched it – are now carrying something they may never fully examine or correct. That’s where the story sits: not in the technology, not in the legislation, but in the unresolved question of what people now believe, and why.
