A fake clip of a candidate can pull millions of views, get fact-checked within hours, and still change almost nothing about how people vote. The one kind of fake that does move people is the kind almost nobody is worried about.
That is not a hunch. It is what the largest cross-platform study of a real election found, after researchers at Mila, the University of Montreal, and McGill University tracked every election-related image shared during the 2025 Canadian federal election.
So what actually happens when a fake candidate clip goes viral? Usually this: it spikes fast, it reaches an audience that was already on its side, corrections travel about as far, and the persuasion washes out. The cases that matter are narrower, and stranger, than the headlines suggest.
Here is the evidence, myth by myth.
Myth 1: A viral deepfake can flip an election
Reach is real, but it is lopsided, brief, and mostly aimed at people who already agree.
The Mila-led team analyzed 187,778 posts across X, Bluesky, and Reddit during the week around the April 28, 2025 Canadian vote (April 24 to April 29). Their result, first posted in December 2025 and published at the ACM Web Conference 2026: 5.86% of election-related images were deepfakes.
That number sounds alarming until you follow where the fakes actually went:
- Political deepfakes accounted for just 0.52% of all views on X.
- Politically harmful deepfakes, the defamatory or hateful ones, accounted for 0.12% of views.
- Deepfake posts drew fewer views than ordinary posts across every political leaning the study measured.
- The accounts sharing the most fakes had small followings, not large ones.
In other words, a fake can trend inside a niche and still be invisible to the electorate at large. Volume is not influence.
One more detail undercuts the conspiracy version of events: fake and real posts were removed at nearly identical rates (13.8% against 13.0%). The big platforms were not quietly deleting deepfakes while leaving everything else standing. The material simply did not travel.
Myth 2: We are drowning in a flood of fakes
Fakes are common. Harmful fakes are not.
Of the deepfakes the study detected, about half (49.41%) were benign or non-political: memes, satire, artistic edits, logo swaps. On Bluesky and Reddit, most synthetic images had nothing to do with politics at all. X hosted more of the defamatory and conspiratorial material, which fits that platform’s reputation.
Broken down by platform: X was highest, at 7.9% of election images detected as synthetic, then Reddit at 3.3% and Bluesky at 2.2%. Harmful material clustered on X, while Bluesky skewed benign and non-political.
The authors sum it up plainly: synthetic images were present in the election conversation, but their reach was modest.
One caveat cuts against complacency. The researchers say their detector favors precision, so the true prevalence is probably a little higher than reported. Right-leaning accounts shared more of it (8.66% of their images flagged, against 4.42% for left-leaning users), and that pattern held. But treat the headline numbers as a floor, not a ceiling.
Myth 3: Deepfakes fool everyone
People are better at this than the panic implies, and so are the machines. Both still fail in ways that matter.
A 2022 study in PNAS (Groh and colleagues) showed 15,000 participants real videos and deepfakes and asked them to pick out the fake. Ordinary people matched leading computer vision models. Many deepfakes are simply badly made.
But the same study found pairs where fewer than 65% of people could tell which clip was AI. And detection has a ceiling:
- Detectors trained on today’s fakes get weaker against tomorrow’s generation methods.
- Small, invisible edits can trick computer-based detectors, a documented class of adversarial attack (CVPR 2022).
- In the Canadian study, the detector was 99.2% accurate on real images but only 77.9% on fakes.
The honest takeaway: no tool is a silver bullet, and a clip that is not flagged is not proof that it is real.
Myth 4: The biggest danger is a fake video people believe
The bigger documented danger is a real video that people are told is fake.
Law professors Bobby Chesney and Danielle Citron call this the liar’s dividend. The logic is simple: as the public learns that deepfakes are realistic, a false claim that genuine footage is AI-generated gets easier to sell too.
It is not hypothetical:
- A lawyer for January 6 defendant Guy Reffitt argued in court that the prosecution’s video evidence was deepfaked.
- Tesla’s lawyers argued that Elon Musk’s own past statements should be kept out of court because they could be AI-generated.
- In 2023, an Indian politician insisted that authentic audio of him was AI-generated, even after research teams verified it.
Then there is the specific warning buried in the Canadian data. The rare “fabricated” deepfakes, the ones built to look like real news photos, drew more engagement than almost any other category. Low volume, high credibility. That combination is the one to watch.
Myth 5: Nobody is doing anything about it
Enforcement is uneven, but it is no longer zero.
- September 2024: the US Federal Communications Commission fined political consultant Steve Kramer $6 million for robocalls that used an AI copy of President Biden’s voice before the New Hampshire primary. The carrier behind the calls, Lingo Telecom, settled for $1 million.
- August 2, 2026: the EU AI Act’s transparency rule (Article 50) took effect. Anyone publishing realistic AI-generated images, audio, or video for an EU audience now has to label them, and that covers ordinary campaign content, not just paid ads.
- In the US: there is still no federal law aimed specifically at election deepfakes. It is a patchwork of state statutes with different thresholds and timelines, and some of those laws target robocalls and ads rather than organic posts.
The practical rule for 2026 is not “trust the label.” It is “check the source, then check who benefits.”
What changed: the tools moved into the apps
The 2025 study flags a shift that makes the next cycle different from the last one. Image generation is no longer a separate tool you go and find. It now sits inside the apps people already use: Grok inside X, Gemini on Android, ChatGPT in a browser tab. The gap between imagining a fake and posting one has collapsed to seconds.
That cuts both ways. It makes more fakes possible, and it makes more of them obvious and forgettable, which is exactly what the 2025 data captured. A cheap flood of low-quality fakes erodes trust without moving votes. The danger stays concentrated in the small number that look real.
What to watch before the 2026 midterms
The US midterms are the next big test. The 2025 pattern suggests what to expect: a steady hum of AI-generated political images, mostly benign or aimed at the already-convinced, plus a small number of realistic fakes timed for a single moment. Add a rising drumbeat of politicians dismissing real footage as fake.
Three things are worth your attention:
- Fabricated news-style images, not jerky face swaps. They are rarer, and they travel further.
- The “it is fake” defense. When a candidate calls real audio fake, the story becomes the denial, not the clip.
- Timing. A fake audio clip of Slovak leader Michal Simecka appeared two days before the September 2023 vote. That late window is when a fake does its damage, before anyone can debunk it.
None of this means deepfakes are harmless. It means the threat is specific, and the loudest version of it is usually the wrong one.
The detection problem is bigger than deepfakes
We wrote a whole post about AI detectors that fail in both directions, flagging real work as fake and missing real fakes. Same lesson, different room.
The Short Version
- Best evidence so far (2025 Canadian election, 187,778 posts): 5.86% of election images were deepfakes, but harmful political fakes drew only 0.12% of views on X.
- The rare, realistic “fabricated” fakes got more engagement than most political fakes. That is the growing risk.
- Detectors hit 99.2% accuracy on real images and 77.9% on fakes. Humans match machines on easy cases and fail on hard ones.
- The larger threat may be the liar’s dividend: real footage dismissed as AI.
- Enforcement now exists: a $6 million FCC fine in 2024, EU labelling from August 2, 2026, and no federal US election-deepfake law.
FAQ
Can a deepfake actually change an election result?
No study has shown a deepfake flipping a national outcome. Research shows a real but modest persuasion effect that fades, with risk concentrated in rare, highly credible fakes and in very close races. Treat any claim of a decisive deepfake with suspicion until it is sourced.
Are deepfake detectors good enough to trust?
No. They score well on known datasets and worse on new generation methods. Use them as one signal, not a verdict. A clip that is not flagged is not proven real.
What should I do when I see a shocking political clip?
Pause before sharing. Find the original source. Check whether major outlets and fact checkers have covered it. Look at who shared it first and what they gain. If the only version is a screen recording with no original, that is a warning sign.
Is AI-generated political content illegal?
It depends where you are. In the EU, realistic AI political content must be labelled as of August 2, 2026. In the US, there is no federal election-deepfake law, so rules vary by state, and some cover robocalls and ads rather than ordinary posts.
About The Author
