Across social media, familiar faces are saying things they never said — and the gap between recognition and belief is narrowing dangerously. AI-generated deepfakes of late-night hosts like Jimmy Kimmel and Jon Stewart are proliferating with increasing ease, exploiting the very qualities that made these figures trusted: years of recorded footage, distinctive voices, and audiences conditioned to believe them. The concern is not merely about individual false videos, but about what happens to the idea of verifiable truth when the most recognizable voices in public life can be convincingly forged b
Late-night AI deepfakes spread online, raising concerns about misinformation
When everything can be faked, nothing feels provable.
Mark
Why are late-night hosts specifically easier targets for deepfakes than other celebrities?
Mimi
They have thousands of hours of footage available online—every episode, every appearance. Their voices are distinctive and well-documented. That abundance of training data is exactly what AI needs to work convincingly.
Luke
But the source material doesn't actually specify how much easier it is. It says they're easier to create, but there's no comparison—no numbers, no technical breakdown. We should be careful not to overstate that claim.
Mark
What makes a deepfake of Jon Stewart more disruptive than, say, a deepfake of a random actor?
Mimi
Trust. People watch these hosts regularly. They feel like they know them. When you see a video of someone you recognize and trust, your guard is lower. You're more likely to believe it's real.
Luke
That's reasonable, but again—the source doesn't actually measure the disruption. It says these deepfakes are "potentially more disruptive," which is a hypothesis, not a fact. We don't know yet how audiences actually respond to them.
Mark
How fast are these spreading?
Mimi
Fast enough that fact-checkers struggle to keep up. By the time a video is debunked, it's already reached thousands of people. The correction rarely catches everyone who saw the original.
Luke
The source doesn't give us numbers on spread or reach. We know they're spreading on social media, but we don't know the scale. That's important context we're missing.
Mark
What's the real danger here?
Mimi
It's about erosion of trust. If deepfakes become common enough, people stop believing video evidence altogether. That makes it harder to prove anything—including real misconduct.
Luke
That's a logical concern, but it's also speculative. We're describing a potential future, not something that's already happened. The source doesn't give us evidence that trust has actually eroded yet.
Le Pouls
- Deepfake videos of late-night hosts are spreading faster than platforms or the hosts themselves can respond, reaching millions before any correction lands.
- These particular fakes carry unusual credibility — viewers who have watched Kimmel or Stewart for years are primed to trust what they see, collapsing the instinct to question.
- The technical barrier to creating convincing deepfakes is falling rapidly, with user-friendly tools and freely available training footage making forgery accessible to nearly anyone.
- A corrosive paradox is emerging: as fakes multiply, healthy skepticism grows — but that same skepticism gives real misconduct a place to hide behind the alibi of 'it could be fake.'
- The damage is cumulative and asymmetric — debunking travels slower than the original lie, and the false version hardens into memory long after the correction has faded.
Across social media, familiar faces are saying things they never said — and the gap between recognition and belief is narrowing dangerously. AI-generated deepfakes of late-night hosts like Jimmy Kimmel and Jon Stewart are proliferating with increasing ease, exploiting the very qualities that made these figures trusted: years of recorded footage, distinctive voices, and audiences conditioned to believe them. The concern is not merely about individual false videos, but about what happens to the idea of verifiable truth when the most recognizable voices in public life can be convincingly forged by anyone with a laptop.
Somewhere on social media right now, Jimmy Kimmel is saying something he never said. The video looks right, sounds right, and the timing feels natural — because AI has had thousands of hours of footage to study.
Late-night hosts occupy a specific vulnerability in the deepfake landscape. Their voices, faces, and comedic cadences are exhaustively documented and freely available online, making them ideal training material. Someone with modest technical skill can now produce a convincing forgery of Kimmel or Jon Stewart far more quickly than they could for a less publicly visible figure.
What makes this especially disruptive is the trust these hosts carry. They are not merely entertainers — they are familiar presences that millions of viewers have invited into their homes night after night. When a deepfake of such a figure circulates, it arrives pre-loaded with credibility. The cognitive leap from 'I recognize this person' to 'this video is real' happens faster than it should.
Social media accelerates the damage. Algorithmic amplification spreads these videos before fact-checkers can respond, and by the time a debunking arrives, the false version has already settled into memory for many who will never see the correction.
The deeper threat is philosophical. As convincing deepfakes become commonplace, viewers may begin to doubt all video evidence — a rational response that nonetheless creates cover for real misconduct. When everything can be faked, nothing feels provable. The question is no longer whether these deepfakes will keep spreading. It is what becomes of public trust when they do.
Somewhere on social media right now, Jimmy Kimmel is saying something he never said. Jon Stewart is making a joke that never landed on his show. These videos look real enough to stop a scroll. They sound right. The timing feels natural. And they are getting easier to make.
AI-generated deepfakes of celebrities have become a fixture of the internet, but late-night television hosts occupy a particular vulnerability. Their faces are familiar to millions. Their voices have been recorded thousands of times. Their cadence, their timing, their way of moving through a joke—all of it is documented, indexed, available for algorithmic study. This abundance of source material makes them ideal subjects for deepfake creation. A person with basic technical knowledge and access to the right tools can now produce a convincing video of Kimmel or Stewart in a fraction of the time it would take to create a deepfake of someone less publicly visible.
The proliferation matters because these hosts are not just entertainers—they are trusted voices in the media landscape. When a deepfake of a late-night host circulates, it carries an implicit credibility that a deepfake of a random celebrity might not. Viewers know these people. They have watched them night after night. The cognitive distance between "this is a video of someone I recognize" and "this video is real" collapses faster than it should.
Social media platforms have become the primary distribution channel for these videos. They spread through shares, reposts, and algorithmic amplification before fact-checkers or the hosts themselves can respond. By the time a video is debunked, it has already reached thousands or millions of people. Some viewers will see the correction. Many will not. The false version will persist in memory, in screenshots, in the corners of the internet where misinformation settles and hardens into something that feels like truth.
The concern extends beyond individual videos. If deepfakes of recognizable public figures become commonplace enough, they erode the basic assumption that video evidence means something. A viewer sees a clip of a late-night host saying something controversial or damaging. Their first instinct, increasingly, might not be to believe it—it might be to wonder if it is real. That skepticism is rational. But it also creates space for actual misconduct to hide. When everything can be faked, nothing feels provable.
The technical barrier to entry continues to lower. Tools that once required specialized knowledge are becoming user-friendly. The computational power needed to generate convincing video is becoming cheaper and more accessible. The training data—hours of footage of these hosts—is freely available online. Each of these trends points in the same direction: more deepfakes, faster creation, wider distribution.
Late-night hosts are not the only targets, but they are among the most vulnerable because of the specific combination of factors that make them ideal subjects: massive amounts of high-quality video footage, distinctive voices and mannerisms, and an audience primed to trust them. The question now is not whether deepfakes of these figures will continue to spread. It is what happens to public trust when they do.
Envie de l'histoire complète ?
Lire l'original sur NPR ↗
Media & Entertainment