That Late-Night Clip Looks Real. But Is It? A Guide to AI Deepfakes.

The short version

AI-generated deepfake videos of celebrities like Jimmy Kimmel are getting shockingly realistic and spreading online. It’s becoming much harder to instinctively trust what you see in your feed. We all need to build new habits for spotting fakes and verifying information before we believe it or share it.

What’s actually happening with these deepfake videos?

I saw a clip the other day that stopped me cold. It was supposedly Jimmy Kimmel, on his show, saying something pretty outrageous about a current political topic. It looked real, it sounded real, and the comments were blowing up with people who were either furious or cheering him on. But something felt slightly off, and after a few minutes of digging, I found out the video was a complete fabrication—an AI-generated deepfake. And I’m not the only one getting fooled. This isn’t just a hypothetical problem; a recent report on deepfake videos from NPR detailed how clips impersonating hosts are already fooling people online. These aren’t like the clumsy, glitchy deepfakes from a few years ago. The technology has advanced so quickly that synthetic videos can now mimic a person’s face, voice, and mannerisms with terrifying accuracy. The clips are often taken out of context, posted without any disclaimer, and engineered to go viral on platforms like TikTok, X (formerly Twitter), and Facebook, racking up millions of views before they can be debunked.

Why does this matter so much?

The immediate problem is misinformation. A fake video of a trusted public figure, whether it’s a talk show host or a political leader, can instantly inject false narratives into the public conversation. Because late-night shows often blend comedy with serious commentary on current events, a fake clip can feel plausible. People are primed to believe a host might make a controversial joke or a sharp political point. When a deepfake lands in that context, it’s like a spark in a tinderbox. The video spreads, outrage builds, and the truth has to play catch-up. By the time a correction is issued, the damage is often done. Millions of people have already seen and formed an opinion based on a lie.

But the bigger, more corrosive danger is the erosion of shared reality. When we can no longer trust our own eyes and ears, what can we trust? This phenomenon is sometimes called the “liar’s dividend.” In a world saturated with convincing fakes, bad actors can dismiss real, inconvenient videos as deepfakes. It gives everyone a get-out-of-jail-free card for their own bad behavior. A politician caught on a hot mic can just claim the audio was AI-generated. A CEO whose company is exposed in an undercover video can sow doubt by calling it a synthetic fabrication. This makes holding people accountable incredibly difficult and poisons the well of public discourse. We lose the ability to agree on a basic set of facts, which is the foundation of any functioning society.

How are these AI deepfakes even made?

It feels like magic, but the process behind creating these deepfakes is surprisingly straightforward to understand, even if the technology is complex. At its core, it’s about machine learning. Developers use a type of AI model, often a Generative Adversarial Network (GAN) or a diffusion model, and feed it massive amounts of data. To create a deepfake of Jimmy Kimmel, for instance, you would need hundreds or thousands of hours of video of him from his show. The AI studies everything: how his face moves when he talks, his specific smile, the way he blinks, his vocal patterns, and his unique gestures.

The AI model essentially has two parts that work against each other. One part, the “generator,” tries to create new, fake images or video frames of Kimmel. The other part, the “discriminator,” acts as a detective. Its job is to look at the generator’s work and decide if it’s real or fake. At first, the generator is terrible, and the discriminator easily spots the fakes. But every time the discriminator catches a fake, the generator learns from its mistake and gets a little better. This cycle repeats millions of times. The generator gets better at faking, and the discriminator gets better at detecting. Eventually, the generator becomes so skilled that its creations can fool not just the discriminator but human eyes, too. The same process is applied to audio, with AI models training on a person’s voice to create synthetic speech that can say anything you type.

Why are late-night hosts a perfect target?

It’s no accident that people are creating deepfakes of figures like Jimmy Kimmel, Stephen Colbert, or John Oliver. There are a few reasons why they are ideal targets. First, there is an enormous amount of high-quality training data available. These hosts have been on television for years, almost every weeknight, filmed in consistent lighting with high-definition cameras. An AI has a massive library of their faces, voices, and mannerisms to learn from, making the final deepfake much more convincing.

Second, the format of their shows provides a perfect, believable container for fake content. We are all used to seeing these hosts sitting at a desk, delivering a monologue to a camera. A deepfake that mimics this exact format feels instantly familiar and credible. The audience’s expectation is already set. Finally, their role as social and political commentators makes them powerful vectors for misinformation. A fake clip of a host endorsing a candidate, spreading a conspiracy theory, or making an inflammatory statement is far more impactful than a similar clip of a random actor. People trust these hosts to be sharp, witty, and often, to speak truth to power. Hijacking that trust is a powerful shortcut to spreading a specific agenda.

How can you tell if a video is a deepfake?

While the technology is getting better, most deepfakes still have subtle flaws if you know what to look for. It requires a new kind of media literacy—a healthy, active skepticism. Here are a few things I’m training myself to watch for:

1. Eyes and Blinking: Humans blink regularly and naturally. AI models often struggle with this. Look for a person who isn’t blinking at all, or who is blinking too much or in a weird, unnatural rhythm. The eyes themselves might also look a little glassy or not quite focused correctly.
2. Facial Movements and Expressions: Pay close attention to the mouth. Does the lip-syncing match the audio perfectly? Sometimes the edges of the mouth can appear blurry or unnaturally smooth. Also, watch how emotions play across the face. A real human smile involves the eyes (the “Duchenne smile”), but a deepfake might only move the mouth, leading to an empty or creepy expression.
3. Skin, Hair, and Edges: Look for weird artifacts. Sometimes the skin texture appears too smooth, like it has a digital filter over it. The edges where the face meets the hair or neck can be a giveaway. You might see some blurring, discoloration, or a strange, wobbly effect as the person moves their head.
4. Audio Quality: Listen closely to the voice. Does it sound a little robotic or flat? Is the cadence or emotional inflection slightly off? Sometimes there are weird digital artifacts or background noises that don’t match the environment. A deepfake might nail the face but fall short on the voice.
5. Check the Source: This is the most important step. Before you react or share a video, ask yourself where it came from. Was it posted by a reputable news organization or the official account of the show? Or did it come from a random, anonymous account with a weird username? Do a quick search to see if any trusted sources are reporting the same story. If they aren’t, it’s a massive red flag.

FAQ

What’s the easiest way to spot a deepfake? Look for unnatural facial features, especially the eyes and mouth, and weird blinking patterns. But the most reliable method is to always check the source of the video—if it’s not from an official or trusted account, be extremely skeptical.

Are deepfakes illegal? It’s complicated and depends on how they are used. Creating a deepfake for parody might be protected speech in some places. However, using them for fraud, defamation, harassment, or election interference is illegal in many jurisdictions, and laws are rapidly evolving to address this threat.

Will AI tools get better at detecting deepfakes? Yes, AI-powered detection tools are constantly improving. However, the technology to create deepfakes is also getting better at the same time. This has created an ongoing arms race between creation and detection, so technology alone is unlikely to be a perfect solution.

Is all synthetic media bad? No, not at all. AI-generated media has positive uses in film (like de-aging actors), accessibility (creating custom voices for people who can’t speak), art, and entertainment. The issue isn’t the technology itself, but its use with malicious intent to deceive or harm.