AI Deepfakes Are Getting Dumb Enough to Be Dangerous: That’s the Real Problem

A fake celebrity does not need to survive inspection on a cinema screen. Sometimes it only needs to survive five distracted seconds on a phone, long enough for somebody to click, share or panic.

That is the uncomfortable question behind cheap AI deepfakes. We can spend all afternoon arguing about a strange ear while the link underneath the video is already doing its job.

Coffeezilla’s new Investigating AI Deepfakes explores scams, propaganda and harassment. In introducing it, he says he first attempted the project four years ago and now sees the threat differently. It is a useful starting point for a wider discussion: how much realism does a fake actually need to cause harm?

The fake does not have to fool everyone

Imagine an invented advertisement in which a familiar actor recommends an investment. The mouth movements are slightly wrong, the background is suspiciously smooth and the actor apparently has a new interest in guaranteed returns. You recognize the scam immediately. Someone else sees a trusted face before they notice any of those clues.

The existence of skeptical viewers does not cancel the second person’s loss. Nor does a visible flaw prove that the whole campaign failed. We should judge the risk by what the material persuades people to do, rather than whether it wins an award for visual effects.

This does not make realistic deepfakes harmless. Better imitations can create additional problems. It means imperfect material already deserves attention, without waiting for every generated hand to pass anatomy class.

A familiar voice can add urgency to an old scam

The US Federal Trade Commission warned in March 2023 that voice cloning could make family-emergency scams more convincing. A short recording can supply material for an imitation, while the story supplies the pressure: a relative is in trouble and needs money immediately.

The FTC recommends contacting the supposed caller through a number you already know, rather than trusting the voice alone. Demands for wire transfers, cryptocurrency or gift-card details are additional warning signs.

The fraud’s structure is familiar. AI can add an impersonation layer, but it does not remove the need to verify the request independently. An urgent voice is not an identity document.

There is also a difference between a convincing demonstration and proof that every fraud operation has replaced its staff with autonomous callers. That sweeping claim would need evidence about real deployments, costs and success rates. The narrower risk is serious enough on its own.

“It could be fake” can protect real wrongdoing

Deepfakes can damage trust in both directions. False footage can be accepted as real; genuine footage can be dismissed as synthetic.

Robert Chesney, Danielle Citron and Hany Farid discussed this in their 2020 analysis of deepfake risks. Chesney and Citron call the second problem the “liar’s dividend”: awareness of convincing fabrications gives people an easier excuse to reject authentic evidence. Their argument also recognizes the harm from simpler misleading edits, which do not require advanced AI.

That leaves us with an awkward task. Blind belief is dangerous, but automatic disbelief is not critical thinking either. “Everything is fake” sounds sophisticated until it becomes permission to believe whichever version was most comfortable in the first place.

A useful question is where a clip came from: the original uploader, the full recording, the surrounding event and independent reporting. Those checks ask about evidence beyond the pixels. Counting fingers alone will not do all of that work.

Fake intimate images cause real harm

Nonconsensual sexualized fabrications do not become harmless when a viewer knows they were generated. The person depicted can still face humiliation, harassment and repeated demands to explain or remove material they never agreed to create.

There is an especially cruel asymmetry here: producing and circulating another copy can be easier than persuading every recipient to stop sharing it. “But it isn’t really you” is a remarkably inadequate response to somebody whose face has been used to target them.

In the United States, the TAKE IT DOWN Act became law on May 19, 2025. It addresses specified nonconsensual intimate depictions, including digital forgeries, and establishes a removal framework for covered platforms.

The law gives covered platforms one year from enactment to establish the required notice process, making May 19, 2026 the deadline. Its framework requires removal within 48 hours after a valid request, alongside reasonable efforts to remove known identical copies. As of this article’s March publication, that implementation deadline is still ahead.

A reporting mechanism can help while still requiring effort from the person targeted. Its existence also cannot guarantee that every copy disappears from every service. Passing a law and ending an abuse are very different achievements.

We do not need a countdown to take this seriously

Claims that every video will become indistinguishable from reality within exactly twelve months make a striking ending. They are forecasts, not a settled deadline. Detection depends on the material, the viewer, the available context and the tools used to examine it.

The practical response does not require deciding when the last visual flaw will disappear. Verify unexpected requests through another channel. Look for the source of a sensational clip before forwarding it. Avoid circulating abusive imagery merely to announce that it is fake.

AI deepfakes can be dangerous while still looking a little ridiculous. The internet has never required a lie to be beautifully rendered before giving it a very large audience.