Deepfakes Are Getting Scary Good. But Are They Really Destroying Reality?

The alarm has been building for years: AI-generated deepfakes are getting so convincing that they will destroy our ability to trust anything we see or hear. Detection tools cannot keep up. Reality itself is at risk. The claim has real foundations, and also some significant blind spots.

The technology has genuinely advanced. Creating a convincing voice clone now requires as little as 20 to 30 seconds of audio, according to the World Economic Forum. Video deepfakes can be produced in under an hour using free software. A 2025 iProov study of 2,000 UK and US consumers found that only 0.1 percent could correctly identify all the fake and real media they were shown, while overall human detection accuracy hovered around 55 to 60 percent, barely better than guessing.

Detection tools are struggling to keep pace. The UK government's own assessment describes the detection market as "nascent," noting that systems typically lose 10 to 20 percent accuracy when deployed outside controlled lab conditions. The Columbia Journalism Review found that available tools are "easily circumvented" by basic countermeasures like adjusting lighting or applying filters, and that their probabilistic outputs (a tool reporting "70% human, 30% artificial") confuse rather than clarify verification work.

The documented harms are real. A Hong Kong finance worker transferred $25.6 million after a video conference call with what appeared to be colleagues and a company CFO, all of whom were deepfakes. Journalists have been targeted: Reporters Without Borders documented 100 cases between late 2023 and late 2025, with women accounting for 74 percent of victims.

Where the doomsday narrative starts to wobble is on elections. The fear of AI-generated disinformation swinging votes has dominated headlines for two years, but the evidence is messier. A Columbia University Knight Institute analysis of 78 election deepfakes found that 50 percent were not actually deceptive, and that traditional editing techniques (what researchers call "cheap fakes") were seven times more common than AI-generated content in the 2024 US elections. The core problem, the researchers argued, was demand rather than supply: polarized audiences want to believe what they see, regardless of whether it was generated by an algorithm or a phone's video editor.

The practical shape of the deepfake problem is more targeted than civilizational. Fraud, non-consensual intimate imagery, and impersonation attacks are where the real damage is concentrated. These are serious harms and they are growing. The specific doomsday scenario of deepfakes making democratic society epistemically ungovernable still looks more like a plausible future risk than a present reality.

The gap between creation and detection is real. So is the tendency to overstate what that gap means.