A video showing a familiar face saying something that was never actually said no longer announces itself as fake the way it did only a couple of years ago. Synthetic video generation has advanced enough that a viewer scrolling a social feed, sitting through a video call, or watching a clip forwarded by a friend can be persuaded on a quick look, not because the fake is flawless under close inspection, but because most viewing happens in seconds rather than minutes. That shift matters because so much of the trust placed in video, as evidence, as an identity check, as a stand-in for meeting someone in person, has rested on the assumption that faking it convincingly was hard and required expertise most people didn’t have.
How a Synthetic Video Is Actually Built
Modern deepfakes are produced by training a machine learning model on real footage and images of a target person, then using that model to generate new frames that swap a face, alter mouth movements, or synthesize an entirely new performance. According to a summary of the underlying technology, deepfakes rely on techniques such as generative adversarial networks and, more recently, diffusion models, which pit a generator against a detector during training until the generator’s output becomes difficult for the detector to flag. Where early versions needed minutes of clean reference video shot from multiple angles, current tools can work from a short clip or a handful of photos, cutting the amount of material a specific person needs to have posted online before a convincing fake of them becomes possible. The computing power required has also dropped sharply, moving what once needed a specialized research lab onto consumer-grade hardware and, increasingly, into simplified apps that automate most of the technical steps for a user who supplies only source images and a script.
The Old Giveaways Are Disappearing One by One
For years, a specific list of flaws separated a fake from the real thing: unnatural blinking patterns, lighting on a face that didn’t match the room behind it, warped ears or teeth, and mouth movements that drifted slightly out of sync with the audio. Newer generation models have closed most of that list. Blinking now looks natural because training data captures full ranges of eye movement, lighting can be matched frame by frame, and audio-driven lip sync has become precise enough that mismatches are rare in the short clips optimized for a feed, a format that also gives a viewer fewer frames in which any remaining flaw might surface. Hands, once one of the hardest details for a generative model to render correctly, have improved as well, though they remain a comparatively weak point in longer or more complex scenes.
Regulators Are Already Building a Case Against AI Impersonation
The scale of the underlying capability has drawn regulatory attention. The Federal Trade Commission has proposed extending its rules against government and business impersonation to cover AI-generated impersonation of any individual, a change aimed at the kind of fraudulent, AI-manipulated video and audio now circulating in scams and fabricated endorsements. The agency has said fraudsters are using AI tools to impersonate real people with enough precision to run extortion schemes against families and to fabricate endorsements from public figures who never appeared in the ad at all. The proposal reflects a recognition that existing fraud rules, written before synthetic media was cheap to produce, did not anticipate a scam built entirely around a fake likeness rather than a fake email address or phone number.
Family Emergency Calls Show How the Fraud Plays Out
The clearest real-world version of this threat so far has involved voice more than video, and it shows how little raw material a scammer actually needs. The commission has warned that a short audio clip pulled from content someone has already posted online is enough to clone a voice convincingly, and criminals have used cloned voices to stage fake calls from a supposedly jailed or injured relative asking for money to be wired immediately. The same short-clip vulnerability applies to video, since a public speech, a media interview, or even a few minutes of footage posted to social media can supply enough material to drive a synthetic performance of the same person. Older adults have been singled out as frequent targets of these calls, in part because a grandchild-in-trouble narrative plays on urgency and family loyalty faster than it allows time for verification.
What Still Gives a Fake Away, and What Doesn’t
Compression artifacts, slightly too-smooth skin, and an uncanny stillness around the eyes can still flag a fake in some cases, but none of those cues are reliable anymore, especially in short, low-resolution clips shared through messaging apps, where compression already degrades quality on real footage too. Relying on a mental checklist of visual tells made sense when those tells were consistent and hard for a generator to fix; it makes far less sense now that most items on that list have already been addressed by at least one widely available tool.
Building a Verification Habit Instead of Trusting a Glance
The more durable defense is procedural rather than visual: verifying a request through a separate, already-known channel rather than trusting the video or call itself, and treating urgency paired with a request for money or sensitive information as the actual warning sign, regardless of how convincing the face and voice attached to it happen to look. A prearranged verification method, such as a phrase or question only a real family member or colleague would know the answer to, adds a check that does not depend on spotting a flaw in the fake at all. That kind of habit holds up whether the manipulation is a blurry clip or a studio-quality synthetic video, because it never requires the viewer to win a technical contest that the underlying tools keep getting better at.
This article was produced with the assistance of AI and reviewed by Morning Overview editors.
More from Morning Overview
