Deepfakes Are Getting Good — Here's How to Spot Them (For Now)

The honest headline: visual artefacts are a shrinking advantage. Counting fingers stopped being useful a while ago. What still works is mostly behavioural, contextual and procedural.

What is actually easy now

Cloning a voice from a short sample. Producing a talking-head video of a real person reading a script. Generating a convincing document, screenshot, invoice or product photo. Faking a live video call well enough to survive a distracted minute. None of this needs a lab — it needs a consumer subscription and patience.

Tells that still hold up

Physics under stress. Hair against a moving background, glasses reflections, teeth during fast speech, jewellery, and anything that occludes and re-appears. Ask a suspected live faker to turn their head fully sideways or pass a hand in front of their face.

Timing and prosody. Cloned voices are excellent at timbre and weaker at breath, hesitation and interruption. Interrupting a synthetic caller mid-sentence frequently breaks the illusion.

Edges and light. Mismatched skin tone at the jawline, hard edges around the hairline, lighting on the face that disagrees with the room.

Text inside images. Signage, badges and small print still degrade under scrutiny, though this is closing fast.

Compression laundering. Heavy re-encoding and cropped-in framing are often deliberate, to bury the artefacts. Very low resolution on a supposedly newsworthy clip is a signal in itself.

Tells that stopped working

Extra fingers. Dead eyes and no blinking. Waxy skin. Uncanny mouth shapes. All were reliable in 2023 and are now unreliable — and treating them as proof of authenticity is worse than not checking at all, because it manufactures false confidence.

The verification habits that outlive the tells

Detection tools exist, and independent testing repeatedly finds them brittle on real-world footage — different from the clean benchmarks they were tuned on. So lean on process rather than forensics.

Trace the earliest copy: search a still frame, find where the clip first appeared, and see whether any organisation with a correction policy has it. Check whether the person or body supposedly involved has said anything on a channel they control. Watch for the pattern that fakes rely on — arriving via a screenshot of a screenshot, timed to a news event, and demanding an emotional reaction quickly.

For anything involving money or credentials, use out-of-band confirmation. Hang up and call a number you already had. Voice and face are no longer authentication factors, and the sooner families and finance teams internalise that, the fewer bad afternoons everyone has.

Provenance standards that cryptographically sign camera output at capture are a promising direction, and their coverage today is thin. Absence of a signature proves nothing yet.

The debate worth knowing

There is real disagreement about the scale of harm. One camp points to election-adjacent incidents, fraud losses and non-consensual imagery as evidence of an urgent crisis requiring labelling mandates and platform liability. Another argues the dominant effect so far is not mass deception but the "liar's dividend" — genuine footage dismissed as fake — and that heavy labelling rules risk entrenching the largest platforms while doing little about small-scale abuse. The evidence base on population-level persuasion is genuinely thin, which is why confident claims in either direction should be read carefully.

If you generate synthetic media yourself, the practical ethic is unfashionably simple: disclose it, do not put words in a real person's mouth, and get consent for likeness and voice. Related: your data, their model and the copyright debate.