Spotting Synthetic Media

How to judge whether something was AI-generated, and — more usefully — what to do when you cannot tell.

Start with the uncomfortable part

Detection is losing. Every visual tell on the old checklists — six fingers, garbled text, waxy skin, mismatched earrings — has been fixed. Automated detectors are unreliable in both directions: they miss real fakes and they flag genuine human work, which has already cost students and writers their reputations.

So the goal is not to become a forensics expert. It is to build habits that make you hard to fool even when the media itself is undetectable.

The reliable method: verify the source, not the pixels

Almost every consequential deception fails this test, regardless of how good the generation is.

1. Where did this come from? An account you have never seen, a forwarded message, an unlabelled repost — the provenance is already weak.
2. Who else is reporting it? A real event of any significance has more than one source. A single-source bombshell is the shape of a fabrication.
3. Does the original exist? Search for the claim, not the file. If a video shows a public figure saying something remarkable, the remarkable thing should be findable elsewhere.
4. Does it want something from you? Money, credentials, urgency, secrecy. The deception is usually the request, not the media.

The single most useful question: what is this asking me to do? A convincing fake that asks nothing costs you nothing. An unconvincing one that asks you to transfer money is the dangerous one.

Weak signals, in rough order of usefulness

These are worth noticing but never conclusive.

Signal	Where it still shows
Physics over time	Video. Hair, cloth, water and reflections drifting or resetting between frames
Hands doing complex things	Video and images. Not "six fingers" any more — watch for grip and object interaction
Audio-visual sync	Talking-head video. Consonants at the edge of a sentence are hardest
Background people	Crowds. Extras often warp, merge, or walk oddly
Lighting consistency	Composites. Shadow direction that disagrees with the light source
Breath and room tone	Audio. Synthetic speech is often too clean, with no breathing between clauses
Text in images	Weakest signal now — mostly fixed, but still fails on dense small text

For text, the tells are stylistic rather than mechanical: uniform sentence length, a lack of concrete specifics, no lived detail, the same three transitional phrases, and a conclusion that restates the opening.

Automated detectors

Use them as one input, never as a verdict.

• They produce false positives on non-native English writers, on formal or technical prose, and on heavily edited human work.
• They produce false negatives on any generated text that has been lightly rewritten.
• Vendor accuracy claims are measured on their own test sets. Real-world accuracy is materially worse.

Never take an action that harms someone based on a detector score alone. If you are an educator or an employer, the detector starts a conversation; it does not end one.

Provenance standards

The durable fix is cryptographic provenance rather than detection. C2PA / Content Credentials attach signed metadata describing how an asset was created and edited. Major camera makers, Adobe and several AI vendors now support it.

Worth knowing because it inverts the question: instead of proving something is fake, you check whether it can prove it is real. Metadata is strippable, so absence proves nothing — but presence is strong evidence.

Protecting yourself and people close to you

Voice cloning needs seconds of audio, and a phone call from a familiar voice in distress is the most effective scam currently running.

• Agree a safe word with family and close colleagues. Something no public post of yours contains. Use it to verify any urgent request involving money.
• Verify out of band. Hang up and call back on the number you already have. Never the number that just called you.
• Slow down. Urgency is the mechanism. Almost no genuine emergency is damaged by a two-minute delay to verify.
• Assume your voice and face are clonable. If you have posted video publicly, they are. Plan accordingly rather than hoping otherwise.
• Tell the least technical person you know about the safe word. They are the target.

If you publish AI-assisted work

• Disclose when the audience would care. A stock illustration needs no label; a testimonial, a case study, or a person's likeness does.
• Never synthesise a real person's voice or face without written permission.
• Keep your sources. If someone challenges a piece, being able to show the process settles it.

Your checklist

• Safe word agreed with family and key colleagues
• Out-of-band verification habit for anything involving money
• I check provenance before I check pixels
• I do not act on detector scores alone
• My own disclosure rule, written down: