How to Spot AI-Generated Photos, Videos and Voices: A Practical Guide

A man holding a magnifying glass up to his eye, examining something closely

Photo: Hippopx (public domain). Spotting a fake is less about superhuman eyesight and more about checking the right details.

AI-generated media has crossed a line: most people, most of the time, cannot tell a good fake from a real photograph by looking at it. That is not a personal failure — the generators are trained to defeat exactly that judgement. What still works is a checklist: specific details generators get wrong, tools that check a file’s origin, and one sourcing habit that catches almost everything. FreshSynapse covers this because the stakes are live — regulators are still chasing networks of apps built to generate fake imagery of real people, as our news coverage of the deepfake app networks showed.

First, change the question

Stop asking “does this look real?” and start asking three better questions:

Where did this come from? Who posted it first, and is that account the original source?
Does anyone independent confirm it? Is a newsroom, the person supposedly pictured, or another angle of the same moment reporting it?
What is it asking me to do or feel? Fakes are usually built to trigger a fast reaction — outrage, fear, a donation, a share. Speed is the tell.

A real photo with no provenance is a maybe. A shocking photo with no provenance is a no, until proven otherwise.

Photos: the details that still give generators away

These are clues, not proof — real photos can have odd artefacts and good fakes can avoid all of them:

Text. Signs, labels, book spines, logos and background writing are frequently garbled, misspelled or dissolve into plausible-looking nonsense. Zoom in on any text in the image.
Hands and jewellery. Extra or fused fingers still occur, and earrings, glasses arms and watch straps often merge into skin or hair. Check every hand, especially in crowds.
Backgrounds. Generators focus on the subject. Look at the edges: warped architecture, repeating faces in a crowd, furniture that ends mid-object, people who fade into each other.
Light and reflections. Shadows falling in inconsistent directions, mirrors and windows showing a scene that does not match, or eyes reflecting different rooms.
Too-perfect skin and hair. A waxy, airbrushed texture with flyaway hair that turns into a painted smear at the edges is a classic generator finish.
The same face everywhere. “Influencer” accounts built on generated portraits often reuse one face with subtly shifting features across posts — scroll the account, not just the image.

Videos: motion exposes what stills hide

Edges and contact points. Hairlines, ears, the boundary where a swapped face meets the neck; hands passing in front of the face; objects that should cast shadows and do not.
Physics. Liquid, smoke, hair and clothing move in ways that are almost right. Watch anything that pours, splashes or waves.
Blinking and teeth used to be reliable tells; on current generators they mostly are not. Do not clear a video just because the eyes look natural.
Audio sync. Lip movements that are close but not exact on hard consonants (p, b, m), and voices with flat emotional range on emotional words.
Context, again. A politician “caught on camera” by an anonymous account, with no newsroom carrying the clip, should be treated as unproven no matter how clean it looks.

Voices: use procedure, not ears

Voice cloning needs only seconds of someone’s audio, and a cloned voice on a phone call is effectively impossible to identify by ear. So make it a procedure:

Hang up and call back on the number you already have saved — not a number the caller gives you.
Agree a family code word for genuine emergencies, and rotate it like a password.
Ask something only the real person would know — then still verify by a second channel for anything involving money.
Treat urgency as evidence against. “Don’t tell anyone, send it now” is the script of both a scam and a clone.

Tools that check for you

Reverse image search (Google Lens, TinEye): if the “photo of today’s event” was circulating in 2023, you will find the earlier copies in a minute.
Content Credentials / C2PA: a growing number of cameras and AI tools attach a signed record of where a file came from and whether AI made or edited it. Where present, it is strong evidence. Where absent, it proves nothing — most of the world’s real photos carry no credentials either.
Platform labels. Major platforms increasingly label detected AI media. Labels help when present; their absence is not a clearance.
Detection websites that promise a verdict from an upload are unreliable in both directions. Use them, at most, as one weak clue among many.

Frequently asked questions

Can software reliably detect AI images?

No. Detectors and generators are in an arms race, and detectors regularly fail on compressed, resized or screenshotted images — which is how most fakes are shared. Provenance (where the file came from) is a much stronger test than detection (what the pixels look like).

Are watermarks the answer?

Partly. Invisible watermarks and signed credentials genuinely help when platforms preserve and check them, but they can be stripped by re-encoding, and open-source generators may not add them at all. They raise the floor; they do not settle individual cases.

What should I do if a fake of me or my family appears?

Screenshot and record the URLs before reporting — content gets deleted and evidence matters. Report it on the platform (non-consensual imagery is against the rules everywhere major), search for copies with reverse image search, and if the fake is being used for fraud or extortion, report it to the police rather than engaging with the poster.

Is it safe to assume a video call is genuine?

Mostly, for now — real-time video fakes are still expensive and glitchy. But “mostly” is doing work in that sentence: for a video call that ends in a request for money or secrets, apply the same rule as voice. Verify on a second channel.

The bottom line

You cannot out-stare a modern generator, and you do not need to. Zoom into text and hands, reverse-search anything suspicious, treat anonymous provenance as a red flag, and make callback verification a reflex for voices asking for anything. Fakes win on speed and emotion; every one of those steps adds a pause, and the pause is the defence.

Sources

C2PA — Coalition for Content Provenance and Authenticity (the Content Credentials standard): https://c2pa.org/
Content Credentials — how to check a file’s credentials: https://contentcredentials.org/
FreshSynapse — “How a Network of Deepfake Apps…” (news coverage, 3 November 2026): https://freshsynapse.com/articles/wired-deepfake-apps-network-2026-11-03/