How can I tell if an image or video is AI-generated?

The visual tells are disappearing fast, and any list of them has a short shelf life. The durable approach is verification of *provenance* — where did this come from — rather than inspection of pixels. Visual signs still worth checking, with the caveat that top-tier generators have fixed most of them: - **Hands, fingers and teeth** — historically the classic giveaway; now often correct, but still worth a look on lower-quality output. - **Text in the image** — signs, labels, book spines. Garbled or nonsense lettering remains common. - **Physical inconsistency** — shadows falling in different directions, reflections that don't match the scene, jewellery or glasses that merge into skin, patterns that don't continue correctly behind an object. - **Backgrounds and crowds** — faces in the distance often melt; repeated architectural elements go subtly wrong. - **Skin and lighting** — an airbrushed uniformity and slightly too-perfect lighting. - **In video** — flickering details between frames, hair and jewellery that shift, unnatural blinking, and lip sync that drifts. Why this isn't enough: each of these is being fixed generation by generation, and the flaws you learn to spot are the flaws of last year's model. Meanwhile, detection tools have the same reliability problems as AI text detectors — unreliable enough that a confident verdict from one shouldn't be trusted either way. The methods that hold up: 1. **Reverse image search.** If a dramatic image is real, other sources will have it. If it appears only in one viral post, be suspicious. This is the highest-yield check and takes ten seconds. 2. **Trace it to a source.** Who published it first? A named photographer, an agency, a news organisation with something to lose? Or an anonymous account? 3. **Check whether anyone credible reported the event.** A genuinely newsworthy image is accompanied by reporting. Silence around a dramatic image is itself the answer. 4. **Check metadata and provenance signals** where available — C2PA content credentials are being adopted by cameras and some AI tools, though stripping metadata is trivial, so absence proves nothing. 5. **Apply emotional-trigger scepticism.** Content engineered to make you angry or vindicated is the most likely to be synthetic or manipulated, because that's what spreads. The strength of your reaction is a useful warning signal. The honest bottom line: within a short time, visual inspection will not be a viable method for anyone. Shifting your habit now from 'does it look real' to 'who is vouching for this' is the adaptation that will still work in five years.