Synthetic media has crossed the threshold where casual viewers can no longer reliably tell real from generated. The AI Deepfake Detection Guide is an interactive, offline checklist of the artifacts and tells that current image, audio, and video generators still tend to leave behind — and how to weigh them.
How it works
Pick the media type and the guide presents a prioritised list of detection checks. As you inspect the media yourself, you mark each tell you observe. The guide weights the signals — some are far stronger than others — and gives an overall likelihood that the media is AI-generated, along with a recommendation.
Crucially, the tool does not analyse files for you. It teaches you to look, because the most durable defence against deepfakes is a trained human eye combined with source verification, not a single classifier that the next model generation will defeat.
What the checks cover
Images: malformed hands and fingers, distorted teeth and jewellery, garbled background text, asymmetric earrings or eyes, lighting that does not match reflections, and overly smooth or plastic skin texture. Audio: absent or unnatural breathing, flat emotional prosody, a suspiciously consistent noise floor, and clipped word boundaries. Video: irregular or absent blinking, lip-sync drift on plosive consonants, flickering at hairlines and edges, and inconsistent lighting across frames.
How to weigh the signals
Not all tells are equal. A few strong signals carry more evidential weight than many weak ones:
Strong signals (each warrants serious suspicion)
- Fingers that merge, split, or have the wrong count
- Background text that spells nothing or changes between frames
- Lip sync that breaks on hard consonants (p, b, m sounds)
- Breathing that stops entirely during long spoken passages
Moderate signals (flag for further investigation)
- Overly smooth skin texture with no pores or blemishes
- Ear geometry that is asymmetric or simple
- Background objects that are geometrically improbable
- Voice prosody that is flat across an emotionally charged passage
Weak signals (common in legitimate media too)
- Missing or stripped metadata (common on social platforms)
- Compression artifacts (common in reposts)
- Consistent background noise (common in studio recordings)
The provenance check is more reliable than the artifact check
Every detection method based on visual or acoustic artifacts will eventually be defeated by better generators. The durable approach is source verification:
- Where did this media first appear, and when?
- Who published it, and do other reliable sources corroborate it?
- Does it carry a C2PA content credential? (C2PA is an open standard for embedding verifiable provenance data at creation time.)
- Does a reverse image or reverse audio search surface earlier, different-context versions?
A piece of media with a clear, corroborated provenance chain is more trustworthy than one that passes every visual check — because visual checks will not keep pace with generative models.
Tips and notes
No single tell is conclusive — generators fix individual flaws release by release, so the absence of artifacts never proves authenticity. The strongest move is always to verify the source and provenance: who published it, when, and whether a C2PA content credential is present. Treat a high likelihood score as a prompt to investigate the source rather than a final verdict. Everything here runs in your browser and nothing about the media you inspect is recorded or uploaded.