How can you tell if a video or voice is AI-generated?

Visual and audio artifacts still appear in synthetic media: inconsistent lighting or shadows, unnatural blinking or mouth movement that drifts out of sync, hair and jewelry edges that blur during motion, hands and teeth rendered oddly, and audio with flat prosody or missing breath sounds. These are worth knowing, but they are a shrinking advantage. Generation quality improves quickly and the artifacts of one year are frequently gone the next, so building your defense on spotting flaws is building it on ground that keeps moving.

Verification is the durable approach because it does not depend on the quality of the fake. For a video, trace it to the original source rather than the account that reposted it, check whether reputable outlets are reporting the same event, and use reverse image search on frames to find earlier versions. For a call or voice message claiming to be someone you know, hang up and call them back on the number you already have, which defeats the attack completely regardless of how convincing the voice was. Agreeing on a family code word in advance is a genuinely effective and nearly costless measure against the grandparent scam and its AI-assisted successors.

The threat pattern to recognize is emotional urgency plus an irreversible payment method. Synthetic voice cloning now requires only seconds of audio, which is available from any social media video, and it has been used in both family emergency scams and corporate fraud where an employee was instructed by a familiar-sounding executive to transfer funds. Requests involving wire transfers, gift cards, cryptocurrency, or payment apps should be treated as fraudulent until verified independently, because those methods are chosen specifically because they cannot be reversed. Provenance standards that cryptographically sign authentic media at capture are being deployed by camera makers and platforms, which will help over time but is not yet something you can rely on.