How to tell if a video is AI: verify the full clip, source and soundtrack

Start with the claim the video is making

Write one sentence describing the claim: a person said something, an event happened at a particular place, or a scene was captured on a certain date. A clip can contain genuine footage while its caption supplies a false date or its soundtrack attributes invented words. Asking only ‘Is this AI?’ can miss the actual reason a post is misleading.

AP’s published verification guidance recommends tracing original sources, reverse-image searching and checking corroborating reporting. Use those as evidence-gathering steps, not as a demand that every genuine video already have a search match. An unindexed recording can still be authentic; a widely reposted one can still be misrepresented.

Sources: Associated Press: AP guidance on generative AI and source verification

Follow the full-recording workflow

Keep a simple log with four columns: timestamp, claim, evidence and unresolved question. This avoids turning a discrepancy in one segment into an unsupported verdict on every frame.

1. Save the post URL, account handle, displayed date and complete caption. Note the clip’s duration and any visible cuts. A downloaded file alone may omit the post’s claim and disclosure.
2. Locate the earliest accessible source and request or find the longer recording. Compare what happens before and after the shared segment. Describe it as the earliest source you found, not necessarily the first upload anywhere.
3. Search distinctive frames from different parts of the clip. Match landmarks, signs and the sequence itself. A matching older upload can establish prior use without proving that every later edit kept the original audio.
4. Read creator disclosures and the platform’s label on the original page. A screen recording or repost may omit that interface. Record the wording, location and date you checked it.
5. Compare the soundtrack with the longer source at matching timestamps. Look for a replaced voice-over, missing questions or a different language track. Lip movement or unusual cadence is a lead to investigate, not a reliable verdict by itself.
6. Check the external context: the stated venue, event date and relevant original statement or independent account. If the claim concerns speech, check the actual words and translation separately from the speaker’s appearance.
7. If a tool offers a supported provenance or watermark check for that exact media type, record the file, tool, scope and result. Keep a processing failure separate from a negative finding; never extend one analyzed segment to an unexamined recording.

Keep a label note with the source record

In your verification log, record the exact disclosure you saw and whether it appeared on the source page or only in a repost’s caption. Link that observation to the particular URL and review date. For platform-specific interpretation, use the related YouTube label guide below. Keep this observation alongside your recording comparison rather than allowing it to replace the comparison.

Worked example: a real event with an uncertain voice-over

Suppose a 12-second clip claims to show a speaker making a new announcement. You locate a longer official recording containing the same stage and movements, but the shared version has different audio. You have established a soundtrack mismatch for that segment. You have not established which software produced the replacement or whether the rest of the official recording involved AI.

A useful note says: ‘The visuals match this recording at this timestamp; the repost’s audio differs. I could not verify the source of the replacement audio.’ Link the source and identify the disputed segment. This hypothetical example shows why matching the picture is only one part of verifying a video.

When an image tool helps—and when to stop

A separately extracted still can help find earlier appearances or inspect the saved image file. It cannot authenticate the soundtrack, transitions or unexamined frames. Read the linked frame-limitations guide before treating an image result as a video finding. isGenAI does not provide a general AI video or voice detector. If the source, context or relevant track remains unresolved, preserve that uncertainty instead of publishing an accusation based on a visual hunch.

Questions about this case