Can you spot a deepfake? Why visual clues are not proof

Separate three different questions

Ask whether the file was generated or altered, whether it shows the person or event claimed, and whether its caption describes that event accurately. These questions overlap but are not interchangeable. An unaltered recording can carry a false caption; an openly edited illustration can explain something accurately.

For example, imagine a real clip of a public speaker at an old event circulating with a new date. Proving the pixels are camera footage would not verify the new date. Start by writing down what the forward wants you to believe, then identify which part would need evidence.

Use visible oddities as leads

A strange hand, uneven lighting or an awkward mouth movement is a reason to investigate. It is not a reliable yes-or-no test. Small forwarded copies can contain blur, compression artifacts and missing detail. A paused frame may look different from the movement around it.

Watch the surrounding sequence and compare a higher-quality version if one is available from a known source. Describe the observation narrowly: the lip movement looks out of sync in this copy. Avoid jumping from that description to a claim that a named person was impersonated. Conversely, smooth movement and natural sound cannot certify a recording.

A detector score needs its own context

Forensic detection involves several different tasks, including locating edits, identifying generated material and tracing origins. Performance on one task does not establish performance on every kind of forward. NIST emphasises testing with representative media, including material affected by the processing typical of redistribution.

Before treating a result as evidence, ask what input the detector accepted, what it is designed to detect and how its errors were measured. A bare percentage without that context is hard to interpret. If tools disagree, do not average their numbers into a new verdict. Record the disagreement and look for a different kind of evidence.

Do not upload private conversations or sensitive media to an unfamiliar detector just to obtain another score. First read what the service does with submitted files and whether you have a reason to share that material. For many public claims, tracing the public original is a more useful next step.

What Content Credentials can tell you

Content Credentials can carry information about an asset’s origin and editing history. Their validation concerns the provenance information and its relationship to the asset; it does not settle whether the scene or its accompanying claim is true. A photograph of a staged event can still have an accurately documented history.

The absence of credentials is not proof of fabrication. C2PA participation is optional, and a viewer may not have the full history available. Read the actual credential details when present instead of interpreting a badge as a truth rating. Keep the source, context and content questions separate.

Follow an independent verification route

Search for the full recording on the relevant broadcaster, event organiser or speaker’s established channel. Compare the disputed sentence and the moments before and after it. A matching short repost only shows that the same clip travelled elsewhere. It does not identify who first recorded it.

For an event claim, look for a dated programme, an original report, other camera angles or contemporary accounts that do not all quote the same uploader. For a quotation, find the full speech or transcript and compare the wording. A denial from the person involved is relevant, but an independently checkable original provides a different kind of support.

• Keep the received caption and the earliest source you can actually find.
• Distinguish earliest located upload from proven original upload.
• Record which date, place or quoted words you verified.
• Leave unresolved parts unresolved rather than certifying the entire file.

Choose a conclusion that fits the evidence

Sometimes you can establish that a caption is wrong without determining how the video was made. Say that directly. Sometimes the only defensible conclusion is that the original recording could not be found. Do not convert either result into a technical diagnosis of a deepfake.

ForwardCheck can help organise a claim and sources, but it is an AI-assisted beta, not a forensic laboratory. Its own reading of a face or voice does not prove that a real person said or did something. When discussing a suspected fake, explain the evidence you found and the limits that remain; avoid spreading the clip with a more sensational label.

Common questions

Do unnatural eyes or extra fingers prove AI generation?

No. Such details can prompt investigation, but a visual clue by itself does not establish how a file was created.

Does missing Content Credentials mean a video is fake?

No. Credentials are not present on all authentic content, and their absence does not establish fabrication.

Can a genuine video still mislead?

Yes. A false date, location, caption or missing context can misrepresent real footage.