Deepfake Detectors Are Signals, Not Verdicts

Do not label media real or fake from one detector score. Performance changes across generators, compression, cropping and testing populations.

Why detectors fail

A detector learns patterns from known examples. New generators remove patterns; platforms recompress files; screenshots alter pixels. Laboratory accuracy may not transfer to unfamiliar media.

NIST evaluates face-analysis systems under defined datasets and conditions. C2PA takes another approach: signed provenance can record creation and edits, but missing credentials do not prove falsity.

A stronger workflow

1. Preserve the best original.
2. Find the earliest source.
3. Check date and location.
4. Inspect credentials.
5. Record detector and version.
6. Seek specialists for consequential decisions.

Label uncertainty

Compression can create oddities, while synthetic media can look clean. A detector percentage is usually model output, not a courtroom probability.

Durable verification asks where the file came from and what happened to it. Detection assists; it does not conclude.

Preserve the evidence before testing

Download the original file when lawful and record its hash, source URL and acquisition time. Running it through messaging apps, editors or converters can change the evidence. Test a copy and keep the original untouched.

Compare like with like

A detector evaluated on studio-quality face swaps may not generalize to synthetic audio, fully generated video or heavily compressed clips. Read its model card or evaluation description and check whether the medium and conditions match the file in question.

Example: a viral speech clip

Before analyzing facial pixels, search for the complete event recording from the speaker, broadcaster or venue. Compare the words, camera angle, weather and audience reaction. A clipped authentic video can be misleading, while an apparent lip-sync mismatch may come from dubbing or transmission delay.

Escalate consequences, not confidence

A low-stakes meme can remain unverified. Evidence used for employment, elections, fraud reports or legal claims needs qualified forensic review and a documented chain of custody. Do not let a consumer score create false certainty.

Do not upload sensitive evidence casually

A detector may retain submitted media or use it to improve a service. Read its privacy and retention terms before uploading private recordings, identity documents or abuse evidence. When possible, use an approved forensic environment or a local tool whose handling has been evaluated, and share only the minimum excerpt needed.

If public communication is necessary before verification finishes, describe what is known: the source is unconfirmed, analysis is ongoing and no authenticity determination has been made. Do not publish the detector's score as a headline.