Chapter ADeepfakes and synthetic mediaPage 3 of 8

Deepfakes and synthetic media

Detection and its limits

Artifacts can tell you where to look, but reliable decisions come from combining content inspection with source, context, and corroboration.

~14 minLimits and tradeoffs

Before you start

Why this matters

A clip appears to show a local official making an inflammatory statement. One viewer points to unusual blinking and declares it fake. Another says the lighting looks natural and declares it real. Both are making a larger claim than their evidence supports.

Compression can create blurry edges. Poor connections can break lip synchronization. A person may blink unusually because of bright lights or stress. Meanwhile, newer generation and editing tools can avoid yesterday’s obvious artifacts. Appearance-based clues are useful prompts for investigation, not universal truth machines.

The goal is not to become certain from a glance. It is to know what each kind of evidence can and cannot establish.

1Learn the idea

Read

Four layers of checking

Use four layers, moving from the item toward its surrounding evidence.

1. Content clues

Inspect the media itself. In video, look for inconsistent reflections, shadows, facial boundaries, jewelry, text, or motion. In audio, notice sudden changes in room tone, cadence, breathing, pronunciation, or background sound. In images, examine repeated textures, impossible geometry, inconsistent writing, and mismatched lighting.

These clues change quickly as tools improve. They can also appear in authentic media after cropping, stabilization, noise reduction, low-light capture, or repeated uploads. A clue should lead to a question such as “Can I find a better-quality original?” It should not end the inquiry.

2. Source clues

Ask who first published the item. A repost with no origin is weaker than a full recording from an accountable institution, journalist, or known participant. Check account history, exact username, publication date, and whether a link truly leads to the claimed domain.

An established account is helpful but not conclusive. Accounts can be compromised, and legitimate publishers make mistakes. Source analysis raises or lowers confidence; it does not replace corroboration.

3. Context clues

Check whether the caption matches the media. Search for earlier versions, longer footage, transcripts, event schedules, and statements from people present. A real recording can be paired with a false date or location. A satirical clip can lose its label when copied. A generated reenactment can be reposted as documentary evidence.

Context checks often solve cases that visual inspection cannot. Finding the original upload with a clear fiction label is stronger than arguing about pixels in a compressed copy.

4. Corroboration and provenance

Look for independent evidence that supports the event: additional camera angles, primary documents, official records, direct confirmation, or reporting from organizations with transparent correction practices. Independence matters. Ten sites repeating one anonymous post are not ten sources.

Some media may carry provenance information recording where it came from and what edits occurred. Standards such as content credentials can strengthen a chain of custody when participating tools preserve and sign the information. Provenance can show useful history, but absence is not proof of fakery. Many authentic cameras and publishing workflows do not yet attach or retain such data.

Read

What automated detectors can do

Automated detectors can estimate whether media resembles examples of generated or manipulated content. They may help platforms prioritize review across enormous volumes or help investigators identify areas that deserve closer inspection.

Their output is not a verdict. Performance can change with a new generator, language, face, recording device, compression level, or editing pipeline. A detector tested on one benchmark may not work equally well on short social clips or noisy phone audio. False positives can wrongly accuse people, while false negatives can create unjustified confidence.

If a detector is used for consequential decisions, ask:

  • What media and populations was it evaluated on?
  • How recent and representative was the test?
  • What are the false-positive and false-negative rates?
  • Does performance change after common platform compression?
  • Is there human review and an appeal route?
  • What other evidence supports the decision?

A percentage such as “87% likely synthetic” is not automatically calibrated probability, and it should not be presented as scientific certainty without explanation.

Read

Reverse searching and metadata

Reverse image search, keyframe search, and ordinary web search can locate earlier appearances. Results may reveal that a supposed current event uses an old photo or that a clip came from a labeled demonstration. These tools work best when you search distinctive frames, quoted phrases, names, places, and dates.

File metadata can contain capture time, device information, or edit history, but platforms often strip it. Metadata can also be altered. Treat it as one part of a chain, not a self-authenticating certificate.

Read

Confidence, not binary certainty

After checking, state what you know with appropriate confidence:

  • “This is confirmed by the full official recording and two independent reports.”
  • “This appears to be a mislabeled older clip.”
  • “The source is unknown, and I cannot verify the claim.”
  • “A detector flagged the audio, but that alone is inconclusive.”

An honest unresolved result is useful. You can decline to share or act without proving exactly how the item was made. The burden of proof should rise with the consequence of the claim.

Checking tutor…

Continue learning · glossary & guides