Synthetic and Manipulated Media
Last updated: October 21, 2026
Manipulated media comes in two kinds, and they call for different checks. Britt Paris and Joan Donovan’s Data & Society report Deepfakes and Cheap Fakes defines a deepfake as video altered through machine learning to “hybridize or generate human bodies and faces”. A cheap fake is made with cheaper, accessible software or none at all: Photoshop, lookalikes, re-contextualized footage, or a clip sped up or slowed down. Their central finding is that deceptive media has never needed advanced technology. For an investigator, that means tracing the source often settles a case before a detector is needed.
| Risk | Indicators |
|---|---|
| Generative images | Hands, text on signs, lighting inconsistencies — weak heuristics; use detector tools cautiously |
| Face swaps / deepfakes | Temporal glitches, unnatural blinking — confirm with specialist tools and source tracing |
| Cheapfakes | Mislabeled real footage from another event — often caught by reverse search |
Order of checks
- Trace the source. Find the earliest upload and who posted it. Keyframe and reverse image search (see Reverse Image Search) catch most re-labelled footage.
- Check provenance data. C2PA’s Content Credentials standard lets publishers and creators attach a record of a file’s origin and edits. C2PA describes it as working “like a nutrition label for digital content”. Credentials that are present and intact are strong evidence of origin. Missing credentials are not, on their own, evidence of manipulation.
- Inspect the file. The InVID-WeVerify plugin reads image and video metadata and applies forensic filters to still images. Treat what it shows as signals to explain, not as verdicts.
- Use a detector last. A detector score is supporting evidence at most.
Disclosure
Detector scores are probabilistic. Pair with provenance (original upload chain) whenever possible. For user-facing outputs, apply ITI inferred / AI-generated transparency rules.
When Citizen Analyst publishes a finding that media is synthetic or manipulated, the piece states which checks were run and what each one showed, so readers can weigh the evidence themselves.
Related
Sources
- Data & Society, Deepfakes and Cheap Fakes (Paris and Donovan) (read 2026-10-21)
- C2PA, Content Credentials (read 2026-10-21)
- InVID project, InVID Verification Plugin (read 2026-10-21)
Expanded with AI research and drafting assistance from the primary sources cited; editorial review pending.
