Images
| Examination dimension | Context |
|---|---|
| Evidence sources | Original images, embedded metadata, source assets, and editing projects. |
| Detection approaches | Format-aware structural and statistical analysis; detector results evaluated for the tested image domain. |
| Provenance opportunities | Signed creation or editing assertions and referenced ingredients. |
| Transformation effects | Cropping, recompression, screenshots, and resizing may change signals and remove records. |
| Limitations / reporting cautions | Ordinary editing and compression can mimic apparent generation artifacts; separate observation from source attribution. |
Video
| Examination dimension | Context |
|---|---|
| Evidence sources | Original container, frame sequence, audio tracks, timing, project files, and exports. |
| Detection approaches | Spatial and temporal analysis, track comparison, and evaluated classifiers. |
| Provenance opportunities | Manifest and project relationships across frames and edits where supported. |
| Transformation effects | Transcoding, frame interpolation, clipping, and platform delivery may alter multiple streams. |
| Limitations / reporting cautions | A selected frame cannot establish the entire clip’s history; record sampling and temporal coverage. |
Audio
| Examination dimension | Context |
|---|---|
| Evidence sources | Original waveform or encoded file, channel structure, session projects, source recordings. |
| Detection approaches | Signal and encoding analysis, with format- and environment-specific detection testing. |
| Provenance opportunities | Creation records, signed assertions, and independent recording-side evidence. |
| Transformation effects | Noise reduction, mixing, resampling, and compression can alter relevant properties. |
| Limitations / reporting cautions | Signal anomalies can arise from conventional processing; state the method’s limits. |
Voice
| Examination dimension | Context |
|---|---|
| Evidence sources | Source speech, alleged clone, call or recording context, and authorized creation records. |
| Detection approaches | Validated synthetic-speech analysis and comparisons with clearly defined questions. |
| Provenance opportunities | Generation records, source references, and workflow assertions where available. |
| Transformation effects | Telephone channels and editing can remove or introduce features. |
| Limitations / reporting cautions | Speaker resemblance does not identify a generator or operator; protect the impersonated person. |
Text
| Examination dimension | Context |
|---|---|
| Evidence sources | Document versions, conversation records, exports, editing history, and surrounding context. |
| Detection approaches | Textual comparison and evaluated statistical classification with explicit language and domain limits. |
| Provenance opportunities | Source-referenced interaction history and document workflow records. |
| Transformation effects | Paraphrasing, translation, formatting, and manual editing can change classification signals. |
| Limitations / reporting cautions | Style alone rarely identifies AI authorship or a model; record alternatives to a textual match. |
Multimodal outputs
| Examination dimension | Context |
|---|---|
| Evidence sources | Linked text, image, audio, video, attachments, and tool outputs in one workflow. |
| Detection approaches | Modality-specific examination plus relationship analysis between components. |
| Provenance opportunities | Shared identifiers, ingredient references, and workflow manifests. |
| Transformation effects | Components may be exported or transformed separately, breaking visible relationships. |
| Limitations / reporting cautions | Do not transfer a finding about one modality to all components without supporting evidence. |
Compare method results without collapsing their meaning
A classifier, watermark check, signature validator, and application record can address different propositions. Preserve their distinct results and connect them only through a documented relationship. Use Synthetic Media Forensics for field scope and the detection-versus-provenance case to examine a common reporting error.
Sources & further reading
Source verification checks the official publication and its scope. It is separate from independent technical review of this reference page.
- NIST AI 100-4: Reducing Risks Posed by Synthetic Content
2024 overview of detection, watermarking, and content provenance approaches. Method performance requires context-specific evaluation.
- C2PA Specifications 2.4
Content Credentials technical specification and supporting documents. Claims and identity trust need separate evaluation.
- NIST SP 800-86: Guide to Integrating Forensic Techniques into Incident Response
IT-oriented forensic process and sources; published 2006. It is not a complete investigative procedure or legal advice.
Canonical reference: https://syntheticforensics.org/synthetic-media/
Report a technical correction