Synthetic Media Detection 2026: Text, Image, Audio, Video and Provenance
Synthetic media detection works best as a layered review, not as a single accuracy score. Start with provenance and watermark signals when they exist, then add fingerprinting, metadata, source history, artifact review and human judgment. NIST frames detection, provenance, labeling, watermarking, testing and auditing as complementary approaches, and the same principle applies across text, image, audio, video and music.
Reviewed June 2, 2026. Source basis: NIST synthetic-content transparency, Google SynthID Detector, Google DeepMind SynthID, C2PA Content Credentials FAQ and EU AI Act Article 50.
Fast answer
What is the safest synthetic media detection workflow?
Preserve the original file, check C2PA or other signed provenance, check SynthID only when supported Google-generated media is plausible, run fingerprinting for rights-sensitive audio or video, review metadata and source-account history, then use detector scores and artifact review as supporting evidence. Do not treat one score, one missing watermark or one missing metadata field as proof.
For music-specific Suno or Udio questions, use the Suno/Udio AI music watermark detection workflow: it separates public watermark checks from platform-scale classifier evidence, distributor metadata, license records and source-account provenance.
1. The Multi-Modal Detection Matrix
| Modality | Best Workflow | Signal Strength | Key Signals | Failure Mode |
|---|---|---|---|---|
| Text | Detector score plus document history | Useful for triage | Perplexity-like predictability, burstiness, repetition, source trail, drafts and edit history | Short samples, edited AI text, templated human writing and non-English text can raise false positives. |
| Image | Artifact review plus provenance | Strongest when provenance is present | C2PA Content Credentials, SynthID where supported, EXIF history, reverse image search and visual artifacts | Screenshots, recompression, crops and stripped metadata can remove or hide provenance clues. |
| Audio and music | Provenance plus fingerprinting | Depends on generator and file history | C2PA when present, SynthID for supported Google audio, waveform review, fingerprint matches, distributor metadata and source account records | A missing watermark is not proof of human origin, especially for Suno, Udio, local models or re-uploaded MP3 files. |
| Video | Frame, audio and provenance review | Best as a layered review | C2PA, SynthID where supported, face artifacts, temporal consistency, lip-sync review, audio match and upload context | Low resolution, heavy edits, filters and partial face replacement reduce reliability. |
| Multi-modal case | Cross-check all evidence families | Best for high-stakes review | Do the voice, face, text, source history, watermark signals, metadata and publishing context agree? | No single detector, watermark, metadata field or visual clue is comprehensive on its own. |
2. The 7 Cross-Modal Evidence Families
3. Detection Workflows by Use Case
| Workflow | Modalities | Best Use | Limitation | Next Step |
|---|---|---|---|---|
| Provenance-first review | Image, audio, video and documents | When C2PA or another signed source trail exists. | Only helps when credentials are present, trusted and correctly bound to the asset. | Verify the credential, signer trust and edit chain before using detector scores. |
| Watermark-specific review | Text, image, audio and video where a supported generator embeds a watermark | When the suspected generator is known and the detector supports that generator or modality. | No public detector covers every AI system, and missing watermarks are not proof of human origin. | Use SynthID checks for supported Google media and separate workflows for Suno/Udio music. |
| Fingerprint and rights review | Audio, music and video | When copyright, catalog matches, samples or previous uploads matter. | A new synthetic work can have no match, and a match does not answer every license question. | Compare fingerprint results with license records and platform/distributor metadata. |
| Classifier and artifact review | Text, image, audio and video | Fast triage when no provenance signal exists. | Classifier scores can be brittle and should not decide high-stakes cases alone. | Escalate to human review and preserve the original file. |
| Cross-modal review | Video, voice, transcript, text and account history | High-risk investigations where multiple evidence families are available. | Aligned signals increase confidence but still require source context. | Check whether face, voice, transcript style, upload path and provenance agree. |
4. Transparency and Compliance Context
| Source / Context | 2026 Status | Requirement | Review Note |
|---|---|---|---|
| EU AI Act Article 50 | Transparency duties are central to EU synthetic-media compliance. | AI systems that generate synthetic audio, image, video or text outputs must make those outputs machine-readable and identifiable as AI-generated or manipulated where Article 50 applies. | For deepfake audio/video/image content, disclosure must be clear unless a listed exception applies. |
| NIST synthetic-content transparency | NIST frames detection, provenance, watermarking, labeling, testing and auditing as complementary technical approaches. | Treat each method as context-specific rather than a comprehensive solution. | Use provenance and detection together, with human process and source context. |
| C2PA Content Credentials ecosystem | C2PA focuses on tamper-evident provenance data that can travel with media assets. | Verify credential integrity, trust list path, asset binding and whether the credential survived editing or publishing. | C2PA is transparency and integrity infrastructure, not DRM and not a universal truth detector. |
| Platform and distributor policies | Policies differ across publishers, schools, social platforms, labels and streaming distributors. | Preserve source exports, account records, platform labels, license documents and review logs. | Policy decisions should not depend on one missing watermark or one classifier score. |
5. Source-Backed References
Frequently Asked Questions
What is multi-modal synthetic media detection?
Cross-checking AI-generated content across text, image, audio, video, provenance, watermarks, fingerprints, metadata and publishing context. It is stronger than one score because every signal has different failure modes.
Can one detector prove content is AI-generated?
No. A detector score, missing watermark or missing metadata field is evidence for review, not proof by itself. High-stakes decisions should preserve the original file and combine several evidence families.
What is C2PA and how does it work?
C2PA Content Credentials are tamper-evident, cryptographically signed provenance data. They help verify source and edit history when present, but they do not prove every fact about the creative process and can be absent or stripped.
What is SynthID and which content does it help verify?
SynthID is Google DeepMind watermarking technology for supported AI-generated media. Google says SynthID Detector can scan Google-generated text, image, audio and video, including supported audio such as Lyria. It is not a universal detector for every generator.
How should Suno or Udio AI music be reviewed?
Preserve the original export, source account, project history, distributor metadata and license records. Then combine those records with C2PA if present, SynthID only where supported, audio fingerprinting and human rights review. A missing watermark does not prove the track is human-made.
How does EU AI Act Article 50 affect synthetic media?
Article 50 includes transparency obligations for certain AI systems and deployers, including machine-readable marking for AI-generated or manipulated outputs and disclosure duties for deepfake audio, image or video content where the article applies.
What forensic signals work best?
The most defensible workflow combines C2PA or other provenance when available, supported watermark checks, fingerprinting, metadata, source-account history, artifact review, cross-modal consistency and human judgment. The best signal depends on the generator, file history, content type and decision risk.
Methodology
This guide avoids universal accuracy claims because synthetic media detection depends on generator, file history, compression, watermark support, metadata survival and review context. The workflow is grounded in NIST digital content transparency guidance, C2PA Content Credentials documentation, Google SynthID references, EU AI Act Article 50 and EyeSift's live synthetic-media review tools.
Related Eyesift Guides
- C2PA Content Credentials 2026 — Cryptographic Provenance
- AI Watermark Detection 2026: C2PA, SynthID and Metadata
- AI Watermarking Standards 2026: SynthID, C2PA, IPTC, ISCC
- Suno/Udio AI Music Watermark Detection 2026
- C2PA Deepfake Detection 2026
- AI Image Detection Tools 2026
- AI Voice Detection & Deepfake Audio 2026