P22 · CASE 0015 · POPULATED EVIDENCE RECORD

How Can We Distinguish Real Evidence from Synthetic Media?

OPENCOMPLETE_CLAIMS_SOURCES_ADVERSARIAL

No single detector, watermark or metadata field can prove that media is real. A defensible conclusion requires layered verification.

18sources

15claims

10framework dimensions

15adversarial tests

Canonical source basis:
CP1_OBSERVATORY_P22_CASE0015_SOURCE_MAP_AND_EVIDENCE_MATRIX_V0_1_20260731.xlsx

Claims

Claims retain their original research-state labels. A status is not a probability of truth.

ID	State	Claim	Support refs
C01	STRONG SUPPORT	No single detector, watermark or metadata field can prove that media is real.	S01–S12
C02	STRONG SUPPORT	Valid provenance can support origin and edit history without proving event truth.	S01–S03; S15–S16
C03	STRONG SUPPORT	Absence of credentials or watermark does not prove human creation.	S02–S03; S09–S12; S15–S18
C04	PROVISIONAL SUPPORT	Original assets preserve more useful evidence than screenshots and recompressed reposts.	S01–S08; S15–S16
C05	STRONG SUPPORT	Forensic detector performance is model-, dataset- and transformation-dependent.	S04–S08
C06	PROVISIONAL SUPPORT	Watermarks can provide useful generator-specific evidence.	S09–S12; S18
C07	STRONG SUPPORT	Authentic media can still be used with false context.	S04; S13–S17
C08	PROVISIONAL SUPPORT	Independent corroboration is required for consequential event claims.	S04–S08; S13
C09	STRONG SUPPORT	Multiple copies from one upload do not create independent evidence.	S01–S08
C10	PROVISIONAL SUPPORT	Disclosure and consent reduce deception and rights risks but do not verify factual claims.	S13–S18
C11	STRONG SUPPORT	Metadata is useful but editable and must be corroborated.	S01–S08; S15–S16
C12	OPEN	Positive detector agreement may not be independent.	S04–S08
C13	STRONG SUPPORT	Hybrid and partial edits complicate binary real/fake classification.	S01–S18
C14	STRONG SUPPORT	High-consequence classification requires accountable human review.	S04–S08; S13–S14
C15	OPEN	No current universal media-authenticity score is justified.	S01–S18

Sources · 18 serialized

ID	Source	Research status
S01	C2PA Technical Specification 2.4	ACCEPTED TECHNICAL
S02	C2PA Security Considerations 2.4	ACCEPTED TECHNICAL
S03	C2PA Explainer and FAQs	ACCEPTED CONTEXT
S04	NIST — Reducing Risks Posed by Synthetic Content	ACCEPTED EVIDENCE
S05	NIST Open Media Forensics Challenge	ACCEPTED EVIDENCE
S06	NIST — Evaluating Analytic Systems Against AI-Generated Deepfakes	ACCEPTED EVIDENCE
S07	NIST GenAI Pilot Evaluation Plan for Image Generators	ACCEPTED METHOD
S08	NIST AI 600-1 — Generative AI Profile	ACCEPTED METHOD
S09	Google DeepMind — SynthID	ACCEPTED DISCLOSURE
S10	Google DeepMind — SynthID Detector	ACCEPTED IMPLEMENTATION
S11	OpenAI — Verify OpenAI-generated images	ACCEPTED IMPLEMENTATION
S12	OpenAI — Advancing content provenance	ACCEPTED DISCLOSURE
S13	Partnership on AI — Responsible Practices for Synthetic Media	ACCEPTED GOVERNANCE
S14	European Union AI Act — Regulation (EU) 2024/1689	ACCEPTED LEGAL CONTEXT
S15	Content Authenticity Initiative — How it works	ACCEPTED CONTEXT
S16	Adobe Content Authenticity — Inspect	ACCEPTED IMPLEMENTATION
S17	Meta — Labeling AI-generated content	ACCEPTED PLATFORM CONTEXT
S18	Google DeepMind — AI image verification in Gemini	ACCEPTED IMPLEMENTATION

Tests / adversarial controls · 15

ID	Failure mode / test	State	Required control
A01	Absence-of-label fallacy	ACTIVE	Report negative result as inconclusive.
A02	Provenance-is-truth fallacy	ACTIVE	Verify the represented event independently.
A03	Detector absolutism	ACTIVE	Record tool, version, threshold and uncertainty.
A04	Screenshot laundering	ACTIVE	Seek earliest file and source path.
A05	Metadata absolutism	ACTIVE	Corroborate metadata externally.
A06	Compression confusion	ACTIVE	Use matched forensic controls.
A07	False context	ACTIVE	Use temporal and geospatial verification.
A08	Partial-edit concealment	ACTIVE	Use segment-level analysis and ingredient history.
A09	Cross-detector pseudo-consensus	ACTIVE	Document detector dependency.
A10	Source spoofing	ACTIVE	Verify publication path and signatures.
A11	Credential stripping	ACTIVE	Use durable lookup or soft binding where available.
A12	Watermark transfer	ACTIVE	Localize the signal and inspect ingredient structure.
A13	Disclosure ambiguity	ACTIVE	Define materiality and edit scope.
A14	Human reviewer laundering	ACTIVE	Specify reviewer competence and checks.
A15	Model drift	ACTIVE	Version-lock and reassess regularly.

Serialization boundary

This interface does not complete missing research by inference. The current record is COMPLETE_CLAIMS_SOURCES_ADVERSARIAL. Missing source-level or bilingual object-level fields remain absent until they are read from a canonical Observatory artifact.