Deepfake evidence and the reliability of perceptual judgement in court: A forensic psychological analysis

Synthetic audio, images and video challenge a long-held assumption in the courtroom: that seemingly direct recordings are unusually transparent evidence. The current narrative and scoping review combines forensic psychology, metacognition, misinformation, digital forensics and evidence law with a focus on literature published from 2020 to September 2026. Human observers often perform poorly at classifying deepfakes and can be poorly calibrated to that limitation.

Accuracy and response bias can be affected by context, repetition, prior belief, emotion, source labels, and warnings. A warning could increase vigilance but also create a liar’s-dividend effect where real material is discounted. Machine detectors beat unassisted observers on matched benchmarks, but accuracy degrades under compression, unseen generators, adversarial alteration and distribution shift; a detector score is not self-authenticating.

The paper presents a source-to-verdict model, a psychologically-informed reliability framework, and a six-stage courtroom workflow: preserve, document, triage, examine, challenge, and instruct. Proposed are threshold authentication hearings for materially disputed synthetic media, original-file preservation, multimodal and provenance-based analysis, transparent error reporting, independent replication where stakes are high, and jury directions that distinguish authenticity from evidentiary weight. The bottom line forensically is that perceptual confidence is no substitute for shown provenance and validated analysis.