The development of deepfakes has transformed the evidentiary problem of video and audio from file authentication into an assessment of the computational method used to determine media authenticity. This study has two objectives: to examine the legal position and limits of the probative value of deepfake detector outputs under Indonesian criminal procedure and to reconstruct an AI Forensic Reliability Test that safeguards authenticity, scientific reliability, and procedural fairness. It applies normative legal research through statutory, conceptual, comparative, and interdisciplinary approaches. The first finding indicates that detector outputs may enter proceedings through electronic evidence, examination reports, and expert testimony, but they do not constitute independent evidence or binary statements of truth. Their value depends on object authenticity, lawful acquisition, methodological validity, error rates, and corroboration. The second finding develops an AI Forensic Reliability Test comprising eight dimensions: object authenticity, model validity, data quality, error rates and calibration, robustness and generalization, traceability and reproducibility, expert competence and independent testing, and procedural fairness and accountability. The framework operationalizes the negative statutory theory of proof, epistemic reliability, machine testimony, due process of law, equality of arms, and accountable algorithms