The rapid introduction of artificial intelligence into the criminal process and the mass proliferation of "deepfake" technology are destroying traditional paradigms of proof. There is an acute need to verify machine conclusions and protect justice from digital falsifications, which is exacerbated by the procedural lag of legislation. The formation of a comprehensive doctrinal understanding of the mechanisms of verification and the conditions for the admissibility of AI evidence based on a comparative legal analysis of foreign experience and an assessment of the legislative novelties of the Republic of Kazakhstan. The study applied a comprehensive approach, including comparative-legal (analysis of the systems of the USA, UK, Germany, EU), formal-legal, formal-logical, and systemic-structural methods. The epistemological dualism of AI evidence ("acknowledged" algorithmic decisions and "unacknowledged" deepfakes) was revealed. It has been established that foreign experience (EU AI Act, US and ECtHR practice) proves the acute need for a strict procedural audit of algorithms to prevent miscarriages of justice (the "liar's dividend" phenomenon). It is proven that the domestic Criminal Procedure Code of the RK objectively lags behind the norms of the new Law of the RK "On Artificial Intelligence" and requires adaptation. The critical need to modernize the doctrinal concept of "special scientific knowledge" is substantiated. The introduction of mandatory mechanisms for the algorithmic audit of intellectual systems will make it possible to eliminate the monopoly of AI developers on establishing the truth, ensuring an ideal balance between the technological efficiency of investigating crimes and strict compliance with the constitutional rights of citizens.