The rapid integration of artificial intelligence (AI) into high-risk domains such as biometric identification, security, and law enforcement has raised critical concerns regarding transparency, reliability, and accountability. Modern AI systems, particularly deep learning models, often operate as opaque black boxes, making it difficult to reconstruct their decision-making processes or assess their integrity in forensic contexts. This thesis investigates whether such systems can be systematically analyzed and treated as digital forensic artefacts. The work proposes and validates an AI forensic framework designed to support traceability, integrity verification, and post-hoc auditability of AI models. The framework integrates explainability techniques, representation-level analysis, adversarial robustness testing, and cryptographic integrity verification within a structured logging and chain-of-custody architecture.