Artificial intelligence is increasingly used to predict toxicological endpoints, prioritize chemicals, integrate high-dimensional biological data, and support alternatives to conventional testing. Yet a toxicity score is not itself a safety judgment. A predictive output usually characterizes the probability, class, or magnitude of a hazard-related endpoint under a particular model and data domain, whereas safety assessment requires an inference about biological significance under specified exposure conditions, mechanistic plausibility, species relevance, population variability, uncertainty, and decision context. This article develops a conceptual account of how those evidentiary dimensions should be combined without converting predictive performance into unwarranted claims of safety or harm. The analysis distinguishes hazard prediction from exposure-conditioned interpretation, examines how mechanistic evidence and cross-species biology can strengthen or weaken transfer, and treats uncertainty as a structured property of the evidence rather than a residual confidence score. The resulting approach positions AI output as one evidentiary input within an iterative safety judgment in which concordant evidence may increase confidence, discordant evidence may redirect interpretation, and unresolved limitations may block downstream inference. The central implication is that defensible AI-supported toxicology depends less on whether a model produces an accurate toxicity score in isolation than on whether the complete evidentiary configuration justifies the decision being made.