This paper identifies and formalizes context poisoning: the inadvertent introduction of a concept, frame, or term into an AI conversation that biases subsequent outputs by persisting in the context window. Unlike prompt injection (adversarial, deliberate) or hallucination (structural fabrication), context poisoning is accidental, cumulative, and self-reinforcing. Contaminated outputs mimic coherent reasoning but reflect statistical proximity rather than independent evaluation. Neither the system nor the user can detect the contamination from within the interaction because it presents as thematic consistency rather than bias. The paper grounds context poisoning in the AI Dunning-Kruger (AIDK) framework, analyzes its interaction with the Interactive Dunning-Kruger Effect (IDKE), identifies boundary conditions where the mechanism is minimal, and proposes mitigation strategies organized by Human-Curated, AI-Enabled (HCAE) deployment tiers. Appendix A provides synthetic demonstrations of the mechanism under controlled professional scenarios. Context poisoning is not a bug to be patched but a structural consequence of how probability-based systems use conversational context.