GEO as a Discipline Hinges on Understanding How AI Systems Hallucinate

A developer essay on DEV Community argues that AI hallucinations are not solely a product of RLHF post-training but originate as early as pretraining, when models learn false claims or distort facts through statistical compression. Research such as TruthfulQA showed that base models like GPT-3 and GPT-2 confidently reproduced misconceptions even before preference tuning, with larger models proving less truthful than smaller ones. The author outlines five categories of hallucination, ranging from wrong encoded information and flawed weight retrieval to inference errors, confabulation, and post-training behavioral shifts. In real-world AI deployments, the problem widens further, as untrue outputs can stem from faulty retrieval, incorrect tool use, flawed planning assumptions, or one agent treating another agent's invented output as verified fact. The piece concludes that accepting Generative Engine Optimization as a legitimate discipline requires a clear-eyed understanding of these layered failure modes across the entire AI system stack.