The model’s knowledge is frozen at training time. After the cutoff, anything it says about current events is interpolated from stale data — or fabricated entirely.

Training data skews toward common knowledge. Niche domains — specific case law, obscure regulations, specialized medical protocols — are sparsely represented, making hallucination more likely exactly where accuracy matters most.

The base model has no live connection to a database, no way to verify claims, and no mechanism to distinguish remembered training data from interpolated output.

Reinforcement Learning from Human Feedback trains models on human ratings. Fluent, helpful-sounding answers often score higher than hedged, uncertain ones — even if the hedged answer is more honest.

The model has no reliable internal signal for “I don’t know this.” Its expressed confidence often does not track its actual accuracy.