Confabulation

Confabulation is borrowed from neuroscience, where it describes a memory phenomenon in which people produce false memories without any intent to deceive — they genuinely believe what they’re saying. Applied to AI models, confabulation is a more technically precise term than “hallucination” because it captures the important point that the model isn’t lying — it’s doing exactly what it was trained to do (predict fluent text), which sometimes leads it to produce confident-sounding statements about things it has no reliable information about. The distinction matters for how we think about solutions: confabulation isn’t a bug to be patched so much as a fundamental characteristic of how language models work, to be managed through design and calibration.