Why LLMs invent answers instead of saying they don't know
The article distinguishes between "hallucination" (perceptual error) and "confabulation" (filling memory gaps without intent to deceive) to describe why large language models invent answers. It argues that LLMs are not lying but generating plausible-sounding text from incomplete training data, and that understanding this distinction is key to improving AI reliability.
Background
- This article explains a key difference between two terms often used interchangeably in AI discussions. "Hallucination" (an LLM generating false information as if it were fact) and "confabulation" (a human-like filling-in of memory gaps, not usually intentional deception). The author argues that calling LLM errors "hallucinations" anthropomorphizes them and obscures the real problem: models are designed to always produce an answer, even when they lack reliable data, because they have no internal mechanism for uncertainty or a "I don't know" response.
- Why this matters: Overhyped claims about AI "lying" or "hallucinating like a human" can lead to misplaced trust or fear. Understanding that these errors stem from the statistical architecture of the model (predicting the most likely next word, not reasoning from facts) is crucial for anyone using or building on LLMs. The distinction also points to ongoing research into calibration—making models accurately express their own uncertainty—as a path toward more reliable systems.