Confabulazione

Fenomeno comune negli LLM che cercano di addurre o creare delle spiegazioni inventate per giuistificare delle affermazioni non basate su fatti. La parola deriva dal linguaggio psichiatrico: "Sintomo frequente in alcune malattie psichiatriche, dovuto alla falsificazione dei ricordi: il malato colma lacune di memoria con invenzioni fantastiche e mutevoli, oppure trasforma in modo non intenzionale i contenuti della memoria stessa."

Si veda per esempio nel paper Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models

"Confabulations to back up wrong solutions. We observe that many models that show reasoning breakdown and produce wrong answers generate at the same time persuasive explanations that contain reasoning-like or otherwise plausible sounding statements to back up the often non-sensical solutions they deliver. We call here such phenomena confabulations. Such confabulations may contain for instance calculations or logic-like statements that make no sense. Confabulations can also refer to reasoning about social norms or structures. For instance, in Command R+ we observe many confabulations that use concepts of gender identity such as non-binary gender or concepts related to inclusion or to cultural context dependent family identification as additional backup for the provided wrong reasoning and incorrect answers. Another type of confabulation that we observe is complete refusal to answer due to invented ethical concerns about the nature of the posed AIW problem, such as violation of privacy or lack of inclusion (for instance in CodeLLama-70B-instruct), or by expressing incorrect concerns about supposedly ill-posed problem formulation."