Confabulation
kon-fab-yoo-LAY-shun
Definition
Confabulation is Geoffrey Hinton's preferred term for AI "hallucinations" - the phenomenon where models generate plausible-sounding but incorrect information. Hinton argues the term better reflects that humans do the same thing.
Why Confabulation, Not Hallucination?
"Hallucination" implies seeing things that aren't there - a pathological condition. "Confabulation" is a psychological term for memory construction, which is normal human behavior.
"Hallucinations should be called confabulations - we do them too." — Geoffrey Hinton
The John Dean Example
Hinton uses Watergate testimony as proof humans confabulate:
John Dean testified under oath about meetings with Nixon. He was trying to tell the truth, but was "wrong about huge numbers of details" - meetings that never happened, misattributed quotes. Yet "the gist of what he said was exactly right."
How Memory Actually Works
Neither humans nor LLMs store information like files:
"We don't store files and retrieve them; we construct memories when we need them, influenced by everything we've learned since."
This construction process can introduce errors - the same mechanism behind both human false memories and AI confabulations.
Implications
- Confabulation is normal - not a bug unique to AI
- Verification matters - for both human and AI outputs
- Confidence calibration - the real problem is not knowing when you're confabulating
Related Terms
- Hallucination - The common (but arguably misleading) term
- Grounding - Connecting outputs to verified sources
Mentioned In
Geoffrey Hinton at 00:20:00
"Hallucinations should be called confabulations - we do them too. We don't store files and retrieve them; we construct memories when we need them."