In an August 12 explainer, Google Research described why an LLM can fail to answer a factual question yet recognize the answer among alternatives. Its WikiProfile benchmark tests 2,150 facts, each paired with ten questions, to distinguish recognition from unaided recall across 13 models.
In the tested Gemini-3-Pro and GPT-5 models, the researchers estimated that 95% to 98% of the examined facts were encoded, yet direct recall failed on 26% to 34% of them. Adding more reasoning reduced the misses but did not eliminate them. These are results for those model versions and the study's question conditions, not measured error rates for newer models or everyday conversations.
The gap was particularly noticeable for rare facts and when a question differed from the context in which the model had encountered the information. The researchers also found that additional thinking recovered many facts classified as encoded, while helping much less when their method found no evidence the fact was encoded.
A model that recognizes an answer after being prompted may still be unreliable when asked to supply that fact from memory. The study infers encoding from behavior; it does not expose literal stored facts inside model parameters or prove that a particular prompt reliably fixes recall.