Google Research: Why LLMs Forget Facts They Actually Know

A new paper from Google researchers found that today's large language models (LLMs — the type of model underlying AI systems like ChatGPT or Gemini) "memorize" 95–98% of facts during training, yet fail to directly recall 26–34% of those same facts when generating a response.

Researchers attribute this to the gap between "encoding" (storing information in a model's internal parameters) and "recall" (reproducing that stored information at query time). In advanced models like Gemini-3-Pro and GPT-5, encoding is nearly complete — 95–98% — but recall is where the bottleneck emerges: more than 70% of GPT-5.2's errors stem from failed recall, and that share grows even higher in more powerful models.

One of the most striking findings is that the order of "subject" and "object" in a fact — for instance, in the sentence "Oasis played their first concert at the Boardwalk club," "Oasis" is the subject and "Boardwalk club" is the object — affects the model's ability to recall it. When a question reverses that order (e.g., "Which band first performed at the Boardwalk club?"), the model has a much harder time recalling the fact — even though it can easily recognize the very same fact when presented as a multiple-choice question.

The researchers found that rephrasing a question doesn't significantly affect recall — the decisive factor is specifically the reversal of subject-object order. They also observed that recall difficulties were more pronounced for rare facts, even ones the model had been trained on. A "thinking" mode helped recover 40–65% of facts that previously couldn't be recalled directly, but at a much higher computational cost.

The researchers note that scaling up a model doesn't fix this recall problem. From an SEO (search engine optimization) standpoint, a practical inference from this research is that presenting facts on web pages in the order people naturally ask about them — subject-then-object — may improve the odds of a page being correctly "recalled" in AI-generated answers, though the study itself doesn't directly prove this.