AI Hallucinations Aren't Ignorance — It's Forgetting
September 30, 2026 · 5 min read
You've seen it: you ask an AI a question it clearly should know, and it answers with total confidence — and total wrongness. The old explanation was "it never learned that part." A new Google study offers a different story: it learned it. It just can't remember it.
The numbers
The researchers tested frontier models, and the results are counterintuitive: the models have already encoded 95–98% of the knowledge into their parameters — the learning went fine. But when you ask for an answer directly, without letting the model "think," 26–34% of that encoded knowledge fails to come out. The bottleneck isn't how much it knows. It's recall.
The good news: the problem has a fix. Give the model more thinking compute and it recovers 40–65% of the knowledge it couldn't recall. In plain English: stop rushing it. Let it think.
Like going blank in an exam
It's exactly like going blank in an exam. The knowledge is in your head; under pressure, you just can't pull it out. Nobody would call that student stupid — we'd say they need a moment to think. Turns out the same applies to a hundred-billion-parameter model.
What you can actually do about it
Next time an AI gives you a shaky answer, don't write it off as dumb. Turn on deep-thinking or reasoning mode, or just add "take your time and show your reasoning." The quality jump is often visible to the naked eye. Problems that scaling data and parameters can't solve, "thinking a little longer" solves about half of.
The limit
It's not a cure-all. For the part it genuinely can't recall, you still need to check sources. AI, it turns out, is just a very smart person who occasionally forgets things.