Google and Technion Study Finds AI Models Often Have the Right Answer, Just Can't Find It

AI companies have spent years telling the public that hallucinations happen because models don't know enough. Bigger models. More training data. Fancier retrieval systems bolted on the side. That's been the fix.

A new study from researchers at Google Research and Technion says that explanation is often wrong.

The paper, titled "Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality," was written by Nitay Calderon, Gal Yona, Eyal Ben-David, Zorik Gekhman, and Eran Ofek. According to Crypto Briefing, the team analyzed more than 4 million responses across 13 large language models to figure out why frontier AI systems still get facts wrong even when they've clearly been trained on the correct information.

The answer: the facts are usually already in there. The models just can't pull them out.

The Numbers

Using a Wikipedia-derived benchmark called WikiProfile, the researchers found that Gemini-3-Pro and GPT-5 encode 95 to 98% of the facts tested, according to Search Engine Journal, which reported directly from the paper. That's near total saturation. The knowledge is parametrically stored.

But when those same models get asked direct questions about facts they've clearly encoded, they fail to answer correctly 26 to 34% of the time, per the same paper. Even when the models are allowed to use chain-of-thought reasoning, they still miss 11 to 12% of facts they demonstrably have stored.

Search Engine Journal quoted the paper directly: "Accordingly, recall failures account for more than 70% of GPT-5.2's errors and a larger share in stronger models, suggesting recall is indeed a bottleneck."

The stronger the model, the bigger the share of its errors that come from failing to retrieve knowledge it already has, not from lacking that knowledge in the first place.

Why Order Matters

One specific mechanism the researchers identified: subject and object order.

Using the paper's own example, "Oasis played their first gig at the Boardwalk club," Oasis is the subject and the Boardwalk club is the object, based on the order they appeared in the source text the model trained on. A "direct question" asks for the object (where did Oasis play?). A "reverse question" asks for the subject (what band played at the Boardwalk?).

Search Engine Journal reported that models struggle more with reverse questions, even though the underlying fact is identical either way. The model learned the fact in one direction and has trouble accessing it from the other.

VentureBeat reported the researchers also found that recall failures disproportionately hit "long-tail" facts, the obscure stuff that shows up rarely in training data. Common facts encountered from many angles get retrieved fine. Rare facts encountered once, in one phrasing, get stuck.

Thinking Longer Helps, But Doesn't Fix It

VentureBeat's headline points to a real number: models can recover up to 65% of facts they initially failed to recall just by "thinking longer," meaning extra inference-time computation before answering.

Crypto Briefing noted a complementary paper from June 2026, "Thinking to Recall," found that reasoning strategies can pull up buried facts through intermediate computational steps. That same process can generate new hallucinations along the way, since the model sometimes invents false intermediate facts en route to the correct answer.

What This Means for AI Spending

The practical implication, laid out by both VentureBeat and Crypto Briefing, cuts against how the AI industry has been solving hallucinations. If a model already encodes the fact, throwing more parameters or more training data at the problem doesn't help. That's a pre-training fix aimed at a problem the model doesn't have.

The researchers write that recall failures call for post-training interventions instead, meaning techniques that improve how a model accesses what it already knows, not techniques that give it more to know. That's a cheaper, more targeted fix than the industry's default playbook of scaling up compute and data center spending.

Google Research published this work but did not announce a product change tied to it. Neither OpenAI nor any other major lab has issued a public response to the paper as of this writing.