A joint study by Google Research and the Technion Institute, published Monday, tested how much of what frontier language models learn during training they can actually access during inference. The results are both reassuring and practically inconvenient. GPT-5 and Gemini-3 encode 95 to 98 per cent of tested facts parametrically — the knowledge is in the model, established during training. But standard inference retrieves only 65 to 74 per cent of that encoded knowledge. The gap — 26 to 34 per cent of encoded facts that cannot be directly recalled — maps closely to observed hallucination rates in structured knowledge tasks.

The mechanism the study proposes is the tip-of-the-tongue phenomenon transposed to artificial systems: the model has the information but fails to surface it through the inference pathway activated by the query. Extended reasoning — chain-of-thought and inference-time compute — recovers 40 to 65 per cent of initially inaccessible facts. The implication is that many hallucinations are not failures of knowledge but failures of retrieval, and that reasoning steps create alternative inference pathways that can access what direct retrieval misses.

This finding is inconvenient for the dominant enterprise AI architecture of 2024-2025: Retrieval Augmented Generation. The standard RAG deployment rationale has been that models hallucinate because they lack the relevant knowledge, so providing that knowledge via external retrieval prevents hallucination. If most hallucinations arise from retrieval failure of knowledge the model already has, then augmenting retrieval with more external documents may not address the root cause. The study's practical recommendations reflect this: avoid reflexively using RAG for all hallucination problems; apply inference-time reasoning selectively to the 10 to 20 per cent of facts that standard retrieval misses; implement query reformulation and verification loops rather than defaulting to external knowledge injection.

A secondary finding compounds the practical challenge: scaling models worsens recall failure rates because larger models encode vastly more facts, creating a larger pool of inaccessible knowledge from which retrieval failures can occur. The models that hallucinate least are not necessarily the ones that know the most; they are the ones whose inference architecture most reliably surfaces what they know. That observation has direct implications for model selection in enterprise deployments where factual reliability is the primary criterion. The biggest model is not automatically the most reliable retriever of its own knowledge.

For enterprise architects building RAG systems, the study does not argue that RAG is wrong — it argues that RAG solves a different problem than hallucination reduction. RAG's genuine value is providing context the model was never trained on, for current events, proprietary documents, or rapidly changing information. For the broader hallucination problem, the study suggests the correct interventions are reasoning elicitation, query reformulation, and verification loops. These are architectural choices that most RAG deployments have not made because the retrieval framing made them seem unnecessary. They are not.