AI's "Forgetfulness" is Not a Knowledge Deficit but a Retrieval Failure

When large language models (LLMs) generate incorrect information in a phenomenon known as "hallucination," most developers assume the model simply lacks the necessary knowledge. Consequently, common responses have included enlarging the model, increasing training data, or building mechanisms to incorporate external information (RAG: Retrieval-Augmented Generation). However, a new study published by researchers from Google Research and Technion - Israel Institute of Technology fundamentally challenges this premise.

According to experiments conducted by the research team, cutting-edge models such as GPT-5 and Gemini-3 have been confirmed to retain 95-98% of tested facts as parameters (memory regions within the model). This suggests that the majority of the problem is not "knowledge deficiency" but rather "inability to effectively retrieve stored knowledge." This finding could serve as a basis for reassessing the direction of hallucination mitigation strategies.

The research team proposed "Fact-Level Profiling" as a new framework for evaluating model knowledge. While conventional evaluation looked at whether questions were answered correctly, this method repeatedly probes a single fact from multiple angles and phrasings to thoroughly examine whether the model actually possesses that information internally. The researchers point out that "encoding (memory) failures" and "recall failures" cannot be distinguished by standard accuracy metrics, yet their causes and solutions are entirely different.

Based on this framework, the research team classified the state of knowledge into five profiles. These include "direct recall," where facts are remembered and answered immediately; "encoding failure (empty shelf)," where neither memory nor recall is possible; and "recall failure (lost key)," where information is stored but cannot be retrieved. For example, using the fact that "Oasis held its first live performance at the Boardwalk club," the researchers demonstrated a case where the model can reproduce the information in its original context but cannot answer the question "Where did Oasis hold its first live performance?"

The practical implications of this research are significant. If the primary problem is "retrieval difficulty" rather than knowledge deficiency, improvements may be possible through increased computational resources during inference or optimized questioning strategies, rather than model retraining or massive data augmentation. The researchers themselves note in their paper that "for recall failures, post-hoc learning approaches that enable more effective utilization of knowledge the model already possesses are effective."

For development teams aiming to improve AI system reliability, this shift in perspective has meaningful implications for cost-effectiveness. If dependence on larger models or external databases can be reduced, it could lead to infrastructure cost savings. However, regarding "facts where encoding itself fails," data enrichment and model expansion remain necessary, and accurately distinguishing between the two failure types is a prerequisite. Going forward, how this classification will be integrated into actual systems deserves close attention.