Why can't a medical AI learn medicine once and answer from memory, and why does it need live access to current sources?

This explores why medical AI can't learn medicine once and answer from memory forever, and why it seems to need ongoing access to outside sources. One caveat first: the collection doesn't directly study how fast medical facts go out of date. What it does explain is why medicine leans so heavily on facts in the first place, and why stuffing more facts into a model's weights is a poor way to keep up. When researchers split model performance into 'knowing the right facts' and 'reasoning well,' accuracy on medical tasks tracks knowledge much more closely. Math shows the opposite pattern (Does medical AI need knowledge or reasoning more?). In medicine, the answer usually depends on having the right fact, not on clever reasoning.

That has a surprising side effect. Knowledge and reasoning seem to sit in different parts of the network. Fact retrieval happens in the lower layers and reasoning adjustments in the higher ones. This helps explain why training a model to reason harder can make it better at math but worse at medicine (Why does reasoning training help math but hurt medical tasks?). So the instinct to make a medical model 'think more' can backfire. What it often needs is to know more, and to know it correctly.

Why not just keep fine-tuning new findings into the model? A formal result shows that the number of facts a model can memorize in its weights is capped by its size. Each new fine-tuning round also tends to overwrite older knowledge and erode general ability. Tool use, meaning looking things up, has no such cap: a small model can recall an unlimited number of facts through a simple lookup mechanism (Can models store unlimited facts without growing larger?). In a field where guidelines, drug data and evidence keep changing, the case for constant access to sources is partly a capacity argument, not only a freshness one.

Access alone isn't the whole answer, though. Models do better when they learn, step by step, when to retrieve and when to trust what they already know. Unneeded retrieval adds noise (When should language models retrieve external knowledge versus use internal knowledge?). Retrieval can also fail quietly: when a question doesn't share words with the fact it depends on, retrieval systems surface that fact far less often than a model that already has it in context (Why do retrieval systems fail on queries that never mention needed facts?). A clinical question phrased around symptoms may never match a source written around a diagnosis. One safeguard from another field is to answer only when the retrieved evidence supports the answer and to refuse otherwise. This gives up some coverage in exchange for not inventing things (Can RAG systems refuse to answer without reliable evidence?).

The less obvious takeaway is that current sources also keep people honest, not just models. In a study on real NEJM cases, clinicians with LLM help produced broader, more accurate lists of possible diagnoses than clinicians using search alone (Does LLM assistance help clinicians build better differentials?). But a fluent model can also become a loop that echoes back what you already believe. Powerful models increase the need to anchor answers in real data rather than reducing it (Do foundation models actually reduce our need for real data?). Seen this way, ongoing access to sources is less about patching a model's memory than about keeping AI-produced medical knowledge tied to evidence someone can actually check.