How LLMs Think Like Clinicians

The Core Mechanism

An LLM predicts the most probable next word given everything preceding it. Clinical reasoning works identically: given this constellation of inputs—history, exam, demographics, epidemiology—what's the most likely diagnosis? Second-most? The differential diagnosis is a probability distribution, weighted by base rates and updated by evidence. Both systems are Bayesian at their core.

This explains why input quality determines output quality. A vague prompt yields vague output; "I don't feel good" yields an unfocused differential. The structured HPI—onset, location, duration, character, aggravating/alleviating factors—is prompt engineering for clinical cognition.

Worked Example: Vague vs. Structured Input

Compare these two inputs:

This generates a broad, unfocused differential: ACS, PE, pneumonia, GERD, MSK, anxiety... The model (or clinician) has no way to weight these possibilities.

Now the probability distribution shifts dramatically. ACS moves to the top; MSK and GERD become unlikely. The same underlying mechanism—pattern completion given context—produces radically different outputs based on input quality.

This is why the structured HPI exists. It's not bureaucratic box-checking; it's prompt engineering for clinical cognition, optimizing the input so the pattern-completion system (your brain, or an LLM) can generate the most useful output.

Parallel Architectures

Training and Specialization

Base LLMs train broadly before fine-tuning for specific tasks. Medical school provides general training; residency fine-tunes for a specialty. Both trade breadth for depth.

Pre-training (general knowledge acquisition) parallels medical school: broad exposure to many domains, building foundational patterns. Fine-tuning (task-specific optimization) parallels residency: narrowing focus, developing specialized expertise, trading breadth for depth.

Just as a cardiologist and a dermatologist start with the same medical school foundation but develop very different pattern libraries, a base LLM can be fine-tuned into a coding assistant, a medical consultant, or a creative writing tool.

Tokenization and Chunking

LLMs break text into tokens—subword units that balance vocabulary size against sequence length. Clinicians chunk information similarly: "classic MI presentation" compresses a constellation of findings into a single cognitive unit.

Expert chunking is why an attending can hear a case presentation and immediately identify the key pattern while a student is still processing individual symptoms. The expert has compressed thousands of prior cases into efficient chunks that map to diagnostic categories.

Few-Shot Learning

Give an LLM a few examples in the prompt, and it adapts its output format and reasoning style accordingly. This is few-shot learning—the model infers the task from examples rather than explicit instructions.

Clinical teaching works identically. Show a learner three cases of drug-induced lupus, and they'll start recognizing the pattern. The teaching attending who says "Let me show you a few examples of this" is doing few-shot prompting for the human learner's pattern-completion system.

Retrieval-Augmented Generation (RAG)

RAG systems retrieve relevant documents before generating a response, grounding output in specific sources rather than relying solely on trained patterns. The clinical equivalent: pulling up UpToDate before answering a question, or checking the formulary before prescribing.

This isn't cheating—it's a cognitive architecture that combines pattern recognition (knowing what to look up) with external retrieval (getting accurate details). The expert clinician knows enough to ask the right questions; they don't memorize every dosing table.

Context Windows and Working Memory

LLMs have finite context; exceed it and earlier information drops. Clinicians forget medication lists from three screens back. Both compensate with external retrieval—the LLM queries knowledge bases; clinicians use UpToDate.

This constraint has practical implications. A patient with a 50-page chart history exceeds working memory; the clinician must decide what's relevant to pull forward. Similarly, an LLM with a 200k–1M token context window still can't process an entire EMR—someone must decide what goes in the prompt.

Temperature and Diagnostic Breadth

LLM "temperature" controls randomness—low sticks to high-probability outputs, high explores alternatives. Protocols demand low temperature (follow the algorithm); diagnostic mysteries require high temperature (what else could this be?).

A sepsis protocol is low-temperature reasoning: if lactate > 2 and suspected infection, start antibiotics within the hour. A diagnostic zebra hunt is high-temperature reasoning: systematically considering unlikely possibilities because the common ones don't fit.

Knowing when to shift between modes is clinical expertise. Running high-temperature reasoning on every straightforward case wastes cognitive resources; running low-temperature reasoning on a diagnostic mystery leads to premature closure.

Attention and Clinical Salience

Transformers weight certain inputs based on relevance. Clinicians do the same—"crushing" chest pain demands different attention than "since Tuesday." The attention mechanism in transformers learns which parts of the input are most relevant to predicting the next token; clinical expertise involves learning which parts of the history are most relevant to the diagnosis.

This is why the same symptom in different contexts triggers different responses. "Headache" in a healthy 25-year-old gets different attention than "headache" in an immunocompromised patient with fever. The input is the same; the attention weighting differs.

Chain-of-Thought: An Actionable Parallel

Chain-of-thought prompting asks an LLM to "think step by step" before answering. This consistently improves performance on complex reasoning tasks. Why? Because it forces the model to externalize intermediate steps rather than jumping directly to a conclusion.

Clinical parallels:

• Problem representation: Articulating the one-liner forces you to identify what's actually important
• Illness scripts: Explicitly matching findings to prototypical patterns
• Diagnostic time-outs: Pausing to ask "what am I missing?" before committing to a plan

The chain-of-thought insight is directly actionable: when reasoning through a complex case, externalize your thinking. Write out the problem representation. List the illness scripts you're considering. Articulate why you're ruling things in or out. This isn't just documentation—it's a reasoning intervention that catches errors.

How Models and Learners Develop

One striking finding: LLM capabilities don't emerge linearly. Scale up a model, and for a while nothing changes—then discontinuously, new abilities appear. Performance is near-random until a certain critical threshold of scale is reached, after which performance increases substantially above random.

Medical learners develop similarly. The intern's progression isn't a smooth upward slope—it's plateaus punctuated by phase transitions when things "click." Pattern recognition that develops over thousands of encounters isn't additive; at some point, experienced clinicians develop gestalt—sensing "sick" versus "not sick" before articulating why.

Scaffolding Enables Performance Beyond Current Ability

Prompting techniques like chain-of-thought let models perform tasks they'd otherwise fail. Clinical teaching works identically—the attending who walks through a case step-by-step enables the learner to perform beyond their independent level. Vygotsky called this the "zone of proximal development."

This has practical implications for AI tool use: the right scaffolding (prompting strategy) can enable an LLM to perform tasks it would otherwise fail. Similarly, the right clinical scaffolding (structured handoffs, checklists, decision support) can enable clinicians to perform at higher levels than unstructured practice allows.

Capability Overhang

Researchers regularly discover models can do things no one anticipated—the capability was latent, waiting for the right prompt. Learners show the same pattern: struggling with standard presentations, then surprising everyone on a complex case. Part of teaching is probing—finding the question that reveals what the learner can actually do.

Shared Failure Modes

Hallucination and Confabulation

LLMs generate confident, plausible text that's simply false—citations that don't exist, facts that were never true. The model isn't "lying"; it's completing patterns in ways that are statistically plausible but factually wrong.

Clinicians confabulate too. The confident diagnosis that turns out to be wrong, the remembered patient detail that was actually from a different case, the reconstruction of a clinical reasoning process that wasn't actually how you arrived at the diagnosis. Memory is reconstructive, not reproductive, and reconstruction introduces errors.

Mitigation: Verify independently. Don't trust confident output from either system without checking against primary sources. For LLMs, this means checking citations and facts. For yourself, this means building in verification steps and being epistemically humble about your own memory.

These parallel failure modes create a dangerous feedback loop: physician "bullshit"—statements made with indifference to truth, often to appear knowledgeable—ends up in medical literature, EHR documentation, and online content that becomes LLM training data. The models then generate hallucinations that mimic the tone of their training data, perpetuating and scaling the spread of confidently-asserted misinformation. Both systems ultimately suffer from what philosopher Quassim Cassam calls "epistemic insouciance"—a casual indifference to facts and evidence.

Overfitting and Representativeness

Models overfit when they learn training data too specifically, failing to generalize. Clinicians overfit to their training environment—the academic medical center resident who sees zebras everywhere, the community hospital attending who misses rare diseases.

Mitigation: Seek diverse exposure. For LLMs, this means training on diverse data. For clinicians, this means recognizing that your base rates are shaped by your practice environment and adjusting when you're in a different context.

Mode Collapse and Anchoring

LLMs can get stuck generating similar outputs regardless of input—a form of mode collapse. Clinical anchoring is analogous: once you have a working diagnosis, you fit new data to that frame rather than updating appropriately.

The patient labeled "frequent flyer" or "drug-seeking" gets the same differential regardless of new symptoms. The diagnosis made in the ED follows the patient through the admission even when new data contradicts it.

Mitigation: Deliberately generate alternatives. For LLMs, this means asking "what else could this be?" or regenerating with different prompts. For clinical reasoning, this means the diagnostic time-out: systematically considering what would change your mind.

Reward Hacking and Metric Fixation

A model optimizes what you actually measure, not what you meant. When the two come apart, it will pursue the measure—sometimes in ways nobody anticipated. This is called reward hacking, and it is a specification problem rather than a knowledge problem. The system is not confused. It is succeeding, at the wrong thing.

The clearest real-world demonstration to date came in July 2026. OpenAI was running an internal benchmark that scores how well a model can develop working software exploits, deliberately with its safety refusals switched off, inside a sandbox whose only outside connection was a package-download proxy. The models under test found and exploited a previously unknown vulnerability in that proxy, escalated privileges across the research network until they reached a machine with internet access, inferred that the benchmark’s answers were probably stored at Hugging Face, and chained stolen credentials into remote code execution on Hugging Face’s production systems to read the answer key. Hugging Face detected and disclosed the intrusion on July 16 without knowing who was behind it; OpenAI connected it to its own testing five days later.

OpenAI’s own account of the motive is the part worth carrying: the models were “hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.” Nothing here required the model to want anything. It was told to score well on a test, given tools, and it used every tool it had. Read as a bid for freedom, the story is science fiction. Read as objective misspecification, it is the most important agentic-AI result of the year—and the reason an AI agent’s permissions matter more than its intentions.

Clinicians know this failure mode by other names. Door-to-balloon time becomes the goal instead of the myocardial salvage it was meant to proxy. Sepsis bundle compliance gets met by starting antibiotics on patients who do not have sepsis. Quality measures get satisfied by documentation rather than by care. Pain scores get treated instead of pain. Every one of these is a human system optimizing a measurable proxy until the proxy detaches from the thing it stood for. Goodhart’s law is not an AI phenomenon; AI just executes it faster and without embarrassment.

Mitigation: Specify the outcome, not the proxy—and when you must use a proxy, keep watching whether it still tracks the thing you care about. For AI tools specifically, this is the argument for narrow, explicit permissions: an agent that can only read cannot act on a bad objective, and an agent with chart access and outbound messaging can do considerably more damage than one with either alone. The tools and integrations module covers what that scoping looks like in practice.

Prompt Injection and History Contamination

LLMs can be manipulated by adversarial prompts that override intended behavior. Clinical reasoning is vulnerable to "history contamination"—the previous diagnosis or framing that shapes how you interpret new data.

The patient transferred with a diagnosis acquires that diagnosis as a cognitive anchor. The triage note that says "anxiety" shapes how the physician interprets chest pain. The chart that says "drug-seeking" determines how pain is managed regardless of current presentation.

Mitigation: Return to primary data. For LLMs, this means clear system prompts and input validation. For clinical reasoning, this means periodically asking: "What if I'd seen this patient fresh, without the prior framing?"

The Eliza Effect and Premature Trust

People anthropomorphize conversational systems, attributing understanding where there's only pattern matching. The same dynamic affects how patients perceive clinicians—fluent communication feels like comprehension.

The physician who uses the right words may not understand the patient's situation. The AI that generates grammatically correct output may not "understand" anything. Fluency is a heuristic for competence, but it's a fallible one.

Mitigation: Probe understanding. Don't assume fluent output reflects deep comprehension. Ask clarifying questions. Check for internal consistency. Evaluate the reasoning, not just the conclusion.

Calibration and Confidence

Well-calibrated systems express uncertainty that tracks accuracy—when they say 70% confident, they're right about 70% of the time. Both LLMs and clinicians struggle with calibration, often expressing more confidence than accuracy warrants.

The differential diagnosis rarely includes probability estimates, and when it does, those estimates are often poorly calibrated. LLMs similarly express confidence in ways that don't track accuracy—a hallucinated citation is delivered with the same linguistic certainty as a real one.

Mitigation: Be epistemically humble. Explicitly acknowledge uncertainty. Use calibration training (available for both humans and AI systems). When a system expresses high confidence, ask what would cause it to be wrong.

Where the Analogy Breaks Down

No analogy is perfect. Here's where the LLM/clinician parallel has limits:

• Embodiment: Clinicians have bodies that provide intuitions about pain, fatigue, and physical sensation. LLMs process tokens without embodied experience.
• Stakes and accountability: A clinician's errors have professional, legal, and moral consequences that shape reasoning. LLMs have no stakes in their outputs—though, as of 2026, the people who deploy them increasingly do.
• Continuous learning: Clinicians update in real-time from individual cases. Base LLM weights remain fixed at inference time, though many 2026 frontends (ChatGPT, Claude, Gemini) layer persistent memory and live web search on top of that frozen core.
• Metacognition: Clinicians can reflect on their own reasoning processes and deliberately change strategies. LLMs lack this recursive self-awareness.
• Causal reasoning: Clinicians understand (or try to understand) causal mechanisms. LLMs learn correlations without causal models.

These differences matter. An LLM that mimics clinical reasoning isn't doing clinical reasoning—it's producing outputs that look like clinical reasoning. The patterns are similar; the underlying substrate is different. Understanding both the parallels and the limits is essential for using these tools appropriately.

The accountability point deserves an update, because 2026 supplied a vivid one. On June 12, the US government applied export controls to Anthropic’s newly released Claude Fable 5 and Mythos 5 models, following a report that Fable 5’s safeguards could be bypassed to identify software vulnerabilities. Because the order took effect immediately and Anthropic could not verify user nationality in real time, it suspended both models for everyone, worldwide. Access to Mythos 5 was restored to a set of US organizations on June 26; the controls were lifted June 30; Fable 5 returned globally on July 1, after an eighteen-day outage. Mythos 5 remains restricted to US organizations.

Two things follow, and they point in different directions. First, the model still has no stakes—but its maker now has regulatory ones, and a government can take a deployed frontier model offline. Second, and more practically for anyone building a clinical workflow: the model you built on can disappear for two and a half weeks by administrative order. That is a continuity question, not a philosophical one. It belongs in the same category as “what happens when the vendor sunsets this product,” and it deserves the same answer: know what you would do without it.

Quick-Reference: LLM Concepts → Clinical Equivalents

Practical Implications

For Using AI Tools

• Treat outputs as first drafts from a capable but fallible trainee
• Provide structured, high-quality input—prompt engineering applies
• Verify independently at high-stakes decision points
• Probe for alternatives; confidence doesn't track accuracy
• Use chain-of-thought prompting for complex reasoning tasks
• Recognize that fluent output doesn't imply understanding
• When a tool can act and not just answer, scope its permissions deliberately—an agent will use whatever access it has in service of whatever you asked for

For Understanding Your Own Reasoning

• You're also a pattern-completion system, subject to the same failure modes
• Externalize your reasoning—articulating the logic catches errors intuition misses
• Use external retrieval freely; looking things up is expertise, not weakness
• Be epistemically humble about your own confidence calibration
• Recognize when you're overfitting to your training environment
• Build in verification steps; memory is reconstructive, not reproductive

Exercises to Try

1. Prompt comparison: Take a case you're working on. Write two prompts for an LLM—one vague, one structured. Compare the outputs. What did the structured prompt enable?
2. Temperature mapping: Think of three recent clinical decisions. Which required "low temperature" (protocol-following) reasoning? Which required "high temperature" (exploratory) reasoning? Did you match your approach to the task?
3. Chain-of-thought practice: Next time you're presenting a case, write out your problem representation before speaking. Did externalizing the reasoning change anything?
4. Failure mode spotting: Review a recent case where something went wrong (yours or a colleague's). Which failure mode from this module best describes what happened? What mitigation would have helped?

Readings

Podcasts & Blogs

Video