How AI Works (The Clinical Perspective)

Chapter 1: The 90-Second Referral Letter

---

Story: The Skeptic's 90-Second Conversion

Dr. Anita Mehta had been practicing cardiology for seventeen years when her hospital's CMO announced they were "integrating AI into clinical workflows." She remembers the exact words she muttered to the colleague sitting next to her in that all-hands meeting: "Great. Another EMR rollout that'll double my documentation time."

She had good reason for skepticism. She'd lived through the transition from paper charts to electronic medical records. She'd watched "revolutionary" clinical decision support systems generate alert fatigue so severe that physicians clicked through 90% of warnings without reading them. She'd sat through vendor presentations where non-clinicians explained how their technology would "transform" medicine, only to discover it transformed nothing except the number of clicks required to order a CBC.

So when a colleague suggested she try using a general-purpose AI assistant -- not a medical device, just a language tool -- to help with a particularly complex referral letter, she was unmoved. "I've been writing referral letters for seventeen years," she said. "I don't need a chatbot to do my job."

The letter in question was for a 62-year-old patient with a challenging presentation: new-onset atrial fibrillation with a rapid ventricular response, discovered incidentally during workup for a syncopal episode, complicated by a history of mechanical mitral valve replacement, chronic kidney disease (stage 3b), prior GI bleeding on anticoagulation, and newly diagnosed moderate aortic stenosis on echocardiography. The patient needed a referral to a cardiac electrophysiologist, but the clinical picture required careful contextualization -- the referring physician needed to understand the anticoagulation dilemma, the renal dosing constraints, and the question of whether the syncope was arrhythmogenic, valvular, or multifactorial.

Dr. Mehta estimated the letter would take 20 to 25 minutes to draft properly.

Her colleague pulled up an AI assistant on a laptop (not connected to their EHR -- this was just a demonstration) and said: "Give it the clinical details. See what happens." Dr. Mehta, more to prove a point than anything else, typed a prompt. Not a sophisticated one. Something like:

Write a referral letter from a cardiologist to an electrophysiologist for a 62-year-old male with new-onset A-fib with RVR found during syncope workup, history of mechanical MVR on warfarin, CKD 3b (GFR ~38), prior GI bleed on anticoagulation, and new moderate AS on echo. Need EP evaluation for rhythm management strategy given the anticoagulation challenges.

Ninety seconds later, she was reading a draft that was -- and these are her words -- "about 85% of the way there." The letter correctly identified the anticoagulation dilemma (bridging considerations with a mechanical valve, bleeding history, and renal dosing implications). It raised the appropriate clinical questions for the EP: rate vs. rhythm control strategy, consideration of catheter ablation to potentially reduce anticoagulation burden, and whether the syncope had an arrhythmogenic etiology requiring monitoring. It even noted the relevance of the aortic stenosis to hemodynamic tolerance of atrial fibrillation.

It was not perfect. It included a sentence about direct oral anticoagulant options that was irrelevant (the mechanical valve mandates warfarin). It used a slightly awkward phrase about "optimizing the therapeutic window." And it obviously lacked the specific vital signs, lab values, and echo measurements that would make the letter complete.

But the clinical reasoning scaffold was sound. The letter demonstrated an understanding of why this case was complex and what questions the receiving physician needed answered. Dr. Mehta spent four minutes editing the draft -- adding specific values, removing the DOAC reference, adjusting tone -- and had a better referral letter than she would have written from scratch, in a quarter of the time.

"I didn't change my mind about AI because it was impressive," Dr. Mehta said later. "I changed my mind because it gave me back fifteen minutes. Multiply that across a day of documentation, and you're talking about an hour. That's an hour I can spend with patients, or an hour I can leave the hospital earlier to see my kids."

This is not a story about AI replacing clinical judgment. Dr. Mehta caught the DOAC error because she is a cardiologist who knows that mechanical valves require warfarin. The AI didn't know her patient, didn't examine him, and couldn't make clinical decisions. What it did was eliminate the blank-page problem -- the cognitive load of organizing complex clinical information into a coherent narrative from scratch.

That distinction -- between AI as a drafting tool and AI as a clinical authority -- is the foundation of everything you will learn in this course.

---

Concept: Understanding AI Through a Clinical Lens

What Is a Large Language Model, and Why Should You Care?

A large language model (LLM) is the technology behind tools like ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), and medical-specific systems like Med-PaLM. The term sounds abstract, but the underlying principle is something every physician already understands intuitively: pattern recognition applied at scale.

Here is how it works, translated into clinical thinking:

Training = Exposure to a massive corpus of cases. An LLM is trained on billions of words of text -- books, articles, websites, code, and in some cases medical literature specifically. During training, the model learns statistical patterns: which words, phrases, and concepts tend to appear together, in what order, and in what context. This is conceptually similar to how a physician develops clinical intuition through thousands of patient encounters. You have seen enough presentations of acute coronary syndrome that you can recognize the pattern even when the presentation is atypical. The model has "seen" enough medical text that it can generate clinically plausible language.

Inference = Generating a differential, one token at a time. When you give an LLM a prompt, it generates a response by predicting the most likely next word (technically, the next "token") given everything that came before it. It does this iteratively -- word by word, sentence by sentence -- constructing a response that is statistically consistent with the patterns it learned. Think of it as generating a differential diagnosis: given the presenting symptoms (your prompt), what is the most likely diagnosis (response), considering the base rates of all conditions in the training data?

Context window = The clinical history it can hold in working memory. Every LLM has a "context window" -- the maximum amount of text it can consider at once (both your input and its response). Current models range from 8,000 to over 1 million tokens. This is analogous to the clinical information you can hold in working memory when reasoning about a case. More context generally means better responses, just as more complete history generally means better clinical reasoning.

The Differential Diagnosis Analogy

This analogy deserves expansion because it illuminates both the strengths and the critical limitations of AI.

When you generate a differential diagnosis, you are performing pattern recognition informed by prior probability (base rates), clinical context (the specific patient in front of you), and discriminating features (what makes one diagnosis more likely than another). You do not reason from first principles every time -- you match patterns and then apply clinical judgment to refine.

An LLM does something structurally similar but fundamentally different. It matches patterns in text and generates the statistically most probable continuation. This means:

Where the analogy holds:

• Both you and the LLM are better at common presentations than rare ones (training data, like clinical experience, is weighted toward prevalence)
• Both can generate plausible explanations for ambiguous presentations
• Both benefit from more context -- additional history improves both your differential and the model's output
• Both can exhibit "anchoring" -- you on a memorable case, the model on dominant patterns in its training data

Where the analogy breaks down -- critically:

• You update your assessment based on physical examination, real-time lab data, and the patient in front of you. The LLM has no access to any of this unless you explicitly provide it.
• You have a causal model of disease pathophysiology. You understand why troponin rises in myocardial injury. The LLM knows that "troponin" and "myocardial infarction" co-occur frequently in text -- but it does not understand the underlying biology.
• You know what you don't know. You have calibrated uncertainty. When a case exceeds your expertise, you consult. An LLM has no reliable mechanism for recognizing the boundaries of its competence. It will generate a confident response even when it is wrong.
• You are accountable. The LLM is not.

What AI Can and Cannot Do in Medicine Today

Let us be precise about the current state of the technology, as of early 2026. This is not speculative futurism -- these are tools in production use in clinical settings right now.

What AI Does Well in Clinical Practice

1. Documentation and administrative work This is the current sweet spot. AI excels at tasks where the output is text, the input is structured clinical information, and the physician reviews the result before it reaches a patient or a chart.

• Referral letters (as in Dr. Mehta's case): Given key clinical details, AI can organize them into a coherent, properly structured letter in seconds.
• Discharge summaries: Synthesizing admission course, procedures, medications, and follow-up instructions.
• Prior authorization letters: Generating clinical justification narratives that match payer requirements.
• Progress notes: Drafting SOAP notes from dictated or typed clinical details.

2. Ambient clinical documentation (AI scribes) Products like Nuance DAX Copilot, Abridge, and Nabla record the physician-patient encounter and generate a structured clinical note in real time. Early evidence suggests these tools reduce documentation time by 40-60% and improve physician satisfaction. The physician reviews and signs the note -- the AI does not write to the chart autonomously.

3. Clinical decision support (with guardrails)

• Differential diagnosis generation: Given a symptom complex, AI can generate a broad differential that may include diagnoses the physician hadn't initially considered. This is a cognitive aid, not a diagnostic tool.
• Drug interaction checking and dosing guidance: Some AI systems integrate with formulary data to provide real-time decision support, though these are distinct from general-purpose LLMs.
• Literature synthesis: AI can rapidly summarize recent publications on a clinical question, though citations must be verified.

4. Patient communication

• Translating clinical language into patient-friendly explanations at appropriate reading levels
• Drafting responses to patient portal messages
• Creating custom patient education materials for specific conditions and treatment plans

5. Research and education

• Summarizing journal articles and systematic reviews
• Generating study protocol drafts
• Creating case-based teaching materials
• Assisting with manuscript preparation and editing

What AI Cannot Do (and Should Not Be Trusted to Do)

1. Make autonomous clinical decisions. No general-purpose AI should determine diagnosis or treatment without physician oversight. Even FDA-cleared diagnostic AI tools (e.g., for diabetic retinopathy screening or certain radiological findings) operate within narrowly defined parameters with physician supervision.

2. Replace the physical examination. AI has no sensory input. It cannot hear a murmur, palpate an abdomen, or observe a patient's affect. Clinical information that exists only in the encounter room is invisible to AI unless you describe it.

3. Guarantee factual accuracy. Even the most advanced models produce errors. In medicine, error has consequences. Every AI output used in clinical work must be verified.

4. Maintain continuity of care. Each AI interaction starts fresh (unless specifically designed with persistent memory). It does not remember your patient from last visit, does not know their preferences, and cannot integrate the longitudinal relationship that defines good medicine.

5. Navigate ethical complexity. End-of-life discussions, goals-of-care conversations, breaking bad news -- these require human judgment, empathy, and the sacred trust of the physician-patient relationship. AI can help you prepare for these conversations. It cannot have them.

Clinical AI Tools vs. General-Purpose AI

This distinction matters enormously, and conflating the two is a common source of confusion.

FDA-cleared clinical AI tools are software devices that have been validated for specific clinical tasks through rigorous testing. Examples include AI algorithms for detecting pulmonary nodules on chest CT, identifying diabetic retinopathy on fundus photography, or flagging critical findings on ECGs. These tools:

• Are regulated as medical devices
• Have been tested against defined performance benchmarks
• Operate within narrow, specified use cases
• Are integrated into clinical workflows (typically within PACS or EHR systems)
• Have documented sensitivity, specificity, and failure modes

General-purpose LLMs (ChatGPT, Claude, Gemini, etc.) are not medical devices. They are language tools that happen to have been trained on text that includes medical content. They:

• Are not FDA-regulated for clinical use
• Have not been validated for specific diagnostic tasks
• Operate broadly across any topic (which is both their strength and their risk)
• Are accessed outside the clinical workflow (typically via browser or app)
• Do not have clinically validated performance metrics

Both categories have legitimate roles in a physician's practice, but the rules of engagement are different. You use an FDA-cleared diagnostic AI the way you use a validated lab test -- within its indicated parameters, with understanding of its limitations. You use a general-purpose LLM the way you use a medical textbook or a knowledgeable colleague -- as a resource to inform your thinking, never as the final authority.

HIPAA and AI: What You Must Know from Day One

This is not optional. This is not a "we'll figure it out later" topic. If you are going to use AI in any capacity that involves patient information, you must understand the privacy implications now.

Practical HIPAA guidelines for physicians using AI:

1. De-identify before you prompt. If you are using a general-purpose AI tool, strip all identifiers. Instead of "John Smith, DOB 3/15/1962, MRN 4427891," write "62-year-old male." Remove dates of service, locations, and any unique identifiers.

2. Use enterprise/healthcare-specific tiers. Some AI providers offer HIPAA-compliant enterprise plans with BAAs (e.g., Microsoft Azure OpenAI with BAA, certain Anthropic enterprise deployments, Google Cloud with BAA). If your organization has negotiated such an agreement, use the approved platform -- and only the approved platform.

3. Institutional AI tools have different rules. If your health system has deployed an AI scribe or documentation tool (like Nuance DAX within your Epic environment), it operates under the health system's BAA and data governance framework. These are designed for clinical use with PHI.

4. When in doubt, de-identify. The cognitive overhead of removing identifiers is minimal compared to the risk of a HIPAA violation. Develop the habit now.

5. Never upload clinical images, documents, or files containing PHI to general-purpose AI tools. This includes photos, PDFs of lab reports, or screenshots of EHR data.

---

Apply: Clinical Exercises

Now that you understand the conceptual framework, let us put it into practice with structured exercises that mirror real clinical work.

Exercise 1: The Referral Letter

Please structure as a formal referral letter with clinical question clearly stated." explanation="The expert prompt provides specific clinical details (nodule size, location, morphology, comparison imaging), relevant negatives (never-smoker, normal PFTs), family history context, failed empiric treatments, and a clearly articulated clinical question. This gives the AI the same information you would give a consulting colleague -- enabling a clinically useful draft rather than a generic template." />

Exercise 2: Patient Education

Format: Use headers, bullet points, and bold key terms. Keep total length under 2 pages." explanation="The expert prompt provides patient-specific context (occupation, education level, cultural background, specific concerns) that transforms a generic handout into a personalized educational tool. It specifies reading level, addresses the patient's actual worries, and requests practical guidance rather than abstract advice. The result is something you could hand to THIS patient, not just any patient." />

Exercise 3: Literature Synthesis

Exercise 4: Prior Authorization

---

Reflect: Calibrating Your Approach

You have now seen what AI can do, understood the mechanics behind it, tried it on clinical tasks, and -- critically -- identified where it fails. Let us consolidate.

The Physician's Framework for AI

Think about AI through the lens you already use for any new clinical tool or technology:

1. What is the evidence base? For documentation tasks, the evidence is strong and growing: AI reduces documentation time, improves note quality metrics, and is associated with improved physician satisfaction in multiple studies. For diagnostic tasks, the evidence is narrower and tool-specific. For treatment decisions, the evidence for general-purpose AI is insufficient. Match your trust to the evidence.

2. What is the failure mode? Every tool fails. Stethoscopes miss murmurs. CT scans have false positives. AI hallucinates. Knowing the failure mode is what separates competent use from dangerous use. AI's failure mode is confident inaccuracy -- and it fails silently, without alarm bells. Your verification is the alarm bell.

3. What is the risk profile of this specific task? Drafting a referral letter that you will review before sending: low risk. Generating a drug dose for a critical care patient and acting on it without verification: unacceptable risk. The same tool has different risk profiles depending on the clinical task and the verification step between AI output and patient impact.

4. Does this serve my patients? If AI helps you spend less time on documentation and more time on direct patient care, it serves your patients. If it introduces errors, violates privacy, or creates false efficiency that masks reduced quality, it does not. Keep the patient at the center of every decision about AI adoption.

What to Do Next

You don't need to transform your practice this week. Start with one task:

1. Pick the documentation task that annoys you most. The prior auth letter you dread. The referral letter for the complex patient. The patient portal message you keep putting off.
2. De-identify the relevant clinical information using the framework above.
3. Write a detailed, specific prompt -- not three words, but a paragraph that gives the AI the context it needs.
4. Evaluate the output critically. What did it get right? What did it get wrong? How long did editing take?
5. Repeat. Your prompts will improve with practice, just as your clinical skills improved with repetition.