AI in Daily Medical Practice — Module 1 of 6

AI Foundations for Clinical Practice

Handout — self-paced study material

Why This Module Starts Here — Augmentation, Not Outsourcing

Before any vocabulary, one distinction matters more than the rest of this module combined, because it decides whether everything that follows helps you or quietly makes you worse at your job.

A tool can be used two completely different ways. A physician can ask an AI system to generate a differential diagnosis and copy it into the chart without a second thought — that is cognitive outsourcing: the thinking has been handed over, not assisted. Or a physician can ask the same system to generate a differential, then interrogate it — what's missing, why did it rank this first, does it fit this specific patient — before deciding anything. That is cognitive augmentation: the tool extended the physician's reasoning without replacing it.

The technology is identical in both cases. The outcome is not. Every module in this curriculum, starting now, is built around one working rule:

The three-layer model behind this course

Medicine has always required physicians to carry a large amount of knowledge in their own heads. That hasn't changed and won't change — a physician who has to look up the danger signs of aortic dissection while a patient is actively dissecting has already lost time that matters. What has changed is how much everything else needs to live in memory. This course sorts knowledge into three layers, and it's worth understanding the sort before you start memorizing anything.

Every section below is labeled so you know which layer you're in and how hard to work at retaining it.

Section A — What AI Actually Is (KNOW)

You do not need to become a computer scientist. You need enough working vocabulary to read a product description, follow a conversation with an IT vendor, and know what question to ask when something goes wrong. Nine terms cover essentially everything you'll need for the rest of this course.

Artificial intelligence (AI)

A general term for software that performs tasks — recognizing patterns, generating text, interpreting images — that would normally require human judgment. It is a category, not a single technology.

Machine learning (ML)

The dominant technique behind modern AI: instead of being explicitly programmed with rules, the system is shown large amounts of example data and learns statistical patterns from it. A tool that learned to flag abnormal chest X-rays did so by being shown thousands of labeled X-rays, not by being given a list of rules.

Large language model (LLM)

A type of AI trained on enormous amounts of text, which predicts what text should come next given some input. This is what powers ChatGPT-style tools, many AI scribes, and most clinical-summary generators. It does not "know" facts the way a textbook does — it generates plausible-sounding text based on patterns, which is exactly why Section B matters.

Generative AI

Any AI that produces new content — text, images, summaries, draft notes — rather than only classifying or scoring existing input. LLMs are the most common form of generative AI in clinical use today.

Multimodal AI

A system that can work across more than one type of input — text and images together, for example. This is what lets a tool read a clinical note and interpret an accompanying retinal photo or ECG trace in the same session.

Training data

The examples a system learned from. This matters clinically because a tool is only as good as the population it was trained on — a diagnostic tool trained mostly on data from one country's hospitals may perform differently on your patients. You'll return to this directly in Module 3 and Module 5.

Prompt

The instruction or question you give a generative AI tool. Vague prompts produce vague, generic answers; specific prompts with relevant context produce more useful ones. This is a skill you'll build hands-on in later modules.

Inference

The moment the trained model actually produces an output for your specific input — for practical purposes, "inference" just means "the tool doing the task," as opposed to the earlier training phase that already happened before you ever opened the app.

Hallucination

When an AI system generates information that sounds fluent and confident but is factually wrong or entirely fabricated — an invented citation, a drug interaction that doesn't exist, a lab value that was never in the chart. Covered in full in Section B, because it is the single most important failure mode in this entire course.

Notice what's not on this list: no neural network architecture, no discussion of parameters or model sizes, no coding. That detail sits in the FIND layer — a curious reader can go deeper, but nothing later in this course depends on it.

Section B — How These Systems Fail (KNOW)

This is the section to read most carefully. Everything downstream in this curriculum — using AI for documentation, for diagnostic support, for anything — depends on you internalizing this before you touch a tool with a real patient.

Hallucination

An LLM generates text by predicting what word plausibly comes next, not by checking a database of verified facts. Most of the time, for common topics, this produces accurate output because accurate patterns dominate the training data. But the system has no built-in mechanism that distinguishes "I am certain" from "I am guessing" — both come out sounding equally confident.

Overconfidence and the illusion of competence

AI output rarely hedges the way a cautious colleague would. A tool will often generate a single confident-sounding answer even when the underlying evidence is thin, conflicting, or simply outside what it was trained on. The danger is not that the tool is wrong — every diagnostic aid is sometimes wrong — it's that the tool's wrong answers look identical to its right ones.

Bias

Because these systems learn from training data, they inherit whatever imbalances exist in that data. A tool trained overwhelmingly on one population may under-perform, or fail in ways that are hard to predict, on patients from a different population — different demographics, different disease prevalence, different presentation patterns. This is covered in depth in Module 4; for now, simply know that "the tool works well" is a claim that depends heavily on whom it was validated against, and is never automatically true for your patients.

Confirmation bias — the trap that matters most for you personally

Every one of the failure modes above becomes most dangerous at exactly one moment: when the AI's answer happens to match what you already suspected. That is precisely when scrutiny drops, because the output feels like confirmation rather than a claim that needs checking. This single insight is the seed of Module 4, which is built entirely around training a habit of extra vigilance at that exact moment.

Section C — The Clinical AI Landscape Today (FIND)

You don't need to memorize product names or pricing — that information changes constantly and belongs in the FIND layer. What's worth carrying in your head is the shape of the landscape: the four broad categories almost every clinical AI tool falls into.

A practical habit worth building now: whenever you encounter a new AI product, the first question is simply "which of these four is this?" — it tells you immediately what kind of scrutiny it needs and which later module's skills apply to it.

Section D — What Actually Fits Your Practice (THINK — worksheet)

This is a self-directed reflection, not reading. There is no live seminar for this section anymore — you work through it on your own, and it's worth doing with an actual pen or a saved document, because you'll draw on your answers again in Module 6 when you build your implementation plan.

You are not expected to have confident answers to all five yet — Modules 2, 3, and 5 will give you more to work with. The goal here is simply to start the habit of asking these questions before adopting anything, rather than after.

Module Summary