Understanding generative AI
Large language models, tokens, hallucinations, and a quick tour of prompting — with healthcare examples and clear rules of the road.
A large language model (LLM) is a very large pattern-prediction engine trained on enormous amounts of text. It doesn't "know" facts the way a textbook does; it has learned which words tend to follow which words in which contexts.
Words → tokens → predictions
Large language models break text into small chunks called tokens. Then they predict the next most likely token, over and over, until a sentence forms.
The model never "decides" the whole sentence at once. It picks one likely token, then conditions on what it just wrote to pick the next.
A chatbot, watching itself work
From source to output: what did the AI change?
Generative AI does not copy your information — it writes something new from patterns, context, and your instructions. Run the experiment below and watch exactly what gets kept, reworded, dropped, or invented.
Jordan is a 10-year-old seen for intermittent abdominal pain for 3 months. Growth has remained normal. There is no vomiting, gastrointestinal bleeding, weight loss, or fever. Laboratory testing performed by the pediatrician was normal.
Let's ask generative AI to summarize this information.
Prompt engineering, in two examples
The same model, two prompts. The clearer the request, the more useful the output.
"Write something for my patient about their asthma."
"Write a 5th-grade-reading-level handout for a 10-year-old with mild persistent asthma. Cover daily controller use, rescue inhaler use, and when to call the clinic. Keep under 200 words."
A hallucination happens because the model optimizes for fluent, likely-sounding language — not for truth. It has no internal "I don't know" signal by default. That's why every clinically meaningful output must be verified against a trusted source.
- Verify medications, doses, and guidelines independently.
- Treat polished tone and length as style, not as evidence.
- The more confident-sounding the answer, the more important the verification.
Where generative AI fits in clinical work
Today's safest uses are tasks where a clinician reviews every word before it reaches a patient or a chart.
Rewrite a discharge plan at the right reading level — you edit and sign.
Distill a long admission course into a one-paragraph handoff for review.
Explain a pathway, drug class, or guideline in plain language for teaching.
Not for independent medical decisions. Generative AI shouldn't autonomously diagnose, prescribe, triage, or decide disposition. It can support the clinician — it cannot replace clinical accountability.
Generative AI should always have human oversight
Every AI-generated word that touches a patient — a discharge instruction, a message, a chart note — must be reviewed by a qualified clinician before it's used. The model writes the draft. The clinician owns the decision.
- LLMs are sophisticated autocomplete: predict next token, then the next, then the next.
- Tokens are small chunks of text — sometimes whole words, sometimes parts.
- Fluent and confident does not mean true — hallucinations are an inherent risk.
- Better prompts produce better output: state the audience, format, and length.
- Use generative AI for drafts and education — never for autonomous clinical decisions.
Generative AI concept cards
Flip each card for a deeper definition and a clinical example.
An ambient scribe drafts a note that includes a medication dose the patient didn't mention. The safest next step is to…
Check your understanding
Five quick questions on LLMs, prompts, and safe use.
You're not signed in — this score won't be saved or counted toward your level average. Sign in first to record it.
