You Already Do This
You already do pattern recognition for a living.
When you look at a CT scan and immediately notice that the tumor is invading the adjacent blood vessel, you are not running through a textbook algorithm step by step. You are drawing on thousands of prior images, years of training, and an intuitive sense of what "looks right" versus what "looks wrong." You cannot fully articulate how you do it. You just... see it.
That is, at a very high level, what AI does. It takes in enormous amounts of data, learns patterns, and makes predictions based on those patterns.
The difference: instead of learning from a few thousand cases over a career, AI systems learn from millions or billions of data points. And instead of images, the AI systems you hear about most often - like ChatGPT and Claude - learned from text.
That is the whole mystique, demystified. Now let's get specific.
AI Is Already All Around You
Before we go further: you have been using AI for years. You just did not call it that.
Adaptive cruise control in your car? That is AI. Autocorrect in Word or your phone? AI. Auto-segmentation in your treatment planning system? AI - and if you have ever used auto-contours, you have already integrated AI into your clinical workflow.
AI Is Not One Thing
"AI" is not a single technology. It is a broad umbrella covering everything from your spam filter to the chatbot writing poetry. Context matters. When someone tells you "AI will change medicine," the useful follow-up question is: which AI? Doing what?
Tools like OpenEvidence use LLMs to provide citation-grounded clinical decision support. That is very different from the auto-contouring you use in treatment planning. Let us get specific.
The Jargon, Translated
| Term | What It Means | Example |
|---|---|---|
| Artificial Intelligence (AI) | Broadest term: any computer system performing tasks requiring human intelligence | Your spam filter, autocomplete, contouring software |
| Machine Learning (ML) | A subset of AI that learns patterns from data instead of following explicit rules | Auto-segmentation, dose prediction, outcome modeling |
| Deep Learning | A subset of ML using neural networks with many layers - powers most modern breakthroughs | Image recognition, auto-contouring, voice recognition |
| Large Language Models (LLMs) | ML focused on text - trained on massive text datasets to understand language patterns | ChatGPT, Claude, Gemini |
| Generative AI | AI that creates new content rather than just classifying existing content | ChatGPT writing text, Midjourney creating images |
Think of it as nested circles: AI contains ML, ML contains LLMs, and LLMs are a type of generative AI. When people say "AI" in casual conversation right now, they usually mean LLMs. But the field is much broader than chatbots.
How LLMs Actually Work (No Math Required)
The core concept is deceptively simple: LLMs predict the next word.
That is it. When you type a prompt into ChatGPT, the model looks at your text and asks itself: given everything I have seen in my training data, what word is most likely to come next? Then it generates that word, adds it to the sequence, and asks again. It does this hundreds or thousands of times until it has produced a full response.
This is why the technical term is "next-token prediction." A token is roughly a word or part of a word, and the model is just predicting them one at a time in sequence.
The Grocery Store Analogy
To understand why this matters - and why models sometimes get things wrong in very specific ways - think of a grocery store.
Words that are related to each other end up in the same aisle. "Milk," "cheese," and "yogurt" cluster together in the dairy section. "Bread," "flour," and "baguette" are in bakery.
The model does something similar: during training, it organizes tokens into a vast mathematical space where similar concepts end up near each other. This is called an embedding space, but think of it as a massive grocery store where words that appear in similar contexts are shelved in the same neighborhood.
This clustering is what makes LLMs powerful - it is how the model "understands" that nasopharynx and oropharynx are related concepts, or that cisplatin and carboplatin serve similar roles.
But it is also what causes a very specific type of error: the model can sometimes reach for the wrong item from the right aisle.
It is like reaching for the cheddar and accidentally grabbing the gouda - they are right next to each other on the shelf, and in the dark, they feel the same.
This is exactly what happens with hallucinated citations. If the model has learned that "Fletcher" is strongly associated with brachytherapy literature, and you ask for a citation about brachytherapy, the model will confidently generate something like "Fletcher et al., Int J Radiation Oncology Biol Phys, 2019" - not because it looked up that paper, but because "Fletcher" + "brachytherapy" + "Red Journal" are all clustered together in its token space.
It sounds perfectly plausible. The author name is real. The journal is right. The topic fits. But the specific paper may not exist, because the model assembled it from nearby tokens rather than retrieving a real reference.
This is not random noise. It is structured confabulation - the model grabs from the right aisle but sometimes picks the wrong product. And that is precisely why it is so convincing and so dangerous when you treat it as a search engine.
The Transformer Architecture
The architecture that makes all of this work is called a transformer. You do not need to understand the math, but the key innovation is something called "attention" - the model can look at all the words in your prompt and figure out which words are most relevant to predicting the next one.
When you ask about "PORT for nasopharyngeal carcinoma," the model attends to "PORT," "nasopharyngeal," and "carcinoma" as the important context, not "for."
This is trained at massive scale. Models like GPT-4 were trained on hundreds of billions of words. Through that training, they absorbed patterns about medicine, law, coding, history, cooking - essentially everything humans have written about online.
The result: a system that can produce remarkably coherent, contextually appropriate text on almost any topic.
Why It Seems Smart but Is Not "Thinking"
This is the most important conceptual point in this entire article.
LLMs do not understand anything. They do not have beliefs, knowledge, or intentions. They are extraordinarily sophisticated pattern-matching systems that produce text which looks like it was written by someone who understands the topic.
When you ask an LLM about the RTOG 0617 trial and it gives you a detailed, accurate summary, it is not because the model "knows" that trial. It is because during training, it saw enough text about RTOG 0617 that it can reproduce the statistical patterns associated with that topic. The output resembles understanding, but the process is fundamentally different.
Set the Right Expectations
An LLM can help you draft, brainstorm, synthesize, and think through problems. But it can also produce confidently wrong text that sounds authoritative. It does not know when it is wrong. It has no internal fact-checker. That is your job.
Think of it this way: the model is an incredibly well-read colleague who has read everything but experienced nothing. It can recall patterns from the literature with impressive breadth, but it has never treated a patient, never seen a complication, and never had to make a judgment call at 2 AM.
You bring the clinical judgment. It brings the pattern matching.
Why Should You Care? (Honestly, This Is Exciting)
Let me be direct: this technology is genuinely exciting. Not in the breathless Silicon Valley "disrupt everything" way, but in the way that feels like the first time you realized what you could do with a CT sim instead of film-based planning. It is a step change in capability, and we are right at the beginning.
The reasons that actually matter:
Your patients are already using these tools. They are asking ChatGPT about their diagnosis, their treatment options, their side effects. They are walking into your clinic with AI-generated questions and AI-generated anxieties. Whether you like it or not, AI is already part of the patient-physician conversation. You can be the informed voice in that conversation, or you can be caught off guard by it.
The industry is coming for your attention. OpenAI launched ChatGPT Health. Anthropic is building Claude for healthcare. Every vendor at ASTRO is selling you "AI-powered" something. Some of it is genuinely good. Some of it is a regression model with a marketing team. If you do not understand the fundamentals, you cannot tell the difference.
It is a technology, not a philosophy. Choosing not to engage with AI in 2026 is a bit like choosing not to use computers in 2010. You can do it. Nobody will stop you. But you are voluntarily excluding yourself from tools that are becoming increasingly integrated into how medicine is practiced, taught, and researched.
There may be a genuine responsibility here. The verdict is still out on whether AI makes us better doctors - the studies are early, the evidence is mixed, and the hype far exceeds the data. But if there is even a reasonable chance that these tools improve clinical reasoning, research productivity, or patient communication, do we not have an obligation to at least explore them?
The people around you are already using these tools. Colleagues, residents, and fellows are using ChatGPT and Claude to study, draft, and research. If you want to collaborate effectively - and if you want to help others use these tools responsibly - you need to understand what they can and cannot do.
And frankly, it is fun. Once you get past the initial awkwardness of talking to a chatbot, there is something genuinely delightful about having a tool that can brainstorm with you at midnight, summarize a paper in thirty seconds, or help you think through a case from an angle you had not considered.
It does not replace your expertise. It extends it.
A Word on Data Privacy (Beyond PHI)
Everyone knows not to put protected health information into public LLMs. But that is not the only thing you need to protect.
When you paste text into ChatGPT, Claude, or any cloud-based AI tool, you are giving that company access to whatever you input. That includes:
- Unpublished research data
- Grant proposals under review
- Manuscripts before publication
- Intellectual property
Even if it is not PHI, if it is sensitive or proprietary, it does not belong in a public-facing LLM.
For truly sensitive work - analyzing preliminary data, drafting patents, working with confidential institutional information - consider local LLMs like Ollama that run entirely on your own machine. They are not as capable as the frontier models, but they keep everything private.
Think of it this way: if you would not email it to a stranger, do not paste it into a chatbot.
The Bottom Line
AI is pattern recognition at scale. LLMs are pattern recognition applied to language. They predict the next word, and they are very good at it - good enough to be genuinely useful for clinical work, research, and education.
But they are tools, not oracles. They do not think. They do not know. They predict.
Once you internalize that distinction, you are already ahead of most people talking about AI right now.