What it is:AI, and specifically generative AI built on large language models (LLMs), is now sitting inside the tools you already use for school and will use in practice. It isn't a future thing to prepare for, it's already embedded in search, documentation, note-taking, and clinical decision support.
The core problem:These tools are fluent, fast, and confident, but they don't reason and they don't know when they're wrong. They predict plausible-sounding text. That means the output can be useful and dead wrong at the same time, and heavy reliance on them has a measurable cost to your own critical thinking.
What you do about it:Treat AI as a tool you direct, not an authority you defer to. Verify everything against a real source, never feed it protected health information, and use it in a way that builds your clinical reasoning instead of quietly replacing it.
The organizing mental model for this whole topic is a set of nested circles: Artificial Intelligence ⊃ Machine Learning ⊃ Generative AI ⊃ Large Language Models ⊃ GPTs(the specific architecture behind ChatGPT). Every buzzword you hear fits somewhere in that nest. Once you can place a term in the right ring, the jargon stops being confusing.
People use "AI," "machine learning," and "ChatGPT" as if they're interchangeable. They aren't. Each term describes a smaller, more specific ring inside the one before it.
| Type | What it means | Examples |
|---|---|---|
| Weak Intelligence | Basic systems that perform one specific task with no real understanding | Spell check, email filters, simple chatbots, recommendation systems |
| Artificial Narrow Intelligence (ANI) | AI designed to do a narrow task well | Siri, Alexa, Google Search, facial recognition, ChatGPT |
| Artificial General Intelligence (AGI) | Would understand, learn, and apply knowledge across a wide range of tasks, like a human | Hypothetical, doesn't exist yet |
| Artificial Super Intelligence (ASI) | Would surpass human intelligence in creativity, problem-solving, and emotional intelligence | Hypothetical, science fiction territory |
ChatGPT, Claude, and Gemini are ANI, not AGI.They feel general because language is flexible, but they are still narrow systems performing one kind of task (predicting text) extremely well. AGI and ASI are not in clinical use anywhere today, no matter how a headline frames it.
Understand this section and the rest of the topic (why prompting matters, why hallucination happens, why drift happens) stops being a list of disconnected warnings and starts being obvious.
Older language models processed text one word at a time, in order. The architecture behind modern LLMs (Vaswani et al., 2017) instead looks at the whole input at onceand calculates which words matter most to which other words, an attentionmechanism. That's stacked in layers to make the process more efficient. This is why a modern LLM can track a pronoun back to a noun three paragraphs earlier, hold context across a long conversation, and generate a coherent multi-sentence answer instead of just a plausible next word.
Underneath the attention mechanism, the model is still doing one thing: predicting the statistically most likely next token given everything it's attending to. It is not reasoning about truth, it is pattern-matching toward fluency.That's exactly why a vague prompt gets a vague, generic answer, and a specific, context-loaded prompt gets an answer anchored to what you actually meant. Better prompting and better output come from the same mechanism.
A trained model is a snapshot. The world keeps moving after that snapshot is taken, new drugs, new guidelines, new events, and the model's performance quietly degrades as the environment it's answering questions about drifts away from the environment it was trained on. Fixing that requires recalibrating (retraining) the model. This is also why an LLM's knowledge cutoff date matters clinically: ask it about a guideline updated after that date and it may confidently give you the old one.
Because the model is optimizing for statistically plausible language, it can produce confident, well-formatted, entirely wrong information, and it will read exactly as convincingly as the correct answer. Never let polish substitute for verification.
Vocabulary you'll hit in the wild, split into who's building it and how you'd actually use it.
| Commercial LLMs | |
| Model | Maker |
|---|---|
| Gemini (formerly Bard) | |
| ChatGPT | OpenAI |
| Claude | Anthropic |
| Command R | Cohere |
| Jurassic | AI21 |
| Open-source LLMs | |
| Model | Origin |
| Alpaca | Stanford |
| DeepSeek | High-Flyer (Hangzhou) |
| Dolly | MIT |
| Gemma | |
| LLaMA | Meta |
Because it's grounded to your uploaded sources, NotebookLM is structurally lower-risk for hallucination than an open chat window when you're studying: it can't invent a fact from nowhere the way a general LLM can, it can only misread or misquote what you gave it. That's a meaningfully different failure mode, and a much easier one to catch.
Three distinct levers exist for improving what an LLM gives you back, and they operate at completely different levels.
| Lever | What it changes | Who does it |
|---|---|---|
| Prompt engineering | Your input to the model. The model itself is untouched. | You, every time, free, instant |
| Retrieval-augmented generation (RAG) | Grounds the model's answer in a specific document or knowledge source before it responds | Built into tools like NotebookLM; you supply the source |
| Fine-tuning | Retrains the model itself on additional data, changing its baseline behavior everywhere | Done by developers, not end users |
Bad output because your ask was ambiguous → fix the prompt.Bad output because the model doesn't know your specific course material or a niche protocol → you need RAG(feed it the document, or use a tool built around retrieval). Bad output because the model's behavior needs to change across every conversation → fine-tuning, and that's not a lever available to you as a user.
Five categories of failure to hold in mind every time you open an AI tool.
| Risk | What it looks like |
|---|---|
| Accuracy | Confident, fluent, factually wrong output. The default failure mode, not the exception. |
| Sensitive information | Data you type into a public tool can be stored, logged, or used to train future models. This is a hard HIPAA problem. |
| Bias and fairness | Models trained on real-world data inherit real-world bias, and can reproduce it in clinical framing, risk language, or recommendations. |
| Emotional intelligence | The model has no actual empathy or judgment about a patient's emotional state, it's simulating the language of empathy. |
| Ethical use | Covers everything from plagiarism to undisclosed AI authorship to using AI in place of your own clinical judgment. |
Never put HIPAA-protected information into an AI toolunless it's a system your institution has specifically vetted and approved for that use. A public LLM is not a covered entity's secure system by default.
A checklist of angles to evaluate any AI tool or use case against, not just "does it work."
| Value | Asks |
|---|---|
| Efficiency | Does it actually save time and effort, net of the time spent verifying it? |
| Validity | Does it measure or answer the thing it claims to? |
| Reliability | Does it give consistent output for the same input over time? |
| Viability | Is it actually sustainable to use in a real workflow, not just a demo? |
| Governability | Can it be controlled, audited, and corrected when it's wrong? |
| Transparency | Is it clear how the output was generated and what it was trained on? |
| Sustainability | Every AI prompt consumes real energy. This is now a measured, publicly reported cost, not a hypothetical one. |
| Equity | Does access to and benefit from the tool distribute fairly, or does it widen an existing gap? |
This is the mechanism section for the "AI and your learning" half of the topic. If attention explains how LLMs generate output, this explains what leaning on that output does to yourbrain.
| Term | Definition |
|---|---|
| Cognitive load | The amount of mental effort a task takes, factoring in the setting and what makes it harder or easier |
| Cognitive offloading | Externalizing a cognitive task to an outside tool to reduce that load |
| Cognitive debt | Accumulated loss of higher-order thinking ability from repeated offloading, especially with AI tools |
| Metacognition | The ability to identify, evaluate, and regulate your own thinking process |
| Benefits of offloading | Risks of over-relying on it |
|---|---|
| Frees mental resources for more complex, dynamic work | Cognitive laziness |
| Shifts focus from memorizing to knowing where to find it (the "Google effect") | Cognitive dependence |
| Improves efficiency and performance in the moment | Loss of cognitive autonomy, reduced critical engagement and accountability, decline in cognitive function and memory retention |
This isn't just intuition. Gerlich (2025) found r = +0.72between AI reliance and degree of cognitive offloading, r = -0.75between self-reported cognitive offloading and validated critical-thinking measures, and r = -0.68between AI reliance and decline in critical thinking. As Gerlich put it, unchecked automated decision-support in healthcare risks "a workforce that is highly efficient, yet potentially less capable of independent problem-solving and critical evaluation."
Neuroplasticityis the brain's ability to reorganize itself, structurally and functionally, in response to learning and challenge, the same way physical exercise strengthens a muscle under load. Skip the load and you skip the adaptation. "The absence of difficulty leads to cognitive debt."Every shortcut that removes a moment of productive struggle is a rep you didn't get.
Metacognition mediates offloading.The four moves, planning, monitoring, evaluation, and self-reflection, are what determine whether reaching for AI is a smart delegation of a tedious task or a quiet erosion of a skill you actually needed to build. Offloading isn't inherently bad; offloading withoutmetacognitive awareness of what you're giving up is the problem.
Not all AI-assisted work is created equal. The structure of the assignment determines how much offloading risk it carries.
No AI assistance at all. Maximum skill-building, zero efficiency gain.
Task completed entirely by the AI. Maximum offloading risk, minimal learning.
AI used without any structured instruction on how. The default, and the riskiest hybrid.
Trained specifically to avoid offloading and stimulate critical thinking. The target.
Notice the order: reflection comes first, AI-as-critic comes last. You're never asking the model to generate your first draft or your original thinking, only to fact-check and pressure-test thinking you already did. That single sequencing choice is what separates guided use from the unguided default.
The same AI tool carries very different risk depending on who's holding it.
| Experts | Novices | |
|---|---|---|
| Can they catch AI errors? | Easily spot duplicated ideas, irrelevant information, and wrong information | Lack the skills to recognize inconsistencies, especially early in the learning curve |
| Overall risk of using AI here | Low-risk integration | High-risk usage |
Heading into APPE rotations, you are a novice in every new domain you land in, by definition. That's exactly the condition under which AI use is highest-risk, and exactly when it will feel most tempting to lean on it because you don't yet have the pattern recognition to know when it's wrong.
| Low risk | High risk |
|---|---|
| Using AI to refine a first draft you already wrote | Starting with AI instead of thinking first |
| Getting suggestions for improvement | Replacing your own thinking or drafting with full AI-generated content |
| Offloading tedious, well-defined tasks | Over-relying on AI for decisions in unfamiliar domains or tasks |
A simple three-step structure for any single AI interaction, low-effort enough to actually use every time.
Generate the prompt yourself, with real detail and actual thought behind it.
Let the model generate its answer.
Judge the response for both content andtone before you use any of it.
There's no lab value for this. The monitoring here is a habit you run on yourself, every time you open an AI tool.
| Check | When | Watching for |
|---|---|---|
| Metacognitive pass(plan, monitor, evaluate, reflect) | Every time you use AI on a task | Whether you actually understood the output or just accepted it |
| Task difficulty | Before reaching for AI | Whether you're skipping productive struggle you actually need for the "absence of difficulty" reason |
| Domain risk level | Before starting any AI-assisted task | Novice + unfamiliar domain = don't start with AI, start with reflection |
| PHI content | Every single prompt, no exceptions | Any patient-identifiable information present |
| Attribution | Every assignment or paper | Whether AI use is disclosed per your program's policy |