← Master Index· Foundations & Skills

AI and Pharmacy Informatics

Foundations & Skillslarge language modelsprompt engineeringcognitive offloading

30-Second Snapshot

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.

Worth knowing

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.

Key Terms - The Nesting Doll

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.

The four tiers of AI capability

TypeWhat it meansExamples
Weak IntelligenceBasic systems that perform one specific task with no real understandingSpell check, email filters, simple chatbots, recommendation systems
Artificial Narrow Intelligence (ANI)AI designed to do a narrow task wellSiri, Alexa, Google Search, facial recognition, ChatGPT
Artificial General Intelligence (AGI)Would understand, learn, and apply knowledge across a wide range of tasks, like a humanHypothetical, doesn't exist yet
Artificial Super Intelligence (ASI)Would surpass human intelligence in creativity, problem-solving, and emotional intelligenceHypothetical, science fiction territory
The trap

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.

The nesting, defined

How LLMs Actually Work - and Why That Explains Everything Else

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.

Attention is all you need

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.

The unlock

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.

Model drift

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.

Fluency is not accuracy

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.

AI Tools You'll Actually See

Vocabulary you'll hit in the wild, split into who's building it and how you'd actually use it.

Commercial LLMs
ModelMaker
Gemini (formerly Bard)Google
ChatGPTOpenAI
ClaudeAnthropic
Command RCohere
JurassicAI21
Open-source LLMs
ModelOrigin
AlpacaStanford
DeepSeekHigh-Flyer (Hangzhou)
DollyMIT
GemmaGoogle
LLaMAMeta

Two tools worth knowing by name

Why NotebookLM matters more than it sounds

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.

Prompt Engineering, RAG, and Fine-Tuning

Three distinct levers exist for improving what an LLM gives you back, and they operate at completely different levels.

LeverWhat it changesWho does it
Prompt engineeringYour 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 respondsBuilt into tools like NotebookLM; you supply the source
Fine-tuningRetrains the model itself on additional data, changing its baseline behavior everywhereDone by developers, not end users
Which lever to pull

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 habits that make a prompt actually work

  • Be specific."Summarize this" gets a generic summary. "Summarize this in 3 bullets focused on dosing changes" gets a usable one.
  • Provide context.Tell it who the audience is, what you already know, and what you're trying to do with the answer.
  • Mention preferences.Tone, length, format, reading level, all of it steers the output.
  • Keep it relevant.Strip irrelevant detail out of the prompt the same way you'd strip it out of a SOAP note.
  • Request adjustments.Treat the first response as a draft, not a final answer. Ask it to revise rather than starting over.

Warnings, Risks, and Ethics

Five categories of failure to hold in mind every time you open an AI tool.

RiskWhat it looks like
AccuracyConfident, fluent, factually wrong output. The default failure mode, not the exception.
Sensitive informationData you type into a public tool can be stored, logged, or used to train future models. This is a hard HIPAA problem.
Bias and fairnessModels trained on real-world data inherit real-world bias, and can reproduce it in clinical framing, risk language, or recommendations.
Emotional intelligenceThe model has no actual empathy or judgment about a patient's emotional state, it's simulating the language of empathy.
Ethical useCovers everything from plagiarism to undisclosed AI authorship to using AI in place of your own clinical judgment.
The non-negotiable

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.

Values Relevant to AI

A checklist of angles to evaluate any AI tool or use case against, not just "does it work."

ValueAsks
EfficiencyDoes it actually save time and effort, net of the time spent verifying it?
ValidityDoes it measure or answer the thing it claims to?
ReliabilityDoes it give consistent output for the same input over time?
ViabilityIs it actually sustainable to use in a real workflow, not just a demo?
GovernabilityCan it be controlled, audited, and corrected when it's wrong?
TransparencyIs it clear how the output was generated and what it was trained on?
SustainabilityEvery AI prompt consumes real energy. This is now a measured, publicly reported cost, not a hypothetical one.
EquityDoes access to and benefit from the tool distribute fairly, or does it widen an existing gap?

Cognitive Offloading and the Learning Cost

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.

Key terms, precisely

TermDefinition
Cognitive loadThe amount of mental effort a task takes, factoring in the setting and what makes it harder or easier
Cognitive offloadingExternalizing a cognitive task to an outside tool to reduce that load
Cognitive debtAccumulated loss of higher-order thinking ability from repeated offloading, especially with AI tools
MetacognitionThe ability to identify, evaluate, and regulate your own thinking process

The trade, in one table

Benefits of offloadingRisks of over-relying on it
Frees mental resources for more complex, dynamic workCognitive laziness
Shifts focus from memorizing to knowing where to find it (the "Google effect")Cognitive dependence
Improves efficiency and performance in the momentLoss of cognitive autonomy, reduced critical engagement and accountability, decline in cognitive function and memory retention
The numbers behind the warning

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."

Why the brain needs the resistance

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.

Using AI as a Learner

Not all AI-assisted work is created equal. The structure of the assignment determines how much offloading risk it carries.

MODEL 1

Human-Only

No AI assistance at all. Maximum skill-building, zero efficiency gain.

MODEL 2

AI-Only

Task completed entirely by the AI. Maximum offloading risk, minimal learning.

MODEL 3

Human + AI, Unguided

AI used without any structured instruction on how. The default, and the riskiest hybrid.

MODEL 4

Human + AI, Guided

Trained specifically to avoid offloading and stimulate critical thinking. The target.

The guided-AI sequence (minimum offloading)

  1. Initial reflection:work out what you actually think and how you'd say it, before AI touches the task.
  2. Search:use AI strictly to pull facts, not to think for you.
  3. Creation:take those facts and build the document yourself.
  4. Critical review:submit your own draft to AI and ask it to revise, flag missing information, or propose counterarguments.
  5. Final revision:refine your argument using what that critique surfaced.
The unlock

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.

Expert vs. Novice Risk

The same AI tool carries very different risk depending on who's holding it.

ExpertsNovices
Can they catch AI errors?Easily spot duplicated ideas, irrelevant information, and wrong informationLack the skills to recognize inconsistencies, especially early in the learning curve
Overall risk of using AI hereLow-risk integrationHigh-risk usage
Why this matters for you specifically

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 vs. high risk use, concretely

Low riskHigh risk
Using AI to refine a first draft you already wroteStarting with AI instead of thinking first
Getting suggestions for improvementReplacing your own thinking or drafting with full AI-generated content
Offloading tedious, well-defined tasksOver-relying on AI for decisions in unfamiliar domains or tasks

The Sandwich Method

A simple three-step structure for any single AI interaction, low-effort enough to actually use every time.

1

Human prompt

Generate the prompt yourself, with real detail and actual thought behind it.

2

AI response

Let the model generate its answer.

3

Human evaluation

Judge the response for both content andtone before you use any of it.

Self-Check - What, When, Why

There's no lab value for this. The monitoring here is a habit you run on yourself, every time you open an AI tool.

CheckWhenWatching for
Metacognitive pass(plan, monitor, evaluate, reflect)Every time you use AI on a taskWhether you actually understood the output or just accepted it
Task difficultyBefore reaching for AIWhether you're skipping productive struggle you actually need for the "absence of difficulty" reason
Domain risk levelBefore starting any AI-assisted taskNovice + unfamiliar domain = don't start with AI, start with reflection
PHI contentEvery single prompt, no exceptionsAny patient-identifiable information present
AttributionEvery assignment or paperWhether AI use is disclosed per your program's policy

Practical Rules - What You Actually Do

  • Know the policy before you start."Am I even allowed to use AI on this, and if so, how much?" Check your school's or employer's stated AI policy, don't assume.
  • Never type in patient-identifiable information.Not a name, not a MRN, not enough combined detail that someone could identify the patient. Full stop, no public AI tool is a substitute for a vetted, secure system.
  • Reflect before you generate."What do I actually think here, before I ask anything?" Do that first, every time, even for something small.
  • Disclose your use.Reference AI the way you'd reference any other source you drew on, per whatever citation standard applies.
  • Never submit AI output as your own original work.That's true for a class assignment and it's true for a clinical note.
  • Run the Sandwich Method.Your prompt, its response, your evaluation of both content and tone, before anything goes further.
  • Match your use to your risk level.If you're a novice in the domain, use AI to refine your own draft or pressure-test your own reasoning, not to generate the first version.

High-Yield Recall Sheet

  • The nesting doll:AI ⊃ Machine Learning ⊃ Generative AI ⊃ LLM ⊃ GPT. GPT names one specific architecture, not "any LLM."
  • 4 tiers of AI capability:Weak Intelligence, ANI, AGI, ASI. Only Weak and ANI exist in real use today; AGI and ASI are hypothetical.
  • ChatGPT/Claude/Gemini are ANI, narrow systems that do one thing (predict text) well, not general intelligence.
  • Transformers use attention:they weigh the whole input at once instead of one word at a time, stacked in layers, which is why they hold long context.
  • LLMs predict the statistically likely next token,they don't reason about truth. Fluent output is not the same as accurate output.
  • Model driftmeans performance degrades as the world moves past the training snapshot, requiring recalibration.
  • 3 levers to improve output:prompt engineering (you), RAG (grounds it in your documents), fine-tuning (retrains the model, developers only).
  • 5 prompt habits:be specific, give context, state preferences, keep it relevant, ask for adjustments.
  • NotebookLM only answers from documents you upload,structurally lower hallucination risk than an open chat.
  • 5 warning categories:accuracy, sensitive information, bias/fairness, emotional intelligence, ethical use.
  • 8 AI values to evaluate a tool against:efficiency, validity, reliability, viability, governability, transparency, sustainability, equity.
  • Cognitive offloading correlations:r=+0.72 reliance-to-offloading, r=-0.75 offloading-to-critical-thinking, r=-0.68 reliance-to-critical-thinking-decline.
  • "The absence of difficulty leads to cognitive debt."Neuroplasticity requires resistance, same as a muscle under load.
  • Metacognition (plan, monitor, evaluate, reflect) mediates offloading, it's what makes delegation smart instead of corrosive.
  • Guided-AI order matters:Reflect, Search, Create, Critical Review, Final Revision. Reflection is always first, AI-as-critic is always last.
  • Experts catch AI errors easily; novices can't.Every new APPE rotation makes you a novice again, which makes AI use higher-risk exactly then.
  • Sandwich Method:your prompt → AI response → your evaluation of content and tone.
  • Never put HIPAA information into AI. Never submit AI work as your own.Those two rules don't have exceptions.