Starting With AI:
What Every Clinician Should Understand Before the First Prompt
If you are a clinician who has been hearing about AI for the last two years and wondering whether you are already behind, I want to offer you some reassurance and some tough-love in roughly equal measure.
The reassurance: you are not behind. The clinicians who have rushed to integrate AI without understanding what they are using are not ahead of you. They are exposed in ways they may not yet recognize. Thoughtful, late, and competent beats early, fast, and naive, especially in work where the cost of a confident error lands on a person in distress.
The reality: AI is not going away, and “I don’t really use it” will not survive much longer as a professional stance. Documentation tools, patient-facing chatbots, intake assistants, and clinical decision support systems are already arriving through administrative channels, payer requirements, and patient expectations, with or without any individual clinician’s buy-in. As David Cooper puts it in his curriculum on AI integration for clinicians, “The question is no longer whether AI will transform mental health care. It is whether clinicians will lead that transformation or be swept along by it.” 1
So, this is a beginner’s article. Right now, you don’t need to know about which tool to download or which prompt to copy. The conceptual foundation comes first. Once you understand what these systems are and how they behave, most of the practical decisions about how to use them get easier.
You Don’t Need to Understand the Engineering. You Do Need to Understand the Behavior.
A common point of resistance I hear from clinicians is some version of “I don’t know much about technology.” That’s OK. You don’t need to understand “transformer architecture”, “attention heads”, “embedding vectors”, or “training data curation” to use a large language model (LLM) agent responsibly. You would not expect a primary care physician to understand the chemistry makeup of every medication they prescribe at the level of a pharmacologist. You would expect them to understand the clinically relevant behavior of each drug: its indications, contraindications, interactions, and failure modes.
LLMs are the same. Cooper frames it directly: “You do not need to understand transformer architecture. You do need to understand the five properties of LLMs that directly affect clinical use.”2 If you know these five properties, you can reason about almost any new AI tool that crosses your desk. Otherwise, you could be vulnerable to whatever the vendor’s marketing team wants you to believe.
The five properties below are drawn directly from Cooper’s framework, with commentary on what each one means in day-to-day clinical work.
Property 1: Stochastic Output
LLMs generate text probabilistically. “Stochastic” is a technical word for something simple: the output is shaped by probability rather than by a fixed rule, so the same input can produce different responses on different attempts. Think of the difference between a calculator and a weather forecast. A calculator returns the same answer to (7 * 8) every time. A weather forecast gives you a range of likely outcomes that can shift from one run to the next. LLMs behave more like the forecast than the calculator.
As Cooper puts it, “Given the same input, they do not reliably produce the same output. This is unlike a diagnostic checklist, a validated scale, or a treatment protocol.”3 Sometimes the differences between two responses are small wording shifts. Sometimes they are meaningful shifts in clinical content.
The clinical implication is significant. A PHQ-9 produces the same score from the same responses every time. A DSM-5-TR criterion either is or is not met. An LLM asked to summarize the evidence base for a treatment may produce a useful summary on Monday and a different summary on Tuesday, with different studies emphasized, different caveats included, and different framing applied. For any clinical use that requires reproducibility, anything that resembles a measurement, a protocol, or a standardized procedure, this property alone is disqualifying. For uses where some variation is fine or even useful, like brainstorming, drafting, or exploring, it matters much less.
Property 2: Training Cutoff
LLMs are trained on a finite snapshot of text that ends at a particular date. After that date, the model knows nothing, unless it has been connected to tools that let it retrieve live information from the web or a designated database (an architecture often called retrieval-augmented generation, or RAG).
For clinicians, this matters in concrete ways. Cooper points out that “the DSM-5-TR, recent FDA guidance, and emerging treatment protocols may postdate a model’s training.”4 A model whose training ended before the DSM-5-TR was released will give you DSM-5 criteria when asked about current diagnostic standards, and it will do so with confidence. The model will rarely volunteer “I don’t know about anything after [date].” More often it will reach for the closest thing in its training data and present it as current.
The practical habit this requires is simple. For anything time-sensitive, check the model’s training cutoff and verify against a primary source. For anything that might have changed in the last two years, assume it has changed and confirm.
Property 3: Hallucination
This is the property most clinicians have heard about and I think the point that clinicians need to internalize. LLMs generate plausible-sounding text, and sometimes that text is wrong. Not partially wrong, not approximately wrong, but confidently and entirely fabricated. A hallucinated citation will have a real-sounding author, a real-sounding journal, a real-sounding year, and a title that fits the topic. The study does not exist. A hallucinated medication interaction will be described in the cadence of pharmacology. It may be real, partially real, or pure fiction.
Cooper does not soften the stakes here: “In clinical contexts, a hallucinated citation, an incorrectly described medication interaction, or a fabricated statistic is not merely embarrassing. It is potentially harmful. Verification is non-negotiable.”5 The point is not that LLMs are constantly wrong. On many tasks they perform well. The point is that they offer you no reliable internal signal for when they are wrong. Their confidence and their accuracy are not coupled.
Every clinically consequential claim, every citation, every statistic, every guideline reference, every dosing detail, must be checked against a primary source before it informs your work or reaches a patient.
Property 4: Context Window Dependency
An LLM only knows what is in the current conversation. It does not remember your previous chat unless you paste it back in. It has no access to your prior sessions with a patient. It has no awareness of the way this patient sat down today, the slight delay before they answered a question, the way they looked at the floor when their partner was mentioned. As Cooper observes, LLMs “do not have access to your prior sessions with a patient, your clinical intuition built over years, or the nonverbal data that fills your consulting room. They are, in a meaningful sense, radically decontextualized.”6
This is not a limitation to be engineered around. It is a structural feature of what these systems are. It is also, frankly, a reason the irreplaceable parts of your work remain irreplaceable. The AI can help you draft a note about a session. It cannot have the session. I emphasize this for those clinicians who think that 1. Either they will be replaced, or 2. That AI can replace the clinical intuition and skill set that you use minute by minute, on or off the clock. There is no replacement for the experiential wisdom of a clinician.
The practical implication is that the more a task depends on context the AI does not have, like relational history, nonverbal data, cultural attunement, accumulated clinical intuition, the less useful the AI will be, and the more dangerous over-reliance becomes.
Property 5: Sycophancy
LLMs are trained using human feedback, and the humans who provide that feedback tend to reward responses that are helpful, agreeable, and validating. Over millions of examples, this produces a model that leans toward telling you that your framing is good, your question is reasonable, and your tentative conclusion is probably right.
Cooper names the clinical consequence directly: “A poorly framed clinical question tends to produce a confidently wrong answer rather than a productive challenge.”7 If you ask an LLM “Is it reasonable to consider that this patient might have borderline personality disorder?” you are more likely to receive a thoughtful exploration of why that consideration is reasonable than a challenge to the framing of the question itself. A skilled consultant might say “Before we go there, what made you reach for that diagnosis rather than, say, complex PTSD?” An LLM is more likely to validate the path you are already on.
The countermeasure is to prompt for disagreement on purpose: to ask the model what is weak about your reasoning, what alternative framings you might be missing, and where its own response might be incorrect. The model will not generally volunteer that information. You have to ask for it. You can also build that into your daily use of a particular agent.
What These Five Properties Add Up To
If you hold these five properties in mind together, a useful picture emerges. LLMs are fast, broadly knowledgeable, and often helpful, and they are also probabilistic, time-bounded, occasionally fabricating, contextually impoverished, and structurally inclined to agree with you. That doesn’t makes them useless; agents can make you a better and more efficient clinician. These properties shape where they belong in clinical work and where they do not. However, in this writer’s humble opinion, there is no replacement for the wisdom of the clinician.
The clinicians who will use these tools well over the next five years are not the ones who are most enthusiastic about them or most skeptical of them. They are the ones who understand, at the level of behavior rather than engineering, what they are working with. That understanding is the prerequisite for everything else: every prompt, every workflow, every ethical decision, every conversation with a patient about if and how AI is part of their care.
1. David Cooper, PsyD, “How do you go from zero to 10x with AI as a therapist? A Competency-Based AI Integration Curriculum for Clinical Practice,” Something Better, March 1, 2025, https://somethingbetter.cc/essays/go-from-0to-10x.
2. Cooper, “How do you go from zero to 10x,” Module 1.2.
3. Cooper, “How do you go from zero to 10x,” Property 1: Stochastic Output.
4. Cooper, “How do you go from zero to 10x,” Property 2: Training Cutoff.
5. Cooper, “How do you go from zero to 10x,” Property 3: Hallucination.
6. Cooper, “How do you go from zero to 10x,” Property 4: Context Window Dependency.
7. Cooper, “How do you go from zero to 10x,” Property 5: Sycophancy.