Three Kinds of Clinical AI
Distinguish predictive models, regulated device AI, and generative AI in healthcare, and see how the 4D fluency framework applies to each.
“AI” Means Several Different Things in Healthcare
When a colleague says “we’re rolling out AI,” they could mean a deterioration score in the EHR, an FDA-cleared algorithm that flags intracranial hemorrhage on CT, or a chatbot that drafts replies to patient portal messages. These tools fail in different ways, carry different regulatory status, and demand different kinds of oversight. Fluency starts with telling them apart.
| Predictive models | Device AI (SaMD) | Generative AI | |
|---|---|---|---|
| Examples | Sepsis and deterioration scores, readmission risk, no-show prediction | Imaging triage and detection, ECG interpretation, diabetic retinopathy screening | Ambient scribes, inbox reply drafts, chart summarization, general chat assistants |
| Output | A score or alert | A finding, flag, or measurement | Free text (notes, letters, answers) |
| Built from | Historical EHR data, often at one institution | Curated, labeled datasets for a narrow task | Very large general text corpora, sometimes tuned on clinical text |
| Typical oversight | Local governance committee | FDA clearance or approval (US), CE marking (EU) | Varies widely; often none specific to the use case |
| Classic failure | Performs worse at a new site or over time | Narrow: misses what it wasn’t trained on | Fluent, plausible, and wrong |
Predictive models
These estimate the probability of an outcome from structured data: vitals, labs, demographics, utilization. They’ve been in EHRs for years. Their key weakness is dataset shift: a model trained on one hospital’s patients, documentation habits, and coding practices may perform very differently at another hospital, or at the same hospital after a workflow change.
Device AI
Software intended to diagnose, treat, or prevent disease can be regulated as a medical device: Software as a Medical Device (SaMD). The US FDA’s public list of AI-enabled devices passed one thousand entries in 2025, with radiology accounting for roughly three quarters. Clearance means the device met a regulatory bar for a specific intended use and population. It doesn’t mean it will work equally well for every patient in your clinic.
Generative AI
Large language models (LLMs) produce text by predicting likely next words. They power ambient documentation, inbox drafting, chart summarization, and general-purpose assistants clinicians use on their own. They’re remarkably capable with language and can perform well on licensing-exam questions, but they have no built-in mechanism to know when they’re wrong. Most of this course focuses here, because generative AI is where individual clinicians make the most day-to-day judgment calls.
Common mistake: assuming a general chatbot carries the same evidence base as an FDA-cleared imaging algorithm, or that either is validated for your patients.
Ask of any tool: What exactly is it intended to do? On whom was it tested? Who monitors it here? A general-purpose assistant answering a dosing question has no clinical validation for that use at all.
A Framework for Using Any of Them: The 4Ds
This course applies the 4D AI Fluency Framework to clinical work. Each competency answers a question you face whenever AI touches patient care:
- 1 Delegation
Should AI do this task, and which part?
- 2 Description
How do I specify the task precisely?
- 3 Discernment
Is the output accurate, complete, and safe?
- 4 Diligence
Is my use ethical, compliant, and accountable?
| Competency | In the clinic, it sounds like… |
|---|---|
| Delegation | ”An AI draft of this patient education handout is fine. An AI deciding whether this chest pain needs a troponin is not.” |
| Description | ”I gave the model the patient’s age, renal function, and current medications, and asked it to flag uncertainty.” |
| Discernment | ”The scribe note documented a normal abdominal exam I never performed. I removed it.” |
| Diligence | ”I used our BAA-covered tool, not a consumer app, and I signed the note knowing I own every word.” |
The modules follow this structure: the landscape and failure modes first, then one or more modules per competency.
- List every AI-powered tool you encounter in a typical week: EHR alerts and scores, imaging flags, scribe or dictation tools, inbox drafts, and anything you use on your own (including a phone app).
- Classify each as predictive, device, or generative.
- For each, write whether you know (a) who approved it for use in your organization and (b) whether it’s been validated on a population like yours.
- Circle any tool where you can’t answer either question. You’ll revisit this list in Module 6.
1. An EHR model that predicts which inpatients will deteriorate in the next 12 hours is best described as:
It produces a risk score from structured data, which is the hallmark of a predictive model. Its main vulnerability is dataset shift.
2. What does FDA clearance of an AI imaging device tell you?
Clearance is tied to a defined intended use. Local performance and fit for your population still need attention.
3. Which characteristic of generative AI matters most for clinical safety?
LLMs can be confidently and plausibly wrong, which is why discernment is central to using them safely.
About the Framework
The 4D AI Fluency Framework was created by Rick Dakan and Joseph Feller with Anthropic and is released under CC BY-NC-SA 4.0. This course’s clinical applications, examples, and exercises are original. Learn more at aifluencyframework.org.
Up Next
Before deciding what to delegate, you need a clear picture of how these tools fail. Next: the failure modes every clinician should recognize.