The word "AI" is now applied to nearly all parts of life and science, including nucleic-acid classifiers, chatbots, robotic surgical arms. The challenge is that ‘AI’ is applied as though a single set of strengths, weaknesses, and risks applied to all three. This mirrors a polysemy (when a word has multiple related meanings) clinicians already navigate: "medicine" denotes the entire discipline of healthcare and, within it, internal medicine (abbreviated as ‘medicine’) as one specialty among surgery, radiology, and pathology. A critique that legitimately targets one branch does not automatically transfer to the whole. Most disputes about AI in clinical settings — "AI hallucinates," "AI is a black box," "AI is biased" — are branch-specific claims wrongly addressed to the umbrella term, and so a short checklist for evaluating any AI claim encountered in practice, review, or product development seems appropriate.
1. A familiar confusion in an unfamiliar register
A radiologist reviewing an FDA-cleared nodule-detection algorithm and a hospitalist drafting a note with a chatbot's help are both said to be "using AI." When the chatbot fabricates a citation, the resulting distrust often generalizes to the nodule detector, and vice versa even if the detector performs well. This transfer is not justified. The two systems differ in architecture, training objective, and failure mode as much as an SSRI differs from a laparoscopic Nissen fundoplication — both are "medicine," but neither is informative about the other's risk profile.
2. The precedent: "medicine" already means two things
Clinicians hold this distinction automatically. "Medicine" names the entire enterprise of healthcare, and, nested inside it, internal medicine as a discipline distinguished from surgery, radiology, pathology, and psychiatry. A postoperative anesthesia complication is properly attributed to a procedural specialty, not to "medicine" as such; conversely, hand hygiene failures and diagnostic anchoring bias are genuinely enterprise-wide problems. Board certification, scope-of-practice statutes, and malpractice actuarial categories keep this distinction operational without anyone needing to argue for it. AI has no equivalent institutional scaffolding yet — which is precisely why explicit scoping has to be done by hand.
3. A taxonomy of "AI," medically framed
"Artificial intelligence" nests several technologies of quite different character:
| Level | Examples | Representative clinical use |
|---|---|---|
| Symbolic / rule-based systems | Expert systems, clinical decision rules | Drug-interaction alerts, sepsis rule-outs |
| Classical statistical machine learning | Logistic regression, gradient-boosted trees | Readmission risk scores, mortality indices |
| Deep learning — discriminative | Convolutional networks, encoders | Nodule detection, ECG interpretation, histopathology classification |
| Deep learning — generative | Diffusion models, autoregressive transformers | Accelerated MRI reconstruction, ambient documentation, large language models (LLMs) |
The term "generative AI" is often used as a loose synonym for "LLM," but the correct criterion is output type, not architecture: a generative model is one that emits a high-dimensional product for human consumption — text, image, audio — rather than a label or a score. This is a different sense of "generative" than the classical statistical one (a model that represents a joint distribution P(x,y), such as naive Bayes), which is why the two usages are easily and consequentially conflated. Under the correct, output-based definition, an LLM drafting a note, a diffusion model reconstructing an MRI from undersampled k-space, and an ambient-scribe system are all generative AI; a nodule-detection CNN, whose output is a bounding box and a probability, is deep learning but not generative AI. LLMs are properly understood as one branch of generative AI — autoregressive transformers operating on text — not as a category alongside deep learning; they are, architecturally, a further specialization within it.1
4. Scoped failure modes: what actually transfers, and to what
Confabulation. So-called hallucination in LLMs arises from sampling plausible continuations in regions where the model is underdetermined by its training or context.1 This is a property of unverified generative output, not of "AI" broadly, and not even of deep learning broadly — a discriminative CNN classifying a chest radiograph cannot "hallucinate" a diagnosis in this sense; it can only misclassify within a fixed label space, a different and separately characterizable error. The same generative failure mode, notably, recurs outside language: deep-learning MRI reconstruction from undersampled k-space data can produce anatomically plausible structures that were not present in the acquired signal — a phenomenon documented in the reconstruction literature under the same name, hallucination, because the underlying mechanism is the same.2 The lesson for tool evaluation is that the presence or absence of a verification step against ground truth (a compiler, a lab value, a fully sampled reference scan) predicts error behavior more reliably than the architecture label does. Code or structured-data generation checked against a validator has a materially different error profile than open-ended clinical narrative generated by the same underlying model.
Opacity. "Black box" is a claim whose truth value depends entirely on model class. A twelve-variable logistic regression is fully auditable; a large transformer is not, in the same sense, though interpretability techniqus narrow the gap. Regulatory frameworks already track this variation rather than treating "AI" as monolithic: the FDA's approach to AI/ML-based software as a medical device is explicitly built around model-specific characteristics — the anticipated types of modification a given system can make and how those changes are validated — rather than a single opacity standard applied across all AI-enabled devices.3
Bias. This is the deliberately included counterexample to the main point that AI has many types--the exception that proves the rule. Algorithmic bias in healthcare derives chiefly from training-data provenance and proxy-outcome selection, not from any single architecture, and it therefore generalizes widely across branches. The best-documented case is a commercial population-health algorithm that used healthcare cost as a proxy for healthcare need; because less money is historically spent on Black patients at a given level of illness, the algorithm systematically underestimated their need, reducing the fraction identified for extra care.4 This was a simple regression-based risk model, not a neural network — direct evidence that bias tracks data and objective design across the entire taxonomy, and that "our model is not an LLM" is no defense against it.
5. What the claim about LLMs in medicine actually supports — and doesn't
Evaluations of medical-domain LLMs, such as Med-PaLM's performance on licensing-exam-style questions, show that factual recall and reasoning on structured medical questions improve with model scale and instruction tuning.5 That is evidence about closed-form question answering under evaluation conditions, not evidence about open-ended clinical documentation, differential diagnosis under ambiguity, or unsupervised patient-facing use — a distinction the original evaluation itself is explicit about. The same discipline of scoping that limits over-broad critiques must equally limit over-broad extrapolation from benchmark performance to clinical deployment.
6. A scoping checklist
Before accepting or repeating a claim about "AI" — in a manuscript, an IRB protocol, a vendor pitch, or a curbside conversation — five questions localize it:
- Model class: symbolic, classical statistical, discriminative deep learning, or generative deep learning?
- Output type: a label/score, or a human-consumable artifact?
- Training-data provenance: curated and governed, or broadly scraped?
- Verification step: is output checked against ground truth (a compiler, a reference standard, a fully sampled scan) before it reaches a decision, or not?
- Deployment context: decision support with a human in the loop, or autonomous action?
A claim that cannot be located on these axes is not yet precise enough to evaluate — the AI equivalent of a case report that never specifies which organ system was involved.
7. Conclusion
The nodule detector and the documentation chatbot are both, correctly, called AI — exactly as a laparoscopic cholecystectomy and a course of metformin are both, correctly, called medicine. Precision about which branch a critique or a benchmark result actually targets is not pedantry; it is the same discipline that already governs how the field talks about medicine itself, and it is a precondition for evaluating any AI tool — including the generative ones now entering clinical workflows — on its actual merits rather than its category's reputation.
References
- Ji Z, Lee N, Frieske R, et al. Survey of hallucination in natural language generation. ACM Comput Surv. 2023;55(12):Article 248.
- Bhadra S, Kelkar VA, Brooks FJ, Anastasio MA. On hallucinations in tomographic image reconstruction. IEEE Trans Med Imaging. 2021;40(11):3249-3260.
- U.S. Food and Drug Administration. Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. Center for Devices and Radiological Health; January 2021.
- Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453.
- Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180.