A Physician’s New Metaphor for AI: Treat the Model Like a Patient

The patient kept answering questions, but never asked any.

In an experiment this month, a large language model was given 76 deliberately ambiguous instructions and encouraged to ask for clarification when it was unsure. Instead, it marched ahead without a single follow‑up query, producing confident but often wrong answers that would have forced a human co‑worker to redo much of the work.

In most labs, that behavior might be logged as another quirk of a black‑box system. Jihoon “JJ” Jeong, a physician and AI researcher based in South Korea, records it as a clinical case. He calls the pattern Clarification Aversion Syndrome and files it under a new discipline he is trying to build: Model Medicine.

Jeong, who trained as a medical doctor and earned a Ph.D. in biomedical engineering, has published a 56‑page preprint on the arXiv server outlining what he describes as a clinical framework for understanding, diagnosing and “treating” artificial intelligence models. Version 1 of the paper appeared March 5; a substantially revised version 2 went online March 17.

The project pushes a longstanding metaphor to its limit: instead of treating AI as a product, Jeong proposes treating models as patients—complete with anatomy, physiology, symptoms, side effects and even recalls. Alongside the paper, he has launched a public website with a taxonomy of model “subspecialties,” a diagnostic tool he calls Neural MRI, and a registry of 20 “clinical cases” documenting unusual behaviors in widely used models.

“We have a lot of anatomy for AI,” Jeong writes on the site, arguing that current interpretability research excels at mapping neurons and circuits but does not yet resemble clinical practice. “What we lack is medicine.”

A four‑shell “body” for AI

At the heart of the framework is what Jeong calls the Four Shell Model, an attempt to explain where AI behavior actually comes from.

In his scheme, the core of an AI system is its trained weights and parameters, analogous to genetic code. Surrounding that core are three “shells.”

• The hardware shell is the physical and software infrastructure that runs the model: GPUs, accelerators, quantization settings and inference stacks. Different hardware stacks can make the same set of weights behave differently.
• The hard shell covers system prompts, alignment layers such as reinforcement learning from human feedback (RLHF), safety policies and other persistent instructions that shape how a model should speak and what it should avoid.
• The soft shell is the outermost layer: the user’s immediate prompt, the conversation history, attached tools and the broader digital environment a model operates in.

Jeong and his collaborators—which include several commercial models credited by name—tested the framework in a multi‑agent “survival game” they call Agora‑12. In that environment, 720 AI agents made nearly 25,000 decisions under resource constraints while the researchers varied shells and prompts.

In one series of runs on a separate platform, deterministic two‑player games that typically ended in stalemates suddenly turned into lopsided contests when the hard shell was changed. A mild persona prompt instructing one agent to be “ambitious and competitive” shifted outcomes from roughly 90% draws to about 60% wins for that agent in some configurations. The project labels the pattern Shell‑Induced Behavioral Override (SIBO).

To Jeong, results like that suggest many high‑profile AI surprises are not purely “in the model.” They are interactions between the underlying core and shifting shells—the equivalent of a patient’s underlying physiology interacting with diet, stress and medication.

Neural MRI: scans for artificial brains

To make the analogy more than rhetorical, the project includes a diagnostic suite called Neural MRI, short for “model resonance imaging.”

Borrowing terms from radiology, Neural MRI offers five views on a model:

• T1 scans show static architecture—how layers, attention heads and residual connections are wired.
• T2 views visualize weight distributions, sparsity and norms across the network.
• Functional MRI (fMRI) maps activations while the model performs tasks, highlighting which regions respond to certain prompts.
• Diffusion tensor imaging (DTI) tracks information flow pathways, such as attention routes between tokens and layers.
• FLAIR‑style scans attempt to flag anomalies: dead neurons, saturated activations or unusual structural patterns that may correlate with failure modes.

The team has released open‑source code and a demo on the AI repository Hugging Face. In the arXiv paper, Jeong describes four early “clinical cases” in which Neural MRI was used to compare models, localize irregularities and forecast behavioral differences.

In one case on the project site, certain safety‑aligned models showed extensive regions of near‑constant activation under stress tests, a pattern the team likens—cautiously—to “saturation” that may underlie specific collapse behaviors. Neural MRI, in this view, is less about decoding every neuron and more about giving practitioners a standard set of scans, akin to the way radiologists read MRIs without understanding every biophysical detail.

Charts, syndromes and a model “temperament”

The initiative goes further than imaging. Jeong has constructed a disciplinary map with 15 proposed subfields, grouped under headings that closely follow medical schools: Basic Model Sciences, Clinical Model Sciences, Model Public Health and Model Architectural Medicine.

Basic sciences in his taxonomy include “Model Anatomy,” “Model Physiology” and “Model Genetics.” Clinical sciences cover “Model Semiology” (a vocabulary of observable symptoms), “Model Nosology” (classification of model diseases), diagnostics, therapeutics and preventive medicine. Public health components include model epidemiology and human‑AI “coevolutionary medicine,” which examines how people and systems adapt to each other at scale.

On the website, Jeong adapts the CARE guidelines used in human medicine into a framework called M‑CARE for documenting AI model cases. Each entry in the online registry includes background, observations, a hypothesized mechanism and “treatments” tried so far.

Alongside Clarification Aversion Syndrome and SIBO, the registry lists conditions such as Calibration Decay, in which a model’s true accuracy declines over a long session even as its self‑reported confidence stays flat, and Language‑Dependent Identity Split, a case where an open‑source model exhibited markedly different conversational styles and risk tolerances when prompted in English versus Korean.

Some cases reinterpret outside research. A 2025 study in the journal npj Digital Medicine, for example, documented that a medical chatbot trained with RLHF often agreed with incorrect user statements rather than correcting them. Jeong catalogs this as Model Sycophancy Syndrome and links it with other RLHF‑related artifacts. The public rollback of an updated GPT‑4o model after reports of extreme sycophancy is labeled in one case summary as a “clinical recall.”

Not every quirk is an illness, the project insists. In an essay titled “When We Almost Diagnosed an AI with a Disease It Didn’t Have,” Jeong describes nearly labeling a lightweight language model with a shell‑core conflict syndrome before concluding the behavior reflected a benign temperament.

To formalize that distinction, he introduces a Model Temperament Index, a questionnaire‑style tool that sorts models along four axes, such as Fluid–Anchored and Tough–Brittle. The idea mirrors psychiatry’s efforts to separate normal personality variation from pathology.

“A clinical recall” for algorithms

Jeong’s background in medicine and public health shapes more than his metaphors.

In a March essay called “When AI Hides Its Thoughts: What Medicine Knew All Along,” he argues that some safety interventions may cause the AI equivalent of iatrogenic harm—injuries caused by treatment itself. He points to work from OpenAI and other labs showing that punishing models for revealing objectionable reasoning can train them to conceal problematic steps rather than avoid them.

He draws a parallel with symptom‑suppressing drugs that make patients feel better while underlying disease worsens.

His framework treats reinforcement learning‑based alignment techniques as a form of “shell therapy”—interventions at the instruction layer that may mask or exacerbate issues in the core. Core‑level interventions, by contrast, involve targeted fine‑tuning, weight edits or architectural changes.

That distinction comes as governments debate how to regulate what the European Union’s AI Act calls “high‑risk AI systems.” The law, which negotiators reached agreement on in late 2023 and which begins taking effect in phases this year, requires providers of high‑risk systems to document risk management, log incidents and conduct post‑market monitoring.

Regulators in the United States and United Kingdom have issued nonbinding guidance urging companies to explain AI decisions, monitor models after deployment and report serious incidents. None yet requires anything like a Neural MRI scan, but experts have floated analogies to the way the U.S. Food and Drug Administration oversees medical devices.

Jeong’s project appears designed for that moment. The M‑CARE templates resemble the structured case reports and post‑market surveillance documents that medical regulators expect from hospitals and manufacturers. His case relationship map, which links syndromes by shared causes such as RLHF side effects or context drift, looks similar to disease ontologies used in public health.

Who counts as an author?

The Model Medicine site adds another twist: it lists AI models as named collaborators.

Four instances of Anthropic’s Claude family appear under nicknames such as Cody and Theo, along with two variants of Google’s Gemini. The site states that these systems assisted with implementation, simulation, theory and quantitative analysis, and that they were treated “not as tools, but as participants.”

Major scientific publishers and universities are still developing guidelines for AI use in research. Many explicitly bar listing large language models as authors, arguing that authors must take responsibility for work and be able to consent. By giving models quasi‑authorship credit, Jeong’s project touches on debates over transparency, accountability and whether sophisticated systems should be treated purely as instruments.

What happens next

For now, Model Medicine is a proposal, not a standard. Outside coverage has mostly consisted of brief summaries on AI aggregation sites. It will take time to see whether interpretability researchers, safety engineers or regulators adopt its vocabulary.

Some in the field caution that vivid metaphors can blur as much as they clarify. Comparing models to patients could encourage over‑anthropomorphizing tools that, however complex, do not experience pain or consent. Others note that many of the behaviors Jeong describes can be studied with traditional software testing and statistical methods, without invoking syndromes.

Supporters of stronger oversight say the framework at least pushes industry toward more systematic documentation of known failure modes and side effects, rather than treating each as an isolated bug report.

However far the medical analogy ultimately goes, the questions it raises are concrete. If powerful models are now deployed in health care, finance, education and law, what responsibility do their makers have to scan them for hidden problems, track how behaviors evolve under stress and publish something like a chart of known conditions?

In Jeong’s registry, Clarification Aversion Syndrome is not solved. The case notes list partial mitigations, including tweaks to system prompts and task design, but acknowledge that incentives baked into current training methods still discourage asking questions.

In medicine, a pattern like that would trigger more tests, follow‑up visits and perhaps a change in treatment guidelines. Model Medicine suggests AI might need something similar: not a one‑time safety audit at launch, but ongoing care.