A Different Kind of Review
The ARISE network, a Stanford-Harvard research network, set out to do something the clinical AI field badly needed: organize the actual evidence base, not just the funding announcements, press releases, or benchmark leaderboards.
The inaugural State of Clinical AI Report 2026 synthesizes the most important developments, evidence, and open problems in clinical AI. The authors - Peter Brodeur, Ethan Goh, Adam Rodman, and Jonathan Chen - frame the problem carefully. AI is already moving into clinical workflows. The question is not whether medicine will use it. The question is which parts are holding up when the setting gets messy.
What they found is more nuanced than either enthusiasts or skeptics will find comfortable.
The Bottom Line
The report does not say clinical AI is fake. It says the strongest evidence is concentrated in narrower, better-specified workflows, while the weakest evidence often comes from benchmark-style evaluations that look much cleaner than real medicine.
Where AI Actually Works: Prediction at Scale
The clearest signal in the report is that AI is strongest when the task is prediction at scale: finding patterns across large datasets, monitoring risk trajectories, and surfacing early warnings before a human clinician would reasonably detect the signal.
Examples highlighted by the report and related literature include:
- Hospital-based AI systems that predicted patient deterioration earlier than standard clinical alerting systems
- AI-derived biological age and risk markers that predicted mortality more accurately than chronological age alone
- Large-scale prognostic models that forecast future diagnoses across diverse conditions
These are not chatbots pretending to be doctors. They are population-scale pattern recognition systems. That distinction matters.
This is where AI is genuinely better than physicians: not empathy, not clinical accountability, not bedside judgment, but scale. It can ingest thousands of data points per patient across thousands of patients and find signals that no individual clinician can manually track.
Where AI Struggles: Uncertainty and Multi-Step Reasoning
The most important cautionary finding concerns ambiguity.
Large language models can perform impressively on structured diagnostic tasks. They can match or outperform physicians on some standardized cases. But real clinical work is not a standardized case. It is incomplete information, unclear histories, competing diagnoses, missing labs, time pressure, and a patient who says "I just feel off."
When evaluations start to look more like that, performance gets less clean.
The deeper problem is not simply that AI systems are sometimes wrong. Clinicians are sometimes wrong too. The problem is that AI systems often state wrong answers with the same tone and surface confidence as correct ones. In medicine, that is a dangerous failure mode. A system that says "I am not sure" creates space for human correction. A system that sounds certain while being wrong creates over-reliance.
This is why the evaluation literature matters. A systematic review by Bedi and colleagues categorized 519 healthcare LLM evaluation studies and found that only 5% used real patient care data. The most common task was medical knowledge assessment, often exam-style question answering, and calibration or uncertainty was rarely measured.
The gap between benchmark performance and clinical performance is not a defect in one model. It is structural. Benchmarks are cleaner than medicine.
The Over-Reliance Problem
The ARISE report repeatedly returns to a pattern that should make every clinician pause: humans plus AI can outperform humans alone, but only when the workflow is designed well.
That last clause is doing a lot of work.
If the AI is optional, transparent, and easy to challenge, it can function like a second reader or cognitive support tool. If the AI is embedded as a confident recommendation that no one has time to interrogate, it can create automation bias. Clinicians may follow incorrect outputs because the system feels validated, the interface looks polished, and the recommendation is plausible enough.
This is why AI safety is not just about model accuracy. It is about workflow design, incentives, institutional governance, and how easy it is for a busy clinician to say no to the machine.
Optional AI assistance with final human judgment is very different from AI as a quiet decision-maker.
The Evaluation Problem
If there is a single structural issue the report identifies as most consequential, it is the poverty of evaluation methodology in the clinical AI literature.
Exam benchmarks dominate because they are easy to build, standardize, and publish. Real patient data studies are harder. Prospective clinical trials are harder still. Measuring uncertainty, bias, harm, and workflow failure adds complexity that many papers avoid.
That leaves us with a literature that often tells us how AI performs in favorable test conditions and much less about how it performs where it actually matters.
The right evidence bar is not mysterious. Drugs that work in vitro but fail in trials do not get treated as validated therapies. Clinical AI that works on clean benchmarks but fails in deployed care should not be treated as clinically validated either.
For a practical version of this principle, see when not to use ChatGPT. The tool matters, but the task and evidence standard matter more.
Patient-Facing AI: Promising and Under-Evaluated
One area the ARISE report flags as expanding rapidly but still under-evaluated is patient-facing AI: chatbots and conversational systems that interact directly with patients outside clinical encounters.
In simulated settings, patient-facing AI can answer questions, provide education, guide symptom assessment, and help patients navigate complex information. That is promising. But the clinical risk is different from clinician-facing decision support.
Patients cannot be expected to serve as safety monitors for the system. They may not know when an answer is incomplete, outdated, or subtly wrong. They may also over-trust a polished interface because it sounds calm, fluent, and authoritative.
The two questions I care about most are simple:
- What happens when the AI is uncertain?
- What happens when the patient needs escalation to a human?
If a patient-facing tool cannot answer those questions clearly, it is not ready for high-stakes use.
The Honest Scorecard
The ARISE report resists the temptation to give clinical AI a simple grade. That is the right move. The field is not one thing.
AI is ready, or close to ready, for:
- Large-scale risk prediction and early warning
- Ambient documentation and administrative summarization
- Specific imaging tasks with well-defined endpoints
- Research support: literature synthesis, data structuring, pattern identification
AI requires caution in:
- High-stakes clinical decisions with incomplete information
- Any setting where confident errors have serious consequences
- Patient-facing applications without clear human escalation pathways
The field still needs:
- Far more studies using real patient data rather than exam benchmarks
- Consistent evaluation of uncertainty and demographic bias
- Prospective and post-deployment studies, not just retrospective performance claims
- Workflow design that treats human behavior as part of the system
None of this is a reason to slow adoption. It is a map of where to invest attention as adoption accelerates.
The Report's Real Message
The naive debate is whether AI in medicine "works."
That is the wrong question.
Some of it works. Some of it works only in demos. Some of it works technically but fails because no one uses it. Some of it helps clinicians reason. Some of it makes clinicians overconfident. Some of it should be deployed now. Some of it needs trials, governance, and a much harder look at harm.
The ARISE report is useful because it does not flatten those distinctions. It asks the question clinicians should be asking: what holds up in practice?
That is the question that matters.
Sources: State of Clinical AI Report 2026, ARISE Network; Stanford Department of Medicine summary; Bedi et al., A Systematic Review of Testing and Evaluation of Healthcare Applications of Large Language Models.