Log inSign Up
In March 2026, ARISE, the AI Research and Science Evaluation network led jointly by Stanford and Harvard, published the State of Clinical AI Report 2026. The authors are Peter Brodeur, Ethan Goh, Adam Rodman, and Jonathan H. Chen, the same Stanford lead behind MedAgentBench and the senior author on NOHARM. It is the most carefully synthesised assessment of the medical AI field that has been published to date, and it is now the reference document that researchers, clinicians, and regulators are pointing to when they describe what is wrong with the current state of medical AI.
We read it on the day it came out. What struck us is that almost every weakness it identifies is a problem BloodGPT was specifically built to solve. That overlap is not a coincidence. The report is written by the people who run the benchmarks that have driven our architectural choices for two years.
This post is our walk through the gaps the report names, and what we did about each one before the report told us we had to.
The headline finding of the report is short. Model capability is no longer the bottleneck. Frontier models already match or beat physicians on controlled clinical reasoning tests. The unsolved problems are now reliability, provenance, and safety. That is the layer the field has not figured out, and it is the layer everything downstream depends on.
The report breaks the unsolved problems into a small number of categories, each backed by a specific published benchmark or study. Brittleness on messy data. Lack of provenance in AI outputs. Errors of omission. Nondeterminism. Architectural failures where the best individual components produce a worse overall system than less optimised ones. And the near-total absence of validation on real patient data.
Each category below cites the underlying benchmark.
Brittleness. Models that perform well on clean inputs degrade sharply when data is missing, ambiguous, or structured differently from training distribution. The report cites "None of the other answers" and script-concordance findings: when answer patterns shift, accuracy drops in a way that is inconsistent with the model actually reasoning. The benchmark scores reflect pattern-matching more than they reflect clinical reasoning. Clean-room performance does not survive production data.
No provenance. A Nature Communications study, SourceCheckup (Wu et al., 2025), evaluated seven popular LLMs on 800 medical questions and 58,000 statement-source pairs. Between 50% and 90% of responses were not fully supported by the sources they cited. Even GPT-4o with retrieval-augmented generation left roughly 30% of statements without backing evidence. Claim-level grounding is the field's stated path to clinical trust. Almost no current system implements it.
Omission. The single most important finding from NOHARM, which we covered separately, is that 77% of severe potential harms across 31 tested AI models came from missing recommendations, not wrong ones. A test not ordered. A medication interaction not flagged. An escalation not made. Top models still produced potentially severe harm in up to 22% of cases. Standard knowledge benchmarks did not predict any of it.
Nondeterminism. Same input, different output. The report treats inconsistency and overconfidence as core failure modes, citing MetaMedQA and related work on metacognition. Models rarely recognise their own uncertainty and almost never abstain when they should. An interpretation that varies run-to-run cannot be audited and cannot be deployed in a regulated setting.
Architecture failure. The most counterintuitive finding in the report. The Optimization Paradox in Clinical AI Multi-Agent Systems (Bedi et al., Stanford, 2025) built a "best of breed" diagnostic system from the strongest individual AI agent on every component task. Information accuracy at the component level reached 85.5%. Overall diagnostic accuracy collapsed to 67.7%. A less optimised multi-agent system, using the same kinds of models but with different information flow between them, scored 77.4%. The failure was not in any of the components. It was in how they were wired together. The report's conclusion: architecture is the product.
No real-data validation. The most damning finding for the field as a whole. 5% of AI evaluations used real patient data. Under 1% reported actual patient outcomes. Around 95% relied on accuracy alone, ignoring calibration, bias, and fairness. The report identifies real-data, end-to-end validation as the single biggest missing piece in clinical AI today.
We did not start the company in response to this report. We started it because we kept running into these exact failure modes when we tried to use existing AI on actual laboratory data. The report puts numbers on what we saw firsthand.
Here is what we built, mapped to what the report says is missing.
The same patient data, run through BloodGPT repeatedly, produces the same abnormal set, the same recommended tests, the same structured output. We verify this by hashing the results across runs. This is not a property we tuned. It is an architectural property: most of the computation in BloodGPT is not LLM-based. The language model is a component, not the engine.
This is why we hit 100% on Stanford's MedAgentBench across both benchmark generations using a lightweight model, while frontier models with much larger context and more reasoning power topped out at 98%. The same architectural choice is what closes the nondeterminism gap.
Every output produced by BloodGPT is traceable to its inputs. The audit log captures the requesting actor and timestamp, the exact model inputs and outputs with version pinning, and the specific patient data that drove each interpretation step. This is enforced at the pipeline level, not added through logging middleware.
The corresponding capability for source citations on interpretive output is grounding against the patient's own structured record. Every flag has a provenance path back to a specific lab value, a specific reference range, and a specific stage in the pipeline. That is the operational counterpart to what SourceCheckup found missing in general-purpose medical LLMs.
Critical lab values trigger mandatory urgent-review language and a clear action flag, not as a feature you can switch on, but as a stage gate the pipeline cannot skip. No new diagnoses and no medication start/stop recommendations are produced before a dedicated, evidence-bound stage runs. The system is designed around omission as the leading safety risk, not around accuracy as the headline metric.
The architectural premise is that omission is the leading safety risk, not accuracy. A system can be 92% accurate on labelled data and still miss the test that catches the cancer in the 8% of cases where it matters. The pipeline is designed around the 8%.
Language models do not fail loudly on bad input. They produce plausible but wrong output. To prevent that, the pipeline runs a non-LLM validation layer before any interpretation stage: structural completeness checks against the FHIR R4 schema, duplicate detection by patient and timestamp, and a heuristic gate that rejects out-of-range values inconsistent with the unit declaration (a glucose value of 200 in mmol/L versus 200 in mg/dL triggers different behaviour). The language model never receives inputs the validation layer has not vouched for.
BloodGPT is a sequential, multi-stage pipeline. Retrieval, harmonisation, canonical model construction, validation, abnormality detection, contextual interpretation. Each stage is validated end-to-end, not in isolation. This is exactly what the Optimization Paradox study warns against, picking the best component for each subtask and assuming the whole works. Two years of architectural work went into making the connections between stages the part we optimise for, not just the stages themselves.
The MedAgentBench result is one expression of this.
We are currently running a structured, multi-stage clinical validation on real hospital FHIR R4 data with Assuta Medical Centers, approximately 100,000 anonymised patient records. The protocol covers deterministic extraction and normalisation, abnormality detection with precision/recall/F1, and grounded interpretation scored by clinicians against ground-truth labels.
This is the gap we are most explicit about still being inside. The validation is in progress, not complete. We mention this in the same place we mention the 100% on MedAgentBench, because they answer different questions. MedAgentBench tested whether the pipeline executes structured EHR tasks correctly under benchmark conditions. The Assuta study tests whether it produces clinically reliable output on the heterogeneous record formats of an actual hospital system.
The ARISE report did three things at once. It synthesised the field's most important problems into a single reference document. It made those problems addressable, with named benchmarks and numbers behind each one. And it made the architectural conclusion explicit: better models will not close these gaps. Better architecture will.
That conclusion is now the field's reference position. The companies building real clinical AI infrastructure now have a shared vocabulary for what counts as a serious product and what does not.
If you are operating a clinical lab, a hospital, or a healthcare system, the questions to ask any medical AI vendor now have specific names. Does the system produce the same output for the same input? Is every claim traceable to a source? Does it handle omission as a first-class failure mode, not an edge case? Is the architecture validated end-to-end, or only at the component level? Has it been tested on real patient data?
For BloodGPT, the answer to each of those is yes, with the validation work on real data ongoing rather than complete. That is the honest version. The dishonest version, the one the report's authors are quietly tired of hearing from medical AI vendors, is that any of these problems are about to be solved by the next model release. They are not. They were never going to be. They are architectural problems, and the architecture is now the thing the field is building around.
Or should be.
If you are working on medical AI infrastructure and want to compare notes on how to actually close these gaps, we'd be glad to talk. Reach out at [email protected].
About BloodGPT BloodGPT is a B2B healthcare AI platform that turns unstructured laboratory data into standardised FHIR R4 records with clinical interpretation. We operate in Israel, the UK, Nigeria, Greece, and Poland, with enterprise deployments including white-label, API, and on-premise options. In April 2026 we achieved 100% on Stanford's MedAgentBench clinical benchmark across both generations, the first AI system to do so. Learn more at bloodgpt.com.
References
Transform how you deliver lab results to patients while enhancing your practice's efficiency and value. Try BloodGPT or contact us for a personalized demo
Book a Demo
