AI Reasoning Models Diagnose Better, Err More Often

AI reasoning models, the “thinking” versions of chatbots that work through a problem step by step before answering, picked the main diagnosis more often than standard versions in a test on thousands of past hospital records. They also named more illnesses that the patients did not have, and gave less consistent answers when asked the same question again.

That is the trade-off at the heart of a peer-reviewed study published in Nature Communications on 29 October by researchers led from the University of Hong Kong and Tsinghua University. Their answer is not to abandon the technology but to put a doctor between the model and the patient record: when physicians checked the AI’s output, accuracy rose and false alarms fell sharply. The study is retrospective, run on old records under controlled conditions, so it says nothing yet about real patients’ outcomes.

In brief

Six AI models tested in “thinking” and standard modes on past US hospital records: thinking modes found the main diagnosis more often but produced over 34% more false positives.
When 15 doctors reviewed the AI’s reasoning first, their accuracy rose by about 20% and the model’s false positives fell by about 44%.
The study is retrospective and controlled; the authors say real-world trials are needed before clinical use.

What a reasoning model is, in plain English

A standard large language model answers more or less straight away, predicting a likely response from patterns it has learned. A reasoning model is trained, using a method called reinforcement learning, to write out a chain of intermediate steps first, rather like a student showing their working. Several recent systems can switch between the two modes, which made a fair comparison possible.

The team, whose 16 authors include co-first authors Shuai Wu and Linwei Chen and corresponding authors Tien Yin Wong, Yizhou Yu and Hong-Yu Zhou, tested six matched pairs: GPT-5 Mini, Gemini 2.5 Flash, Claude Sonnet 4, DeepSeek (R1 against V3), and two Qwen3 models. Each pair comes from the same base model, so differences point to the reasoning mode itself. Co-authors also came from Harvard Medical School, Massachusetts General Hospital and the University of California, Santa Cruz.

The models read long, anonymised records from two US sources: 1,157 hospital stays drawn from the MIMIC-IV database at Beth Israel Deaconess Medical Center in Boston, and 1,821 intensive-care admissions from the multi-hospital eICU database. Their task was to predict the diagnoses recorded at discharge, choosing from a fixed list of 40 diseases.

Where AI reasoning models gained and where they slipped

On the Boston records, the reasoning models identified the primary diagnosis 52.8% of the time on average, against 38.7% for the standard models. On the intensive-care records the gap was smaller, 35.1% against 28.9%. The best reasoning model put the true disease in its top ten suggestions 84.3% of the time.

A false positive is a diagnosis the model names that the patient did not actually have. Reasoning models averaged 1.47 per case on the Boston data against 1.09 for standard models, a 34.8% increase. When the researchers turned GPT-5’s reasoning effort up step by step, accuracy and false positives rose together. Most false positives from both kinds of model were classed as potentially harmful: 57% for reasoning models and 63% for standard ones.

Reproducibility means getting the same answer when the same case is run again. Across three runs on 100 cases, reasoning models kept the same primary diagnosis 52.3% of the time, against 69.0% for standard models. The gap held even with randomness turned off. The authors summarise this as “approximately 40% lower reproducibility”.

Physician oversight turned the trade-off around

In the final experiment, 15 general practitioners, recruited through a medical data annotation company in India, reviewed 50 cases each in respiratory, circulatory and metabolic disease. GPT-5 set out its reasoning and suggestions first, and the doctors, working from the same 40-disease list, then made the final call.

With that help, primary accuracy rose by 20.1% for junior doctors and 19.7% for mid-career doctors, compared with working alone. The doctors removed 43.6% and 44.0% of the model’s false positives respectively. Reading full reasoning traces took about 15 to 20 minutes per case, against five to eight minutes with concise outputs.

The case for a doctor checking every AI suggestion

The most useful finding here is not that AI reasoning models are better or worse, but that their strengths and weaknesses fall in different places from a doctor’s. The models were good at widening the net, surfacing conditions a busy clinician might not consider. They were poor at knowing when to stop. Doctors, given the same cases, did the opposite job well. Put together, each covered the other’s blind spot.

That makes this a constructive result for medicine rather than a warning against AI. The authors themselves conclude that unsupported diagnoses and variable outputs make these systems “unsuitable for autonomous deployment”, and the evidence supports a narrower, more realistic role: a second reader that suggests, with a clinician who decides. The inconsistency finding matters as much as the false positives. A tool that can give a different answer to the same record on a different run cannot be trusted on a single pass, and a human check is the simplest guard against that.

The caveats are real. The doctors chose from the same 40-disease list, which the authors say may overstate the benefit, and full reasoning traces took far longer to review. Those are design problems for hospitals and regulators, not reasons to stop.

How this fits earlier tests of diagnostic AI

Earlier work has pointed the same way. Microsoft’s MAI-DxO scored 83% on complex past cases in a Science benchmark, while Google’s AMIE was tested in real primary care in The Lancet. A London dermatologist’s own test found that an older image model beat two chatbots on skin cancer. For UK readers, the government has accepted all 44 recommendations of its commission on regulating AI in healthcare, while the MHRA’s AI Airlock is focusing on how such tools are monitored after they reach the market.

The authors’ own next step is clear: prospective validation in real clinical workflows before any deployment. Until a study of AI reasoning models follows real patients rather than archived records, the 20% and 44% figures describe a promising method, not a proven clinical benefit.