ChatGPT-5, Gemini 2.5 Flash and DeepSeek-V3.2-Exp all scored highly on a small set of paediatric infectious-disease cases, according to a study reported by Bioengineer.org. The researchers also flagged gaps in age-specific decisions, treatment detail and safety escalation, so near-perfect scores here are not a licence to hand over the clinic.
Watch Desk analysis
What happened
The study, published in Pediatric Research, tested the three models on 12 scenarios covering six paediatric infectious diseases. Each case prompted questions about causes, clinical features, diagnosis, treatment and possible complications. Two paediatric infectious-disease specialists assessed each response.
ChatGPT-5 received the highest reported accuracy and completeness scores, with DeepSeek-V3.2-Exp close behind and Gemini 2.5 Flash somewhat lower. None of the model-to-model differences reported by the study reached statistical significance. Reviewers nevertheless identified weaknesses including age-specific thresholds, escalation criteria, precise dosing and follow-up planning.
Why it matters
The results offer a useful, bounded comparison of how current chatbots handle structured paediatric cases. They do not show how reliably the systems would perform across the variety of real consultations, or whether their answers remain safe when information is incomplete or changing.
The study used only 12 predefined scenarios, and most scores clustered near the top of its scales, limiting the test’s ability to distinguish strong systems from one another. The authors also note that consumer platforms update continuously and do not disclose exact model versions, making the results harder to reproduce.
Our read
High marks on a carefully prepared test are encouraging evidence about a narrow task, not a paediatrician in a browser tab. The practical takeaway is that these systems may help with supervised education or early reasoning, while clinicians remain responsible for checking the answer and acting on it. The ceiling effect is doing quite a bit of work here; a tougher test may tell us more than another parade of impressive-looking scores.
What to watch
- Whether future studies use more cases and more challenging, less structured presentations.
- Whether researchers publish more sensitive evaluation methods that can separate high-performing models.
- Whether later testing uses identifiable model versions and reports how often answers miss safety-critical details.
Discussion spark: Should high scores on structured medical cases be enough to support AI use in supervised clinical reasoning, or should deployment wait for stronger evidence from messier, real-world cases?
Sources and evidence
- ChatGPT-5, Gemini and DeepSeek ace pediatric infection questions, study finds – Bioengineer.org (17 October 2026, 22:41 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.