Most people instinctively give more details to their doctor than to an app. Headaches are more detailed, including where they hurt, how long they last, and whether they are accompanied by nausea or photosensitivity.
In contrast, digital tools typically obtain compressed versions. Researchers now say this practice has spread to medical chatbots as well.
A new study measured exactly how much information people omit when they think an AI is reading their symptoms, and whether those omissions change the usefulness of their reports.
AI reduces human input
A team led by Moritz Reis, a researcher at the Institute of Psychology at the University of Würzburg (JMU), conducted a simple test on 500 adults in the UK.
Each person wrote two symptom reports, one for an unusual headache and one for a flu-like illness.
Half of them were told that their doctor would read the instructions. The other half were told that an AI chatbot would do it for them.
The wording on the page changed, but the steps remained the same. Reports written for human doctors averaged 256 characters. Chatbot reports averaged 229 characters, about one sentence shorter.
Review reports blindly
To see if shorter also means worse, the team ran all of their reports through a scoring system.
The question was how useful the descriptions were in determining who needed emergency treatment.
A higher score means that doctors can read you and give you advice with confidence.
Reports sent to chatbots scored 8% lower on average. Four licensed physicians, two neurologists and two pulmonologists reviewed the data.
They looked at a random subset without knowing whether the reports were written for doctors or chatbots. Their judgments were consistent with the AI’s scoring.
Small omissions quickly add up
What is cut out is the kind of background on which the doctor builds the picture. For example, how long has your headache lasted or what does your cough sound like at 3am?
None of this is difficult to write. But when people thought a machine was reading them, they simply wrote less.
The researchers tracked the decline in quality directly to length. Fewer characters means the report is less useful for self-triage, which is an early filter for determining who needs a doctor right now.
Chatbot accuracy may decrease
AI tools are typically tested based on standardized scenarios, rather than the messy paragraphs that people actually type.
This problem is often hidden. Even if a chatbot passes a benchmark, it can mistakenly transfer a real patient if the patient only tells half the story.
This quality gap existed not only among participants who had imagined the symptoms, but also among participants who had associated symptoms at the time.
Another paper on the accuracy of online symptom checkers reported similar caveats. Laboratory-level accuracy cannot withstand everyday user input.
be misunderstood by the machine
So why are people so stingy with chatbots? The researchers explain a phenomenon called uniqueness neglect. This is the belief that AI sees you as a category rather than an individual.
If the tool just matches a pattern, you’re left wondering why you need to go through all the weird details.
“Many people think that AI cannot grasp the individual nuances of a personal situation, but simply matches standardized patterns,” explained Professor Wilfried Kunde.
Privacy concerns may also be part of it. So may general skepticism about whether algorithms can actually diagnose anything.
Previous research by the same group found that people rated the same medical advice as less trustworthy and not worth following the moment they were told it was written by an AI.
Design better questions
The fix is not a smarter model, but a smarter interview. The researchers argue that medical chatbots should actively prompt users for details that a doctor might ask.
Rather than waiting for users to guess what’s important, show details like duration, severity, and what went better or worse.
Providing specific examples of strong explanations can help strengthen medical advice.
You can also describe what the system does with the information. Understanding the logic of a tool can help you type more, not less.
“If we don’t trust machines to understand our uniqueness, we may unconsciously withhold the information machines need to provide accurate assistance,” Reis says.
Behavior may change depending on the actual illness
Participants wrote about situations they were asked to imagine, rather than situations in which they were actually sick and needed emergency care.
The researchers note that real-world reports, where emotional risks are higher, may differ in ways that were not captured in this experiment.
Whether this gap persists in actual clinical practice remains to be verified and requires further research.
The human side of AI triage
Until now, no one had measured what patients omitted before the AI recognized their questions. Evaluations of medical chatbots focused almost entirely on the model’s side of the conversation.
This study reversed that. It showed numbers on the human side, showing that when healthy adults spoke to the machine, the quality of the report was 8% lower due to 27 fewer characters.
8% per person may sound modest. Multiplied by the millions of queries hitting symptom checkers and consumer chatbots, missing details confirm that triage decisions were made based on incomplete information.
Developers now have specific problems they must design for. Patients have a reason to input more rather than less, even when humans aren’t reading it.
This research natural health.
—–
Like what you read? Subscribe to our newsletter for fascinating articles, exclusive content and the latest updates.
Check us out on EarthSnap, the free app from Eric Ralls and Earth.com.
—–
