Almost everyone has tried it. A report arrives, the appointment is weeks away, and you type the values into whichever general-purpose chatbot is already on your phone. It answers immediately, fluently, and with complete confidence. Sometimes it is useful. Often it is wrong in ways that are very hard to see.
The failure is structural, not accidental
A general assistant is trained to be helpful across every subject at once. Asked about a haemoglobin value it will produce an answer, because producing an answer is what it is for. It has no way of knowing that the reference range printed on your report may differ from the one it learned, that the units on your page may not be the units it assumed, or that the question you asked is one nobody should answer without seeing the document.
So it answers anyway. The result reads exactly like a correct answer, because fluency and accuracy are not the same thing and nothing in the output distinguishes them.
Three mistakes we see repeatedly
- Answering against the wrong range. Laboratories differ. Two can print different ranges for the same test and both be right for the people they serve. A general model answers against whatever it absorbed in training, which is often from a different country entirely.
- Restating numbers that are not on the page. Asked to summarise a report, a general model will sometimes produce a plausible value that was never in the document. On a shopping list that is an irritation. On a blood test it is a different category of problem.
- Never saying it does not know. This is the most damaging of the three. A health answer that arrives with the same certainty whether it is well founded or invented gives the reader no way to calibrate.
What a health-specific system does differently
It starts from the document rather than from memory. It reads the values and the range printed beside them, and it explains the result against that printed range rather than against a figure of its own. Where it has no reliable answer it says so, which is a feature and not a limitation.
None of this makes it cleverer than a general model. It makes it narrower, and narrow is the correct shape for this problem.
This article explains how our technology works. It is health information, not medical advice.