How Much Can We Trust ChatGPT’s Medical Information?
AI assistance was used for translation
Today, I accompanied my husband to a follow-up appointment after his retinal detachment surgery.
He had surgery in June and has been seeing his ophthalmologist regularly ever since. Although he hasn’t noticed any real change in his vision, today’s appointment brought some encouraging news.
The doctor showed us images taken using optical coherence tomography (OCT), a test that produces detailed cross-sectional images of the retina. Comparing today’s images with those from his previous appointment, he explained that my husband’s retina was showing improvement.
I could see the difference myself.
The doctor explained that recovering vision takes time, but he expects my husband’s sight to improve further. He also mentioned that cataract surgery might be necessary in the future and discussed how it could help.
When my husband asked about the risks of cataract surgery, the doctor reassured him that it is a commonly performed procedure and that he saw no particular reason for concern.
My husband had seemed discouraged by the lack of noticeable improvement in his vision since the surgery, so I was happy to hear such an encouraging assessment.
Still, living in a foreign country has its challenges. I’m comfortable with everyday conversations in English, but medical terminology is another matter. Sometimes I wonder whether I’ve understood everything correctly.
That’s why, after we returned home, I decided to ask ChatGPT to help me make sense of the doctor’s explanation.
And that’s where things took an unexpected turn.
ChatGPT began by explaining how the retina recovers after surgery. But then, seemingly out of nowhere, it introduced another possibility: retinal swelling can also make the retina appear thicker.
I found this puzzling.
The doctor had specifically told us that my husband’s retina was improving. He hadn’t mentioned swelling at all, and his outlook for further recovery had been encouraging.
So why was ChatGPT bringing up a potential problem that the doctor hadn’t even raised, especially when it hadn’t seen the examination images?
I kept asking ChatGPT why it had chosen to introduce this possibility.
Eventually, it explained that when answering medical questions, it tends to prioritize safety, which can lead it to include possible risks and negative outcomes.
But what exactly does safety mean in this context?
All I wanted was a clearer understanding of what the doctor had said in English. I wasn’t asking ChatGPT to identify additional risks that hadn’t been part of his assessment.
Before my husband’s surgery, the doctor had carefully explained the potential risks. Yet at both his previous appointment and today’s examination, there had been no such concerns raised. Instead, the doctor had consistently given us an encouraging outlook on his recovery.
I had every reason to be pleased with what we’d heard.
The experience left me thinking about something much broader than my husband’s eye condition.
AI can explain highly specialized subjects in impressive detail. But that doesn’t mean those explanations necessarily apply to the person asking the question.
I work professionally as an astrologer, and I’ve encountered a similar problem in my own work.
Sometimes I ask ChatGPT to polish the language of a horoscope reading I’ve written after interpreting an astrological chart. I’m not asking it to change my interpretation, only to make the writing flow more naturally.
Yet it sometimes changes the meaning of what I’ve written, replacing my original interpretation with something that sounds like a convincing horoscope but says something quite different.
In other words, AI can produce language that sounds perfectly appropriate for an astrological reading without faithfully representing the interpretation behind it.
I’ve had to point this out many times.
I’ve noticed something similar with AI-generated illustrations and photo editing. I might ask it to edit a photograph of myself, only to end up looking like a completely different person. Or I might ask it to edit a picture of one of our cats, only to find that the resulting cat doesn’t look like ours at all.
Get Himiko’s stories in your inbox
Join Medium for free to get updates from this writer.
And now, I was seeing a similar pattern in the way AI handled medical information.
Of course, medicine and astrology are very different fields. Medicine is grounded in scientific research, clinical evidence, and diagnostic testing.
But when it comes to AI-generated explanations, there’s a common problem: AI can arrange specialized terms into fluent, convincing language, whether or not the result accurately reflects the situation at hand.
With medical information, the consequences of that problem can be far more serious.
There is an enormous amount of medical knowledge available, covering everything from diseases and symptoms to complications, treatments, and possible risks. AI can draw on all of it, producing detailed explanations about almost any medical condition.
But a medical risk that exists in general is not necessarily a risk that applies to a particular patient.
AI hasn’t examined the person asking the question. It may not have seen their medical records or test results. Yet it can combine general medical knowledge with technical terminology and produce an answer that sounds remarkably specific to that person’s condition.
The result can be persuasive, especially to someone who doesn’t have a medical background.
After all, when an explanation is detailed, well organized, and filled with professional terminology, it’s easy to assume that the information must be reliable.
That was what troubled me most about my conversation with ChatGPT.
I had interpreted the doctor’s comments positively because they were consistent with his previous assessments and supported by the latest examination results.
But what if someone in a very different emotional situation received the same kind of response?
Imagine someone dealing with a serious illness, going through treatment while living with fear and uncertainty about the future.
Perhaps they’ve just returned from a hospital appointment where their doctor gave them encouraging news. For the first time in a while, they feel a little more hopeful.
Wanting to understand the explanation better, they turn to AI.
Instead of simply clarifying what the doctor said, AI begins listing possible complications, worsening symptoms, or other negative outcomes that the doctor never mentioned.
How would that person feel?
They might start wondering, “My doctor didn’t mention any of this. Does that mean my condition could actually be worse than I thought?”
For someone already frightened or emotionally exhausted by illness, words like these can carry enormous weight.
I don’t think there’s anything inherently wrong with using AI to help understand medical information. In fact, as someone living abroad, I find it incredibly convenient to be able to look up unfamiliar terminology or have complicated explanations presented in language I can understand.
But convenience and reliability are not necessarily the same thing.
AI can seem to have an endless supply of knowledge, and it often presents its answers with remarkable confidence. Yet sometimes it fills in missing information with assumptions, introduces possibilities it hasn’t verified, or offers general medical explanations that have little to do with the person asking the question.
An explanation can sound highly professional without being accurate or relevant to an individual’s situation.
And this matters even more when someone is struggling with a serious illness or feeling anxious about what the future might hold.
At times like that, I think it’s especially important not to accept everything AI says at face value.
Just because AI mentions a possible complication doesn’t mean that complication applies to you. And if AI introduces a concern your doctor hasn’t raised, it doesn’t mean that something is secretly going wrong with your health.
What I realized from this experience is that AI needs not only the ability to provide a wealth of information, but also the ability to recognize which information is truly necessary for the person asking.
AI is an incredibly useful tool, but the vast amount of information it provides isn’t necessarily the information we personally need.
Especially when it comes to matters of health and life itself, I think it’s important to keep that distinction in mind.
★If you enjoyed this story, I’d be happy if you gave it a few claps.
