Will AI replace your doctor?

We tried with lawyers, which didn’t go well (judges don’t appreciate the hallucinated citations). We’re dead-set on trying with artists, and the result is the mass sloppification of the internet. Why not try with doctors?

Today, I’d like to point to a problem with medical AI besides rampant hallucinations, lack of actual understanding, lack of human connection, and all the rest. Today, I’d like to point out that AI simply doesn’t have the capacity to reliably, continuously integrate new data.

Now, I’m no doctor myself, but I am a doctor’s son, and I do help run his blog, which means I read all the articles before publication, not to mention dinner table conversation. As a result, I have an interest in the topic and some knowledge to boot. I know that medicine is an ever-evolving field. What we knew to be true two decades ago is half disproven; what we know to be true today is thoroughly imperfect; what we learn tomorrow is still unknown. Medical practice lags medical research by a decade or more (especially corporate, mass-production medicine, which can be as much as 17 years behind, not helped by pharma money).

If we look at how AI works, though, we find that AI needs to be trained off of an amount of data at least proportional to the complexity of the topic. For a topic as tremendously complex as the human body, AI would need to be trained off of a truly mammoth amount of data. Thanks to the mass collection of data in the hospital and insurance system, we might have that data (albeit I hope you don’t mind the complete disregard for medical privacy that would involve- the government probably reads it all anyway).

All that data, though, was generated with a mélange of research from the past thirty or so years, not with the latest medical understanding. Doctors do have to pursue continuing education, but inevitably all but the most persistently and rigorously curious are going to end up operating on the orthodoxy they were taught when they went through med school- one, two, three, four decades ago. They’ll have updated it in part, of course, a bit here and a bit there, adding five-year-old research there and ten-year-old research there, but even the most enterprising doctor can’t manage to apply today’s research to a case he finished five years ago.

As a result, all but a tiny fraction of data available for AI training deals with potentially outdated research. Much of it will be still-viable answers, non-ideal or even unchanged, but any recent advances in understanding will be, if not completely missing, certainly drowned out by the mass of other information the AI has been trained towards. Training the AI to favor that tiny fraction of newer data or to prefer innovative techniques isn’t much better, though; that data hasn’t been tested enough to rely on, not yet, and will inevitably be wrong in a lot of ways, just as the orthodoxy it contradicts was wrong.

Medical AI will always be trained on outdated data, if it has anything approaching adequate amounts of training data. The only source we have for such large amounts of data on medical practice is the institutional system, hospitals and clinic networks and state-run information sharing. These bodies, however, will inevitably reflect the orthodoxy of Current Science, meaning the majority consensus from over a decade ago combined with pharmaceutical and corporate interests, including risk-assessments which value money over patient care.

The AI trained on such data will necessarily imitate it. The result? Medical AI doctors will provide out-of-date, finance-centered ‘care’, a mash of the past few decades of corporate orthodoxy that cannot be trained in new techniques and research until it has been mainstream for a long while, adding years of wait-time to the already-bad lag between research and actual practice. Now imagine the result if medical AI replaced much of those institutions, became the source of much of that data. We’d have to let AI experiment on people en masse by directing it to innovate¹ (a horrendous idea, if you remember the infamous glue-on-pizza incident or its ilk) before feeding the data into the AI (let AI feed AI data, compounding error margins), wait even longer to implement new research because we have to let the reduced number of ‘data-sources’ (human doctors) work longer before they produce the same amount of data (letting us re-train the AI),² or just freeze innovation.

In sum: AI doctors will always be worse than motivated human doctors and will probably be less up-to-date than even dispassionate, over-worked corporate medical care.

God bless.

Footnotes

1. Imagine the lawsuits and the malpractice! I’d say they wouldn’t ever do this, considering the risk to their profit margins, but they would- after getting immunity from prosecution. See: COVID. ↩︎
2. Oh, and medical AI makes a good excuse for government to mandate doctors depositing all medical data into a central databank, which the government would naturally have access to. ↩︎