I’m actually not clear on what percentage of the doctor’s work-time is spend doing things other than talking to patients (like arguing with insurance or making records).
The solution to doctors arguing with insurance isn't "have an AI do it" it's universal health care so doctors don't need to worry about insurance in the slightest
I worked on software for electronic medical record note taking and I'm not sure how an LLM can help a doctor speed that up tbh. All of the stats need to be typed into the computer regardless. The LLM can't really speed that up?
I’d also prefer single payer, but nothing except ourselves has been stopping us from doing that, and we haven’t changed much. Maybe it’ll happen. But I don’t see any recent tech changes making it so.
Unless somebody manages to make hyper-convincing LLMs and use them for good, I guess. (Note: I think this is a bad path).
I have no expertise and am prepared to be quite wrong, but I wonder if llm's would be good at listening to a session, and/or a doctor dictating, and putting the right stats in the right place, and the dictated case history into a note.
I think llm's are alright at speech recognition and that sort of unstructured to structured text manipulation. At least, in my corner of the customer success world I've seen some uses along those lines
My doctor was in a pilot program for this exact thing. It recorded our conversation and created the after visit summary.
It's summary was that I wasn't taking my antibiotics (I was, neither I nor my doctor said anything to the contrary). Luckily my doctor was very skeptical of the whole thing and carefully reviewed the notes, but this could be an absolute disaster if it hallucinates something more nefarious and the doctor isn't diligent about reviewing
From me talking to Doctors, it seems most of their job is handling records, contacting insurance, prior auth, talking to pharmacists, etc. This is despite having billing specialists and admins out the wazoo.
Also their care is pretty much completely decided by insurance. What surgeries they can perform, what medicine they can give, how much, what materials they can use for surgery, and on and on. Your doctor is practicing shockingly little medicine, your real doctor is thousands of pages of guidelines created by insurers and peer-to-peer doctors who you will never meet.
My experience with the current stuff on the market is you get out what you put in
If you put in a very detailed and high quality, precisely defined question and also provide a framework for how you would like it to reason and execute a task, then you can get out a pretty good response
But the less effort you put in the less accurate the outcome is
If a bad doctor is someone who puts in less effort, is less precise, and less detail oriented, it's difficult to see how AI improves on the situation at all
Especially current iterations of AI that don't really prompt the users for more details or recognize when users need to be more precise
IMO the problem is that, at least right now, the AI can't examine the patient itself, it has to be fed information from the doctor. This step means bad doctors are likely to provide the AI with bad information and reduce it's effectiveness (or cause the AI to re-enforce the biases of the doctor by only feeding it the information they see as relevant).
Not sure about what will happen with software engineers, lawyers, or doctors, but I do know how computer assistance worked decades ago when it took over retail clerks, the net effect was to de-skill and damage the job as a career, by bringing everyone up to the same baseline level management lost interest in building skills above that baseline level.
So until the 1970's shopping clerk was a medium-skill and prestige job. Each clerk had to know the prices for all the items in your store because of the danger of price-tag switching(1). So clerks who knew all the prices were faster at checking out then clerks who had to look up the prices in their book, and reducing customer friction is hugely valuable for stores. So during this era store clerk is a reasonable career, you could have a middle-class lifestyle from working retail, there are people who went from clerk to CEO, and even those who weren't ambitious could just find a stable path to support their family.
Then the UPC code, laser scanner, and product/price database came along in the 1970's. The UPC code is printed in a more permanent way so switching tags is not as big a threat (2). Changing prices is just a database update, rather than printing new tags for every item and having the clerks memorize the new price. And there is a natural language description of every item that the register can display, so you don't have to keep the clerk around to be able to tell the difference between the expensive dress and the cheap dress- it will say the brand and description. This vastly improved the performance of a new clerk, but also decreased the value of the more experienced clerk. The result was a great hallowing-out of the retail sector employment, the so-called "McJob" of the 1990's.
But the result was things like Circuit City (in its death throes) firing all of their experienced retail employees (3) because the management didn't think that experience was worth paying for. This is actually the same sort of process that Marx had noted about factory jobs in the 19th century- he called it the alienation of labor, this is capital investment replacing skilled labor, to the benefit of the owners of the investment- but since retail jobs largely code as female no one really paid much attention to it. It never became a subject of national conversation.
1: This also created a limit on store size: you couldn't have something like a modern supercenter (e.g. Costco, Walmart, Target) because a single clerk couldn't know the prices for such a wide assortment of goods. In department stores in the pre-computer era every section had its own checkout area, you would buy the pots in the housewares section and then go to the women's clothes area and buy that separately, and they would use store credit to make the transaction as friction-less as possible.
2: Because in the old days a person with a price tag gun would come along and put the price directly onto each item when a price changed, so you'd have each orange with a "10p" sticker on it, and now it's a code entry and only the database entry needs to change, the UPC can be much more permanently printed.
3: https://abcnews.go.com/GMA/story?id=2994476 all employees paid above a certain amount were laid off, which pretty much meant they were the ones who had stuck around for a while and actually knew the business well and were good at their jobs.
Considering how little interest doctors have taken in some of my medical problems I'll be happy to have AI help me to investigate things myself. And for a lot of people in the US it may make the difference between not being able to afford a doctor vs getting some advice.
Ok, what about an Average doctor with an AI? Or how about a Bad doctor with an AI?
AI assisted medcare will be good if it catches some amount of misdiagnoses
AI will be terrible if it winds up reinforcing misdiagnoses
My suspicion is that AI will act as a force multiplier to some extent, more than a safety net
Yes, some top percentage of performers will get some percentage of performance gain out of AI
But it will not make average performers great or bad performers good. It will make bad performers worse