Back in September 2023 I noted that the biggest hurdle for AI completely replacing physicians is the physicality of the job. Sure, LLMs are good at giving differential diagnoses and faking empathy once somebody’s problem has been reduced to text, but the art of medicine is in the act of seeing, feeling, smelling, etc. [Note: Although increasingly less so, as doctors and trainees are becoming experts at treating patients in the chart and not those in front of them, making themselves the perfect foils for replacement; queue photo of the old man yelling at clouds. ] If clankers have any hope of replacing humans, they’d better get some senses.
At first glance, a recent Nature Medicine paper aimed to do just that by introducing what the group of authors — all of them Google employees based in the UK and California — call “multimodal reasoning” but is in fact the chatbot being able to interpret images, ECGs and lab reports in addition to the pre-digested clinical pearl. The topline result, one that the journal itself felt obligated to headline, was that “AI had superior performance compared with physicians for almost every metric (29 of 32 axes)”. But at what?
You would think that the question would have been easy to answer, this being a peer-reviewed paper and all, but no. In fact, I am still not completely certain what interactions were performed and whether they completely match what was reported. What is certain is that a set of primary care physicians and patient-actors from Canada and India — countries different from the author’s own countries and let’s wonder conspiratorially for why that may be the case — interacted via an instant messaging-like service. This is the first oddity: even remote health visits are performed using video calls, and yes you may occasionally get a text through the EMR or if you are a VIP/boutique physician maybe your phone, but that is far from the norm.
The primary report is on what happened when the patients uploaded the skin photos, ECGs, lab results, etc. and then asked the physician or LLM on the other end questions about it. Pretty standard fare for a human-to-LLM interaction, but not exactly natural for a doctor-patient relationship which usually starts with questions being asked of the patient. This is the second way in which the setup was made to fit the computer and not the human.
But then the last section of the paper is about what happens when there is, in fact, a back-and-forth by the way of taking a history. The extended figures — “extended” here meaning not worthy enough of being included in the main paper — say it improves the performance of the LLM. They do not say how it affected the human performance, or how the patient-actors rated humans versus LLMs in history-taking. I would call that strike three.
To the journal’s credit, they did not allow Google to get away with it completely. “To evaluate the performance of our finalized system, we conducted a randomized, blinded human evaluation that emulates an objective structured clinical examination”, says the final paragraph of the introduction, only to end with:
We note, however, that our study is not a randomized clinical trial with prespecified endpoints and preregistered statistical analysis. Rather, it is an exploratory study investigating the properties of multimodal diagnostic dialogue.
Peer review is at least good for something, even if it does result in self-contradiction.
Meanwhile, in the world without motivating reasoning, more objective assessments of the usefulness of AI in medicine show that it is in fact still quite bad. This does not prevent the massively funded hordes of AI researchers from flooding the field with sloppy work, creating the impression that the rise of the machines is imminent. Comply or relegate yourself to the permanent underclass, serf MD. But of course, relegation will only be possible to the extent doctors — or any other profession, really — has already debased itself and abandoned its core professional principles in the service of electronic ease.