Emanuel and Khosla argue the physician in the loop now degrades AI performance

Set a simulated case study next to a clinic visit. The simulation hands the model a tidy history: onset, duration, medication list, no hedging. In the clinic the patient has to produce that history, and a February 2026 Nature article found that in real-life cases most patients could not converse effectively enough with large language models to reach their expertise, which is the study John Whyte pointed Wired toward ^[12]. The first task on the paper's list of five is taking a medical history ^[4]. Both describe the same appointment, just at different moments in it.

Now the shipped product. Curai Health, the online company run by coauthor Neal Khosla, uses AI to treat patients and keeps a staff of doctors to prescribe medicine and handle the more complicated issues ^[9]. The arrangement is disclosed in the paper, along with Vinod Khosla's relevant investments and Emanuel's grants and consultancies ^[10]. The human step that survives contact with operations is the one holding prescribing authority, a legal boundary rather than a judgment about who reads a chart better.

The horizon deserves arithmetic. The review opens on January 1, 2024, and the newest study in the argument carries a February 2026 date, so the evidence base runs roughly 25 months of publication ^[16]. From that the coauthors forecast that a superior autonomous AI will pass humans using AI by 2030 ^[6], and Emanuel's defense to Wired is that ChatGPT is less than four years old while the prediction is four years out ^[14]. It is worth noting where the date came from. Khosla was telling Emanuel at conferences in the 2010s that AI would be doing 85 percent of what doctors did by 2035 ^[8], so the published number lands five years earlier than the hallway one ^[17].

There is also a tiering question the paper leaves open. Robert Wachter, head of medicine at UCSF, described first-class care as doctors collaborating with AI and "economy class" as mostly AI alone, and the paper's first paragraph names him as getting it wrong ^[15]. For anyone writing a product spec, that is the spec: which patients ever get a person.

So sort every human touchpoint into one of three buckets. The human supplies information the model cannot obtain, such as the physical exam or what the patient actually wants out of treatment. The human re-ranks output the model has already produced. The human holds an authority software is not given, which is where prescribing sits in Curai's own flow ^[9]. Buckets one and three you keep, staff and schedule. Bucket two you instrument: log the override and log what happened to the patient afterward. If clinicians override often and outcomes hold flat, the step is a review ritual, and accept rate on suggestions is not evidence of safety in either direction. Whyte's answer is that these tools belong inside a care plan governed by a physician ^[13]. That is policy, and the product version follows from it: a team that cannot say which bucket a step belongs to is running bucket two and billing it as bucket one.