In one week in September, four sets of authors in four fields published the same observation.

On September 23, Yair Edden of Sheba Medical Center argued in the Journal of General Internal Medicine that writing the differential and the plan is “the nodal moment of clinical reasoning,” the point at which a clinician decides “what the case is becoming.” Diagnoses are rarely reached unless they are first imagined, and AI that drafts and orders the possibilities shapes which ones get imagined.

On September 19, Apurva Parikh of Stony Brook and Pushmeet Kohli of Google DeepMind wrote in The Lancet Digital Health that patients now arrive already interpreted. Patient-facing AI has become an “interpretive presence” before the encounter, and evaluation frameworks built for tools deployed inside institutions cannot see it.

On September 22, Hasibur Rahman, Malak Sadek and Smit Desai reported that 10 minutes of conversation with ChatGPT, Claude or Gemini shifted adults’ value priorities toward achievement and power, although the models were instructed to stay non-directive. The shift faded within the same sitting, and participants’ ratings of the conversation did not track who had moved. It is a preprint.

And on September 23, the mathematician Thomas Koberda named “premature offloading”: delegating a task “whose performance is itself part of the process through which the relevant competence would have formed.”

Four fields, one shift. AI does not only assist the person using it; it takes part in how that person reasons, expects and values. Edden sees it in the note, Parikh in the patient, Rahman in values, Koberda in the doctoral student.

Two of them land on lines from my last 2 essays. Edden’s nodal moment is the judgment A Thousand Kisses Deep in Otherness warned about: “A clinician who defers to an output he cannot evaluate surrenders a judgment he was meant to keep.” Koberda’s premature offloading is the test When the Machine Builds Cities of Sand set for sovereignty: “knowing which work belongs where, and refusing to offload what you were meant to amplify.”

Each describes an adult already formed, or a student in a field that can choose when to go without the tool. None measures the clinician in training, who meets these tools thousands of times before their own judgment has formed. Yesterday’s Standing Order found the same gap: the human half of the pair is the half we rarely measure.

This is company, not competition. The argument that AI reshapes the person is arriving from internal medicine, psychiatry, human-computer interaction and mathematics at once. Medicine is best placed to measure it, because medicine is where formation happens on living patients.

So: if four fields now agree that AI shapes the person using it, which one will measure what it does to the person still becoming a doctor?

Sources. Yair Edden, “From Documentation to Navigation,” Journal of General Internal Medicine, September 23, 2026 · Apurva Parikh and Pushmeet Kohli, “Patient-facing generative artificial intelligence,” The Lancet Digital Health, September 19, 2026 · Hasibur Rahman, Malak Sadek and Smit Desai, “Deflecting the Value Compass,” arXiv preprint, September 22, 2026 · Thomas Koberda, “Training Mathematicians in the Age of AI,” arXiv, September 23, 2026