AI Should Replace Tasks, Not Psychiatrists, Expert Warns

Christoph Correll sits at the intersection of two pressures that define modern psychiatry: the exhaustion of clinicians drowning in paperwork, and the promise of artificial intelligence to lighten that load. His answer to both is precise and bounded. AI should take the repetitive work off clinicians' desks—the intake forms, the history-taking, the screening questionnaires that eat up appointment time and fuel burnout. It should not, under any circumstances, make the clinical decision itself.

Correll, a psychiatrist at Zucker Hillside Hospital and the Zucker School of Medicine at Hofstra/Northwell, laid out this vision in conversation with John J. Miller, the editor in chief of Psychiatric Times. The model he described begins before the patient ever sits down with a doctor. A patient would complete an avatar-based interview at home, walking through prior hospitalizations, past medications, and other history. The system would generate a summary, the patient would review it and correct errors, and the output would arrive at the clinician's desk already formatted to match the hospital's intake structure. Self-report measures collected in the same session could calculate T scores, flagging whether a patient's symptoms cluster more closely with depression, PTSD, or another condition. The clinician then walks into the room with a full picture and can focus the visit on what actually matters—the chief complaint and the areas where the patient reports problems—rather than repeating screening questions about substance use or trauma when the patient has already said none occurred.

The system has a built-in safeguard: it flags inconsistent responses or symptoms that don't cohere, signaling that a full clinician interview is needed. A patient whose thought disorder may have colored their self-report would trigger this alert. The goal is to catch what an algorithm cannot see.

From there, Correll described a clinical decision support system linked to the electronic health record. It would pull in the full available history, incorporate the intake data and ambient speech from the visit itself, and update its suggested diagnosis and treatment plan in real time. A clinician could review two or three suggested options—say, augmentation strategies for a patient with depression—and choose or reject them. But here is where Correll draws the hard line: the system is barred from making a recommendation when information is insufficient. Instead of guessing, it generates the questions needed to close the gap. Does the patient have hepatic problems? A history of suicidality or mania? Only after those gaps are filled does the system offer guideline-based options.

Correll serves as chief medical officer of MedLink Global, a company developing an AI platform called Comentra. The company's central aim is to demonstrate that the system performs better than usual care—which, he emphasized, requires showing it does not hallucinate. Work with the Mayo Clinic involving approximately 40,000 patients found that when clinicians made the choices the decision system now suggests, outcomes improved. But clinician uptake remains an open question. Doctors need to see the benefit and understand that they retain the power to select or deselect the system's options. Education will be essential.

Correll also raised a harder question: whether off-label treatments with supporting clinical data—treatments an evidence-based system would not typically suggest—yield better outcomes in some patients. That question points toward precision psychiatry, where a system could compare a new patient against large populations sharing the same diagnosis and estimate whether starting an SSRI, an SNRI, or augmentation with an atypical antipsychotic is associated with fewer hospitalizations or less suicidality for that particular person. But that kind of outcome modeling is more of a black box than guideline-based recommendations, and it depends entirely on systematic measurement feeding the system.

The field is nowhere near ready for that. A 2025 review identified 84 AI-enabled clinical decision support tools for mental health care. Of those, only 7 held FDA or European and UK regulatory clearance. Across those 7 tools, only 5 peer-reviewed validation studies existed. Regulatory approval does not guarantee that a system will work the same way in a different hospital, with a different population, using different data. The authors of the review called for more external validation and standardized reporting. Correll and Miller agreed to revisit the topic in a year, given how quickly the technology is advancing. By then, the field may have moved closer to clarity—or further into the fog.