AI should take over burdensome tasks, not clinicians, and guardrails must stop decision support systems from guessing.
In this episode of “Talking With Titans,” Christoph U. Correll, MD, of The Zucker Hillside Hospital and the Donald and Barbara Zucker School of Medicine at Hofstra/Northwell, spoke with Psychiatric Times Editor in Chief John J. Miller, MD, about how artificial intelligence (AI) could support psychiatric assessment and treatment selection. Correll framed AI as a tool clinicians must “harness, but also be careful with,” with firm guardrails.
Correll described the most immediate opportunity as task replacement rather than people replacement. He pointed to the repetitive history taking that contributes to clinician burnout as work patients could complete in advance.
Pre-Visit Avatar Interviews
Under the model Correll outlined, patients would complete an avatar-based interview at home before their appointment, covering prior hospitalizations, past medications, and other history. Patients would review the program’s output and make corrections, and the summary could be formatted to match a hospital’s or clinician’s intake form before it reaches the clinician.
Self-report measures collected in the same session could generate T scores indicating whether reported symptoms align more closely with depression, PTSD, or other conditions. Correll said this would let clinicians focus the visit on the chief complaint and on the areas where patients report problems, rather than repeating screening questions in domains such as substance use or trauma when patients have reported none.
These systems could also flag inconsistent responses or symptoms that do not cohere, signaling that a full clinician interview is needed, for example in a patient whose thought disorder may have affected the self-report.
Guardrails Against Hallucination
Correll then described a clinical decision support system (CDSS) linked to the electronic health record that would pull in the full available history, incorporate intake data and ambient speech from the visit, and update its suggested diagnosis and treatment plan in real time. During or after the visit, the clinician could review 2 or 3 suggested options, such as augmentation for a patient with depression.
Correll serves as chief medical officer of MedLink Global, a Mayo Clinic Platform_Accelerate–incubated company developing an AI-based clinical decision support platform called Comentra.1 He said the company’s aim is to demonstrate the system performs better than usual care, which requires showing it does not hallucinate.
To that end, the system is barred from making a recommendation when information is insufficient. Instead, it generates the questions needed to close the gap, such as whether the patient has hepatic problems or a history of suicidality or mania, before offering guideline-based options.
“So I think these systems can help clinicians [get] the whole 360 view of the past, but also what’s needed in the present, and draw on evidence-based treatment,” Correll said. “So we need evidence-based and also measurement-based treatment, which AI can really deliver.”
Correll cited work with the Mayo Clinic involving approximately 40,000 patients, in which patients whose clinicians had made the choices the decision system now suggests had better outcomes. He identified clinician uptake as the open question and said education will be needed so clinicians see the benefit and can select or deselect the system’s options.
Another question under study is whether off-label treatments with supporting clinical data, which an evidence-based CDSS would not typically suggest, yield better outcomes in some patients.
“The AI shouldn’t make the treatment decision ultimately, but it gives you at least 360 view of the data of the patient and also available data in the field, so that you make the most informed decisions,” Correll said.
Toward Precision Psychiatry
Asked whether such systems could compare a new patient against large populations sharing the same diagnosis, Correll said that approach could move the field toward precision psychiatry, but only if phenotypes are measured. Natural language processing of clinician notes is onetandardized 5- to 10-minute avatar interviews before each visit could provide more consistent symptom data
With that data, a system could estimate whether starting a selective serotonin reuptake inhibitor, a serotonin-norepinephrine reuptake inhibitor, or augmentation with an atypical antipsychotic is associated with fewer hospitalizations or less suicidality for a given patient. Correll described outcome modeling of this kind as more of a “black box” than guideline-based recommendations and said it depends on systematic measurement feeding the system.
His emphasis on validation reflects a wider gap in the field. A 2025 review of regulated AI-enabled CDSS tools for mental health care identified 84 products, of which only 7 held FDA or European/UK regulatory clearance, and found only 5 peer-reviewed validation studies across those 7 tools.2 The authors noted regulatory approval “does not guarantee external validity” and called for more external validation and standardized reporting.2
Miller and Correll ultimately agreed to revisit the topic in a year given how quickly the technology is advancing.
Dr Miller is Medical Director, Brain Health, Exeter, New Hampshire; Editor in Chief, Psychiatric Times; Volunteer Consulting Psychiatrist, Seacoast Mental Health Center, Exeter; Consulting Psychiatrist, Insight Meditation Society, Barre, Massachusetts.
Dr Correll is professor at the Institute of Behavioral Science, Feinstein Institutes for Medical Research; medical director of the Recognition and Prevention Program in the Department of Psychiatry at Zucker Hillside Hospital; and professor of Psychiatry and Molecular Medicine at the Donald and Barbara Zucker School of Medicine at Hofstra/Northwell.
1. MedLink Global. About MedLink Global. Accessed October 3, 2026.
2. Kleine AK, Kokje E, Hummelsberger P, et al. Artif Intell Med. 2025;160:103052.
