In 1966, the Advanced Research Projects Agency of the United States Department of Defense explored how remote computers could share resources. The scientists working on that project explored this technology as an ecosystem for sharing scientific data and never imagined it being commercialized. Sixty years later, the internet has radically transformed the way we live, conduct business, and practice medicine. As a psychiatry resident approaching the transition to attending hood around the advent of AI, it is hard not to draw parallels.

Just in the past year, AI has begun to enter clinical workflows at an accelerating rate. Clinical decision support platforms such as OpenEvidence have partnered with major medical journals and are increasingly used at the point of care, while AI scribes are writing notes in more and more clinics. Meanwhile, I keep reading headlines about AI solving mathematics problems that I couldn’t even begin to understand that were previously considered unsolvable. As a young psychiatrist preparing to launch my career in this uncertain and rapidly changing world, I often wonder how replaceable I will be. AI models certainly “know” more psychoanalytic theory than I do, they can challenge automatic thought processes using CBT techniques quite effectively, and can rattle off clinical trials, dosing regimens, and binding affinities in seconds. AI never gets hangry, tired, or wonders what they’re going to eat for dinner that day. And this is the least capable it will ever be, and it’s getting better every day.

However, the further I get in my training, the more hopeful I feel. I have come to believe that the primary treatment psychiatry offers to most patients is relational. Decades of psychotherapy research point to the quality of the alliance between therapist and patient as one of the most consistent predictors of how well patients do, regardless of the type of therapy delivered. That would come as no surprise to most therapists, but the effect extends beyond the couch to the prescription bottle. One mistake I have made is prescribing a medication before the patient was truly on board, only to watch them develop unexplainable, strange side effects and quit the medication quickly. The research backs up what those moments have taught me. When researchers reanalyzed a large NIMH depression trial, they found that the most effective psychiatrists prescribing a placebo had better results than the least effective psychiatrists prescribing an antidepressant.¹

On a personal level, I’m not a perfect human or psychiatrist. The quality of my patients’ relationship with me fluctuates. Sometimes patients and I aren’t aligned on our goals, sometimes I have to enforce professional boundaries, and sometimes I just straight up say something foolish. These moments of rupture offer the opportunity for repair, which allows an opportunity to strengthen the relationship. In contrast, AI is endlessly patient and agreeable and is frictionless, which provides no opportunities for genuine, interpersonal repair.

I don’t know what AI will be able to do in ten years, but it will probably write my notes and catch medication interactions I might miss. Last month, I spent nearly an entire visit looking through wedding photos with a patient who was stable and doing well, which I felt improved our therapeutic relationship. I feel so lucky that this is my job. It’s hard to imagine a chatbot doing that in a meaningful manner. A large part of why it mattered was that I was a real person who had chosen to be interested.