In medicine, we live in a state of managed tension, guided by two companion virtues: non-maleficence (the obligation to “do no harm”) and beneficence (the imperative to “do good”). Psychiatrists know this delicate balance well. We must at times involuntarily hospitalize a patient or administer medication against their will. This stripping of autonomy is justified only through a greater beneficence: the patient’s ultimate return to stability.

This same tension now sits at the heart of the debate over AI. Picture this now-common scenario: Your patient on a six-month waitlist has been using an AI chatbot for ad-hoc therapy. On the surface, it’s working, helping them manage their panic attacks. However, it has also systematically reinforced their medication anxiety, and they’ve abruptly stopped their lithium. When this comes to your attention, what do you do? How about when a cutting-edge model provides well-reasoned counterarguments that directly challenge your recommendations?

The temptation to recoil is strong. It is also, I believe, a mistake. Our primary duty must be beneficence, which requires that we respect patient autonomy and engage with the world they live in. As clinicians, we are fallible, overworked, subject to cognitive biases, and not available 24/7. We should welcome new tools that help us be better. With proper, clinician-led guardrails, embracing AI is not a threat; it’s the most significant advancement for beneficence in decades. It enhances patient autonomy, mitigates burnout, and helps us communicate more empathetically.

Argument 1: Patient Autonomy:

If the internet first challenged the doctor-patient paradigm, AI is the force that will decisively reshape it. Patients are moving from passive recipients to active, informed participants. The era of “Dr. Google” was a chaotic prologue, flooding patients with information that was frequently useless, false, or terrifying. Is this different? Yes, it’s dramatically different: The new generation of AI offers personalized education and triage that is rigorously benchmarked. One frontier model, GPT-5, scored above 95% on average across USMLEs 1-3 (Yang, et al., 2025). This far exceeds the average physician’s scores, and while “hallucinated” inaccuracies are still possible, it demonstrates a level of medical knowledge that is undeniably sophisticated.

The result is a fundamentally different encounter. Bolstered by AI, patients arrive more informed and ready for true shared decision-making. Now, instead of debunking a vague, internet-fueled fear like “will this medication make me a zombie?,” we can immediately engage in a productive discussion about valid concerns, such as akathisia or metabolic side effects. This democratizes medical knowledge, shifting the old paternalistic hierarchy to a collaborative partnership. We may fear that this jeopardizes the therapeutic alliance, but the opposite is true.

We practice in an environment of fractured trust. While trust in the U.S. health care system plummeted from 71.5% to 40.1% post-pandemic (Perlis, et al., 2024), “my provider” remains the single most trusted source for health information (Edelman, 2025). Simultaneously, the majority of physicians agree that misinformation has increased, creating a toxic “fog” that our patients must navigate (The Physicians Foundation, 2023).

In this context, where patients are drowning in misinformation and trust us personally but not the system, a vetted, evidence-based AI is not the enemy; it is our ally. It is a force multiplier for rational thought, corroborating our counsel long after the patient has left the office. The alliance is broken when we act as gatekeepers, dismissing the patient’s own research. It’s strengthened when we partner with them, collaboratively reviewing the information—even AI-generated—within the safety net of our clinical judgment.