Opinion: Using AI in addiction medicine could be particularly risky

OpinionFirst Opinion Using AI in addiction medicine could be particularly risky Encouraging patients in recovery to form emotional attachment to an algorithm could reinforce dangerous patterns Manage alerts for this article Email this article Share this article By Steve D. KleinMay 14, 2026 Steve D. Klein, M.D., Ph.D., is a triple board-certified addiction medicine physician practicing in Pennsylvania at Caron Treatment Centers. Recently, I saw one version of the AI-enabled future of medicine. It was compelling in its simplicity, thrilling in its potential as a force multiplier, and all wrong. The product misunderstood the needs of the patient, the role of the doctor, and the nuanced dance between them. The AI was technically adept, perhaps even marvelous. But it failed at an essential element of a physician’s job: the art of medicine. What troubled me most was the possibility that patients might mistake simulated empathy for genuine human connection. For me, the art of medicine requires the ability to connect authentically with my patients. That connection is bidirectional, intentional, and therapeutic. With AI, we risk losing that connection. This loss would negatively impact subjective and objective outcomes. Advertisement I am a believer in the potential of AI, and specifically the potential of AI in medicine. I use AI daily in my addiction medicine practice to organize my thoughts and improve efficiency. I help to oversee AI implementation at the treatment center where I work. We’ve deployed an ambient listening AI system that is incredibly effective and, judging by the feedback we receive from providers, transformative. With a well-designed system, ambient listening AI moves the electronic medical record taking into the background, allowing clinicians to focus on their patients, not their computers. This is how medicine should be practiced, with the relationship and time between patient and doctor sacred. While we’ve focused on provider-facing tools, we’re open to new ideas. When a company that makes a patient-facing conversational agent contacted me, I took the meeting. Initially, I was extremely excited about the product: an agent that communicates via text message with patients between visits. It summarizes the data into digestible and usable metrics, reflecting the patient’s mental state for providers to review. I imagined how we could use a system that automatically kept in touch with patients and continued to collect usable data on their day-to-day experiences. In addiction medicine, this type of frequent patient contact would be invaluable for informational purposes as well as successful adherence to treatment plans.Advertisement As soon as the demonstration began, my concerns grew. My biggest? The company’s boast that patients form emotional relationships with their AI agents. Physicians need to be very cautious about with whom and why our patients develop relationships. This isn’t paternalistic; it’s reality. Genuine community and connection are paramount in long-term recovery. If connections to a conversational agent replace connections to one’s sponsor, therapist, home group, loved ones, or other relationships, it could be counterproductive to recovery. What if upon discharge — a time when we overwhelmingly recommend building community and recovery capital — a patient focuses on building a relationship with the conversational agent while ignoring one’s local resources, community, and treatment team? Patients have a limited supply of emotional equity. Medical providers need to be judicious about how and where we ask them to invest this precious resource. Doctors also should be mindful of their return on investment, especially as emerging research from experts at Stanford, Brown, and other institutions shows how psychologically risky large language models can be. Humans are wired to respond to perceived empathy and responsiveness, even when we know intellectually that the interaction is artificial. A chatbot that remembers details, responds instantly, and mirrors emotional language can easily create the illusion of understanding or compassion. But these systems do not carry responsibility, accountability, or genuine concern for the patient’s well-being. In addiction treatment, encouraging emotional attachment to an algorithm risks reinforcing the very patterns of avoidance and substitution that sustain addiction. This is the opposite of rebuilding authentic relationships, tolerating discomfort, and learning to rely on other people whom patients in recovery actually need. Beyond the danger of false relationships, the platform presented other concerns as well. The company described its agents as compassionate and empathetic. Can AI have bedside manner? I don’t believe it can. Empathy is the ability to reflect and relate one’s emotional state to another. AI might emulate empathy. It cannot, however, relate and reflect its experience to a human emotion.Advertisement Similarly, while conversational agents can reflect compassionate language, genuine compassion comes from a place of concern. Concern comes from a place of ownership and buy-in. I’m a compassionate provider because I believe that I can help the patient sitting in front of me. I worry about them. I care because of my empathy. Large language models do not share this ability. They cannot worry about their patients. They can only respond with worried language. Finally, the platform refers to its conversational agents either as “M.D.” or “doctor.” These individual agents did not go to medical school. They are not, in fact, doctors. This is not an issue of pride or ego. The title of doctor, M.D., or Ph.D. degree confers certain expectations, as well as societal and interpersonal expectations. Creating a perception that agents have medical training when they do not delivers the wrong message. It’s misleading. These agents have not taken the Hippocratic Oath. They cannot. AI can make great strides in medical decision support, but unfiltered interaction between agent and patient is dangerous. I especially worry about this in sensitive patient populations such as those affected by mental health and addiction issues. The inability of an agent to read nuance and respond accordingly presents major risks. I do believe there’s a place for AI in medicine. There’s a role for agents. There’s too much potential good and efficiency to ignore the idea entirely. This demonstration was not the solution. My addiction-medicine-providing colleagues and I constantly balance principles such as empathy, fidelity, justice, beneficence, non-maleficence, and feasibility. The ideal AI system is one composed of multiple specialized models that must reach consensus, mirroring how real clinicians deliberate complex decisions. There is strong work being done in areas such as text-messaging interventions for youth diabetes (reminders to check blood sugars) and emerging efforts with biometric data collection. Technological advances including app-based health tracking are widely accepted now, too.Advertisement AI’s role in medicine is augmenting, not replacing, human connection. As Hippocrates once said, “wherever the art of medicine is loved, there is also a love of humanity.” In this moment for transformational change, we must remember to maintain and retain that love. For now, I want AI to write my notes, so I have time to practice and love the art of medicine. I don’t want AI to practice my art, so I have time to write my notes. Steve D. Klein, M.D., Ph.D., is a triple board-certified addiction medicine physician practicing in Pennsylvania at Caron Treatment Centers. His focus is on advancement of the field of addiction medicine through technology, novel therapies, and compassionate care. Letter to the editor Have an opinion on this essay? Submit a letter to the editor. addiction, AI Submit a correction requestReprints Steve D. Klein Newsletter Understand how science, health policy, and medicine shape the world every day Recommended First Opinion May 13, 2026 The hantavirus is a wake-up call. Will the Trump administration answer it? First Opinion May 13, 2026 Hospitals are silencing doctors online, and it’s fueling the health misinformation crisis Advertisement First Opinion May 13, 2026 Access to air conditioning is critical preventive care. But federal policy treats it as optional First Opinion Podcast May 13, 2026 What can ‘blue zones’ really teach us about aging? First Opinion May 13, 2026 RFK Jr.’s antidepressant deprescribing push gets one thing right — and others dangerously wrong Subscriber Picks