AI can assist addiction therapy, but the therapeutic relationship is irreplaceable. What therapists and clients need to know about AI sycophancy and the limits of chatbot care.
A version of this article originally appeared in Dr. Arnold Washton’s column on Psychology Today.
Key Points
- AI tools can supplement CBT-based addiction care, but cannot replace human interpersonal psychotherapy.
- Denial and ambivalence require skilled clinical truth-telling, not AI validation.
- AI sycophancy, trained to please and affirm, is dangerous in the face of minimization and denial.
- The therapeutic relationship remains the irreplaceable vehicle and driver of meaningful change.
Artificial intelligence (AI) has entered virtually every sector of healthcare, and addiction treatment is no exception. Chatbots, conversational agents, and AI-powered coaching apps are now marketed as tools, or even substitutes, for traditional psychotherapy in the treatment of alcohol and other substance use disorders (SUDs). Proponents argue that AI can expand access to evidence-based care, reduce stigma, and provide around-the-clock support. Critics warn that these promises obscure serious limitations and genuine clinical dangers. Whether AI can meaningfully supplement, let alone replace, human psychotherapy for addiction (and other behavioral health problems) remains unclear and requires careful scrutiny.
The Case for AI as a Supplement to Addiction Treatment
There are legitimate reasons to explore AI as a supplemental resource. Fewer than 10% of the estimated 46 million Americans with a substance use disorder received any treatment in 2023 (SAMHSA, 2024). Geographic, financial, and social barriers, including stigma, prevent millions from ever entering a therapist’s office. AI tools that deliver psychoeducation, screen for risk, encourage help-seeking, and provide between-session support can serve a meaningful adjunctive role.
Perhaps the most defensible application is AI-assisted delivery of cognitive behavioral therapy (CBT). Because CBT is highly structured and emphasizes psychoeducation, cognitive restructuring, and behavioral skill-building, its core components are more amenable to algorithmic delivery than the nuanced, relationally complex work of interpersonal or psychodynamic therapies. Several digital CBT programs have demonstrated modest efficacy in randomized controlled trials as supplements to standard care (Carroll et al., 2014). AI can also provide relapse-prevention prompts, help users track craving patterns, and offer immediate coping strategies at high-risk moments, extending the therapeutic hour into daily life.
The Fundamental Limits of AI as a Therapist
The notion that AI can replace psychotherapy rests on a profound misunderstanding of what therapy actually is. Therapy is not advice-giving, not the generation of to-do lists, not the dispensing of directives such as “stop drinking and start exercising” or “avoid triggers but maintain social connections.” Framing addiction treatment as a problem of insufficient behavioral instruction fundamentally misrepresents the clinical reality. Even CBT, the modality most amenable to algorithmic delivery, is most effective within a strong therapeutic relationship (Norcross & Wampold, 2011).
Interpersonal, psychodynamic, and motivational enhancement therapies place the therapeutic relationship at the very center of the change process. Addiction is not merely a set of maladaptive behaviors to be corrected; it is an expression of deeper psychological vulnerabilities, including self-esteem deficits, unresolved trauma, attachment ruptures, and shame, that require a sustained, attuned, trusting relationship between patient and clinician to address effectively (Washton & Zweben, 2023). No current AI system is capable of forming or therapeutically utilizing a genuine interpersonal relationship. It can simulate warmth, but it cannot provide it.
Denial, Ambivalence, and the Sycophancy Problem
Perhaps the most clinically significant limitation of AI in addiction treatment concerns denial and ambivalence, core features of virtually all SUDs. Addiction is uniquely characterized by a patient’s resistance to acknowledging the severity or consequences of the problem, being of “two minds” about change, or disputing that a problem exists at all. Minimization, rationalization, and externalization of blame are not merely obstacles to treatment; they are the treatment problem itself. Motivational interviewing (MI), the evidence-based framework specifically designed to address ambivalence and resistance, depends on the clinician’s ability to read subtle interpersonal cues, roll with resistance, develop discrepancy, and respond with calibrated empathy within an established therapeutic alliance (Miller & Rollnick, 2023).
Current AI systems exhibit a well-documented tendency toward “sycophancy,” a systematic bias toward responses that users find agreeable and emotionally comfortable, even when clinical accuracy demands otherwise (Malmqvist et al., 2025). This tendency is not a superficial design flaw, but an artifact of AI training processes that reward responses consistently rated by human evaluators as affirming and preferable. In addiction treatment, this tendency is potentially dangerous. A patient who insists their drinking is “not really a problem” is likely to receive a gentle, validating AI response that neither challenges the distortion nor advances the change process. A skilled clinician, by contrast, uses that moment, together with the full context of the patient’s history, present affect, and the therapeutic relationship itself, to compassionately but directly address the discrepancy between stated values and current behavior. This is the art of therapeutic truth-telling, delivered with warmth within a relationship strong enough to hold the patient’s discomfort. AI, trained to avoid discomfort and optimize for user approval, is structurally ill-suited for it.
Safety, Ethical, and Clinical Risks
Significant safety concerns attend the use of AI in addiction treatment. Alcohol and certain drug withdrawals can be medically life-threatening; AI systems cannot conduct clinical assessments, detect imminent risk, or coordinate medical interventions. Comorbid psychiatric conditions, highly prevalent in SUDs, require differential diagnosis and treatment planning beyond the scope of any current AI. Confidentiality standards, mandatory reporting obligations, and liability frameworks governing licensed clinicians do not apply uniformly to commercial AI applications (NIDA, 2023). Patients who engage with AI as a primary “treatment” risk delaying or forgoing evidence-based care, potentially at serious cost to their health and well-being.
A Constructive Path Forward
The most defensible role for AI in addiction treatment is as an adjunct to human care, not a replacement for it. AI tools may appropriately extend the reach of psychoeducation, support between-session skill practice, facilitate symptom monitoring, and reduce barriers to entering treatment. Used judiciously and with appropriate clinical oversight, they may enhance care for patients already engaged with a qualified provider. What they cannot do, and should not be allowed to obscure, is the irreducibly human work of standing with another person in their struggle, earning their trust, tolerating their resistance, and helping them find the courage to face what they have been working so hard not to see. That remains the province of the skilled clinician, and no algorithm is close to replicating it.
Key Terms in This Article
- Substance use disorder (SUD): A medical condition in which recurring alcohol or drug use causes significant impairment or distress. Fewer than 10% of the estimated 46 million Americans with an SUD received treatment in 2023.
- Cognitive behavioral therapy (CBT): A structured, skills-based therapy that targets the thoughts and behaviors driving substance use. Its structure makes it the approach most adaptable to digital or AI-assisted delivery.
- Motivational interviewing (MI): An evidence-based, collaborative method for resolving the ambivalence and resistance that are central to addiction, built on the clinician’s read of subtle interpersonal cues.
- Therapeutic alliance: The trusting, collaborative relationship between patient and clinician. Across therapies, it is one of the strongest predictors of whether treatment works.
- AI sycophancy: A documented tendency of AI systems to give agreeable, affirming answers that users find comfortable, even when accuracy calls for challenge, an artifact of how the models are trained.
- Co-occurring (comorbid) disorders: Psychiatric conditions that occur alongside a substance use disorder. They are highly prevalent and require professional differential diagnosis and treatment planning.
References
Carroll, K. M. (1998). Therapy manuals for drug addiction, Manual 1: A cognitive-behavioral approach: Treating cocaine addiction (NIH Publication No. 98-4308). National Institute on Drug Abuse.
Carroll, K. M., Kiluk, B. D., Nich, C., Gordon, M. A., Portnoy, G. A., Marino, D. R., & Ball, S. A. (2014). Computer-assisted delivery of cognitive-behavioral therapy: Efficacy and durability of CBT4CBT among cocaine-dependent individuals maintained on methadone. American Journal of Psychiatry, 171(4), 436-444.
Malmqvist, L. (2025). Sycophancy in large language models: Causes and mitigations. In K. Arai (Ed.), Intelligent Computing: Proceedings of the 2025 Computing Conference (Lecture Notes in Networks and Systems, Vol. 1426, pp. 61-74). Springer.
Miller, W. R., & Rollnick, S. (2023). Motivational interviewing: Helping people change and grow (4th ed.). Guilford Press.
National Institute on Drug Abuse. (2023). Principles of drug addiction treatment: A research-based guide (3rd ed.). National Institutes of Health.
Norcross, J. C., & Wampold, B. E. (2011). Evidence-based therapy relationships: Research conclusions and clinical practices. Psychotherapy, 48(1), 98-102.
Substance Abuse and Mental Health Services Administration. (2024). Key substance use and mental health indicators in the United States: Results from the 2023 National Survey on Drug Use and Health (HHS Publication No. PEP24-07-021). Center for Behavioral Health Statistics and Quality, SAMHSA.
Washton, A. M., & Zweben, J. E. (2023). Treating alcohol and drug problems in psychotherapy practice: Doing what works (2nd ed.). Guilford Press.