AI in Addiction Recovery Apps

Addiction-recovery apps range from journaling and reminders to structured digital therapeutics and AI-enabled assistants. A tool may prompt check-ins, offer evidence-informed exercises, support appointment adherence, or identify a change in self-reported craving. Sensor-based models may explore patterns in sleep, activity, or device use, but these signals are indirect and can reflect many nonclinical factors. Prediction quality and use matter. A false alert can cause distress or stigma, while a missed alert may create false reassurance. A model trained on one group or treatment setting may not transfer to another. Craving and relapse are complex, and a probability score cannot establish that a person has used a substance. Recovery support should include a way to contact human providers or peer supports and a clear response plan. Evidence is specific to the app and study. For example, one observational mixed-methods pilot study used a wrist sensor and self-reports during outpatient SUD treatment to explore stress and craving signals; this design can assess feasibility and associations but does not establish that an AI app prevents relapse. Separately, the FDA's De Novo summary for reSET describes a prescription-only digital therapeutic delivering CBT content as an adjunct under clinician supervision. That example is not an AI chatbot and illustrates that intended use and care context should be explicit. People should not rely on an app during an overdose, medical emergency, or immediate danger. Contact local emergency services or qualified treatment support. Apps should avoid overpromising confidentiality or clinical effectiveness, explain data retention and sharing, and protect sensitive records. Use AI as an optional support layer within a broader recovery plan. Involve clinicians and the person using the tool in deciding whether notifications are useful. Evaluate outcomes, harms, accessibility, and engagement—not just prediction accuracy. Maintain human oversight and a safe route to help when an alert is urgent or uncertain.