AI in substance use disorder treatment is most promising for predicting high-risk moments and prompting support between visits. Research models can forecast next-day alcohol lapses with high accuracy, and smartphone recovery support has reduced risky drinking in a randomized trial. The gap is response: prediction helps only when a person or service can act quickly.
Addiction does not keep office hours. Cravings spike on Friday nights, after arguments, on paydays and anniversaries. A clinic open nine to five, with a counselor who sees a patient every other week, is structurally absent at the moments when recovery is most likely to break.
That mismatch is the best argument for AI in substance use treatment, and also the most common way the argument goes wrong. The technology to detect rising risk between visits is getting good. The services to respond to that risk, especially at night, mostly do not exist. The thesis here is that relapse prediction is valuable only if someone can respond at 2 a.m., and that the smartest investments in this field will be in the response layer rather than in ever-better models.
If you or someone you know is in crisis, please contact your local emergency services or a crisis line. In the US you can call or text 988, and SAMHSA’s National Helpline at 1-800-662-4357 offers free, confidential treatment referral around the clock.
Can AI predict relapse in addiction recovery?
Increasingly, yes, at least in research settings.
In 2024, Kendra Wyant, John Curtin and colleagues at the University of Wisconsin-Madison published a study in the Journal of Psychopathology and Clinical Science that used machine learning to predict lapses back to drinking among 151 adults in early recovery from alcohol use disorder. Participants completed brief smartphone check-ins four times a day for up to three months, reporting things like craving, stress, recent events and confidence in staying sober. The models predicted lapses in the next week, next day and next hour with median areas under the curve of 0.89, 0.90 and 0.93. Factors such as past use and self-efficacy mattered across time frames, while craving, stressful events and arousal mattered most in the next-hour model.
Those are strong numbers, with caveats. The sample was small and mostly white, participants were motivated volunteers who kept answering prompts, and accuracy in a research cohort is not accuracy in a busy clinic. Similar work is under way in opioid use disorder, where smartphone surveys paired with deep learning have shown early promise in forecasting both relapse risk and treatment dropout.
The direction of travel is clear enough. Short, frequent self-reports, possibly combined with passive phone and wearable data, can identify periods of rising risk days or hours before a lapse. That is a genuine clinical opportunity, because relapse prevention works best early.
Do addiction recovery apps actually work?
The best-known trial predates today’s AI but tells us what a good response layer looks like.
In 2014, David Gustafson and colleagues published a randomized trial in JAMA Psychiatry of A-CHESS, a smartphone recovery support system for people leaving residential treatment for alcohol dependence. The 349 participants received usual continuing care, with or without eight months of A-CHESS. Over the intervention and four months of follow-up, the A-CHESS group reported fewer risky drinking days, 1.39 versus 2.75 on average, and higher rates of abstinence at eight and 12 months.
A-CHESS was not primarily a prediction engine. It combined peer support, recovery content, check-ins, a panic button that connected users to support people, and alerts to counselors. Its success is a hint that connection, not computation, did most of the work.
The commercial story is more sobering. Pear Therapeutics won FDA authorization for reSET, a prescription digital therapeutic for substance use disorder, and reSET-O for opioid use disorder. They were landmark products. Pear filed for bankruptcy in April 2023, having struggled to get payers to reimburse, and its assets were sold at auction for about $6 million. The problem, as we argued in Digital Therapeutics Don’t Have a Reimbursement Problem. They Have a Buyer Problem., was less the evidence than the absence of anyone whose budget benefited enough to pay.
Why prediction without response fails
Imagine a model that is 90% accurate at forecasting next-day lapses. What happens when it fires?
In most real programs, the alert goes to a dashboard that a counselor checks during working hours. If it fires on Saturday afternoon for a Saturday evening risk, nothing happens until Monday. The patient may get an automated message suggesting a breathing exercise. That is not nothing, but it is a thin response to a strong signal. The Wisconsin models were most accurate at the shortest horizon, the next hour, which is precisely the window a weekday dashboard cannot reach.
This is the same trap we described in Remote Patient Monitoring Pays for Data. Patients Benefit From Responses. Health systems buy sensing because it is a product. They do not buy response because it is a service, with staff, rotas and night shifts.
There is a second trap: alert fatigue for patients. A person in early recovery who receives frequent “you seem at risk” messages may find them intrusive, stigmatizing or simply ignorable. The best-designed interventions, often called just-in-time adaptive interventions, vary the type and timing of support based on context rather than simply announcing risk.
And there is a stakes question. Substance use is not only about quality of life. US overdose deaths fell by nearly 27% in 2024 to an estimated 80,391, according to provisional CDC data released in May 2025, the lowest level since 2019. That is welcome progress, but it still means tens of thousands of deaths a year, many of them in people who had been in treatment. Missing the moment carries real consequences.
A worked example: what a 2 a.m. response could look like
Here is a hypothetical program design for an outpatient clinic treating opioid use disorder with medication.
Patients opt in to a brief smartphone check-in three times a day. A model combines responses with missed doses and appointment data to estimate next-day risk. So far, this is what many pilots do.
The difference is the response tiers. At moderate risk, the app offers a scheduled call from a peer recovery specialist within four hours, during waking hours. At high risk overnight, the patient is offered an immediate connection to a staffed recovery or crisis line with access to their care plan, and the clinic’s on-call clinician is notified if the patient consents. Every high-risk flag is reviewed at the next morning’s team huddle, and anyone who has missed doses gets same-day outreach.
The model might cost little to run. The response layer requires contracted overnight peer support and clinician on-call time, which is where the budget goes. That budget is the intervention. The algorithm simply decides where to point it.
A program like this should be evaluated on the outcomes that matter: retention on medication, emergency visits, overdoses and patient experience. Not on how accurately the model forecast lapses.
The strongest case against AI in addiction care
The strongest objection is not that the models are weak. It is that surveillance and addiction treatment have a painful history together.
People with substance use disorders have long faced punitive monitoring through drug tests, probation, child welfare investigations and employment consequences. A system that continuously infers when someone is likely to use drugs creates data that could, in the wrong hands, be used against them. In the US, substance use treatment records carry special confidentiality protections under federal rules known as 42 CFR Part 2, precisely because disclosure can cause harm. Consumer apps often sit outside those protections.
There is also a risk of reinforcing a deficit model, in which recovery is framed as a series of predicted failures to be managed, rather than as a process of building a life.
These objections are valid and should shape design. Data should be patient-controlled, used only for support, and never shared with courts, employers or insurers without explicit consent. Patients should be able to see and challenge what the model infers. Programs built this way can be supportive rather than punitive. Programs that skip these steps will lose the trust they depend on, and with it the self-reports that make prediction possible at all.
What this means for clinics, payers and founders
For treatment providers, the priority is to design the response first. Decide what you can deliver at night and on weekends, and with what staff, before choosing a model. Peer recovery specialists are an underused workforce for exactly this role.
For payers and health systems, the business case sits in reduced emergency visits, hospitalizations and treatment dropout. Pay for the response service, not only the software licence, and require evaluation on retention and acute care use.
For founders, the lesson of Pear is that FDA authorization does not create a buyer. Products that bundle prediction with a staffed response, sold to organizations that bear the cost of relapse, have a clearer path than apps sold on accuracy.
For people in recovery and their families, the useful question about any app is simple: when it notices I am struggling, who does it connect me to, and when?
Frequently asked questions
Can AI help with addiction recovery?
It can, mainly by spotting periods of rising risk between appointments and prompting timely support. A University of Wisconsin study published in 2024 predicted next-day alcohol lapses with high accuracy from smartphone check-ins. The benefit depends on what happens after a prediction, such as contact with a peer supporter, counselor or crisis service.
What is the best app for addiction recovery?
There is no single best app, and evidence varies widely. A-CHESS, tested in a 2014 JAMA Psychiatry randomized trial, reduced risky drinking days among people leaving residential treatment. Look for apps with published trial evidence, connection to real people, clear privacy protections and integration with your treatment team. Discuss options with your clinician.
Are addiction treatment apps FDA approved?
Some have been. Pear Therapeutics’ reSET and reSET-O received FDA authorization as prescription digital therapeutics for substance use and opioid use disorders. Pear went bankrupt in 2023, and its products were sold to other companies. Most recovery apps on the market are not FDA-authorized and are not required to be.
Is my data safe in an addiction recovery app?
Not necessarily. US federal rules under 42 CFR Part 2 strictly protect records from many substance use treatment programs, but consumer apps are often not covered by those rules or by HIPAA. Check what the app collects, whether it shares data with advertisers or third parties, and whether you can delete your information.
What should I do if I relapse?
A lapse is common in recovery and is not a failure. Contact your treatment provider, a peer supporter or a helpline as soon as you can. In the US, SAMHSA’s National Helpline at 1-800-662-4357 is free and available 24 hours a day. If there is an immediate risk to life, contact emergency services.
The takeaway
The addiction field does not need a better crystal ball as much as it needs someone to pick up the phone. Fund the 2 a.m. response, and even a modest model becomes useful. Skip it, and the most accurate prediction in the world becomes a record of what nobody did.
Who answers when your program’s risk alert fires at night? We would like to hear how services are solving this.
