A recovery chatbot did no better than emailed education in a randomized trial

The short version

Researchers tested Woebot for Substance Use Disorders, a chatbot app, against emailed educational material. Both groups cut back on substance use by about the same amount. The chatbot showed no advantage.

What they did

Adults with problematic substance use were recruited online and randomly assigned to one of two groups. One group used the Woebot chatbot for 8 weeks. The other group received psychoeducation by email. Researchers checked in at the end of the 8 weeks and again one month later.

The trial randomized 258 people. The analysis included 202 of them: 107 in the chatbot group and 95 in the email group. Most named alcohol as their problem substance (73.3%). Others named cannabis (30.7%), stimulants (17.3%), or cocaine (12.4%). Those figures add up to more than 100%, so some people named more than one substance.

The main measure was the change in how many times a person used in the past month, from the start of the study to the end of treatment.

The trial was registered in advance as NCT04925570, ran from November 2022 to August 2023, and was published in the Journal of Substance Use and Addiction Treatment in 2026.

What it found

There was no significant difference between the two groups on the main measure (beta = −0.675, p = 0.741) or on the secondary measures.

Both groups improved. Past-month use occasions fell by 13.6 in the chatbot group and by 12.9 in the email group by the end of treatment, and those drops held at the one-month follow-up.

No serious adverse events were reported.

What it does not show

A null result is not proof the chatbot is useless. It means the chatbot did not beat a simple comparison. And because there was no untreated group, the study cannot say how much of the improvement in either group came from the program rather than from other things, such as the passage of time.

Follow-up lasted one month after treatment. The study says nothing about whether the changes lasted longer.

About one in five people who were randomized are not in the analytic sample. The trial's results on ClinicalTrials.gov list 56 people who did not complete it: 54 for protocol violations and 2 who withdrew, 24 from the chatbot group and 32 from the email group. The registry does not say what the protocol violations were.

The trial was funded by the National Institute on Drug Abuse through a small-business research grant (R44DA048712) awarded to Woebot Labs, Inc. Woebot Health, the company that makes the app, is listed on ClinicalTrials.gov as the trial's lead sponsor, and five of the authors were Woebot Health employees. The study still reported a null result on its main question.

The participants were adults recruited online with problematic use, mostly alcohol. The results may not apply to people in formal treatment programs or to people with more severe disorders.

Why it matters

The comparison group is what makes this trial worth reading. Because the other group got real material by email, the trial could test whether the chatbot added anything beyond basic education. On its main question, it did not.

If you are in recovery and weighing a chatbot like this one, the takeaway is modest. The paper describes Woebot for Substance Use Disorders as an app in development, and this entry does not establish that it is available to the public. The trial reported no serious adverse events, and people in both groups used less. There is no evidence here that the chatbot works better than simpler support.

Clinicians and programs weighing any chatbot can take one habit from this: ask what the tool was compared against. Improvement over time is not the same as improvement over an alternative.