Chatbots and AI companions in recovery

What studies show, and where they have failed on safety.

1. Study

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

   Source date April 24, 2026 · Randomized trial, registered in advance (NCT04925570) · 202 analyzed of 258 randomized (107 chatbot, 95 email)

   In a randomized trial (202 adults analyzed), people using the Woebot substance use chatbot cut back about as much as people sent education by email.

2. Study

   ChatGPT-4 gave sound answers to common alcohol questions, but rarely pointed people to help

   Source date August 14, 2024 · Content evaluation. Two coders rated every answer against National Institute on Alcohol Abuse and Alcoholism (NIAAA) resources. Not pre-registered; the authors call the results exploratory. · 64 questions, each asked once in a new chat

   ChatGPT-4 answered 64 common questions about alcohol use disorder. 59 answers were fully evidence-based, but only 8 pointed people to outside help.

3. Study

   Chatbots answered real recovery questions well overall, and some answers were dangerous

   Source date June 26, 2024 · Evaluation of chatbot answers. Clinicians rated them without being told the answers came from AI. · 75 questions, 150 answers, rated by 7 clinicians

   Clinicians rated ChatGPT and LLaMA-2 answers to recovery-forum questions as high quality, but some gave wrong helplines or endorsed home detox.