How Well Can AI Provide Addiction Treatment?
By Tom Horvath, PhD
It is too early to describe precisely what aspects of addiction treatment AI could provide. I expect that it will be useful in providing between-session support, helping clients learn basic information about addictive problems (often called psychoeducation), and conducting screening and diagnosis (assuming the individual is being open and honest). If someone has no access to addiction treatment, then AI might be the backup option, assuming we can establish that the risk of harm with AI treatment is low.
Unfortunately, the risk of harm from an AI therapist is potentially high, because AI generally works to support continued engagement with itself. This can be good in addiction treatment, within limits. With AI, addressing the client’s ambivalence, often a part of the client’s presentation, might not be handled well. The work of treatment is often built on interventions like “you love the high you get from coke, but you don’t like how the comedown keeps you from being the father you want to be.” This kind of intervention is potentially painful for the client, and AI might stay away from it. Additionally, even if only online, addiction treatment usually involves “taking a look” at the client. Someone might interact well with a computer but seeing the individual might provide a much different impression.
Kenneth Anderson, a frequent contributor to this website, stated to me in an email:
Large Language Models (LLMs) are not intelligent, and we need to stop calling them AI–we ought to call them Artificial Idiocy. I have asked 11 questions and received nine wrong answers because AI just returned the first Google result. AI will lead to massive increases in unemployment, homelessness, and human misery and hence massive increases in addiction.
Anderson went on to say that “I asked AI itself why it produced such garbage” and was told:
Al is full of inaccuracies, ‘hallucinations,’ and garbage primarily because it is designed to predict what sounds right rather than what is true. It synthesizes language statistically instead of comprehending facts, meaning it can confidently fabricate plausible-sounding falsehoods if its training data contains errors or biases.
Anderson’s perspective is widely shared in the behavioral health field (as exemplified by the links below). On the other hand, if we compared AI (and its potential problems), and standard US addiction treatment (with its well-established problems), AI might do well. If that result were obtained, I hope it would lead to improving addiction treatment, rather than relying on AI.
Given that AI has been created by “the billionaire class” we should not be surprised if AI does not promote the welfare of the public at large. However, just like addictive problems themselves, we should also not be surprised that some individuals use the “quick fix” of AI treatment rather than seeking genuine addiction treatment.
https://pmc.ncbi.nlm.nih.gov/articles/PMC11572328/
https://hai.stanford.edu/news/exploring-the-dangers-of-ai-in-mental-health-care
