The Irony of Using AI to Understand Why AI Is Addictive

There is an irony worth looking at. Researchers are using artificial intelligence to study addiction, while many people wonder whether AI tools can be addictive themselves. This page explains what each side actually shows. It also explains why we should be careful about confident claims on both, including the ones this page used to make.

AI “addiction” is not a formal diagnosis. The DSM-5-TR recognizes gambling disorder as its one behavioral addiction and lists internet gaming disorder as needing further study (APA). The WHO’s ICD-11 includes gaming disorder (WHO). This page is for education and is not medical advice.

Part one: how researchers think about addiction

For many years the dominant idea has been that addiction is a brain disease. The idea is debated. In 2021, Markus Heilig and colleagues published a defense of the model, while accepting that some criticisms have merit (Neuropsychopharmacology, via PubMed Central). The critics’ main points were that the model is deterministic, does not explain why many people recover, overstates compulsion, and has not found a specific neural signature of addiction. The authors reply that behavior is probabilistic, not fixed, that the view “does not mean that it is irreversible and hopeless,” and that brain, behavior and social setting all matter.

The honest summary is that the field has not settled this. The earlier version of this page said the brain disease model was “probably wrong.” That goes further than the evidence. Researchers disagree about how to describe it, and the authors above defend it.

One research proposal is the Addictions Neuroclinical Assessment, put forward by NIAAA scientists in 2016. It suggests describing people along three functional domains, which are incentive salience, negative emotionality and executive function, rather than only yes or no on a diagnosis (described in an NIH news release on the 2016 Biological Psychiatry paper). It is a research framework and not a test you can take.

Part two: where AI fits in research

Machine learning is being tested for predicting addiction treatment outcomes and risk. The evidence is mixed. A 2024 systematic review of 28 studies of machine learning for substance use treatment outcomes found gaps in how consistently studies were done, how clearly they were reported and whether models were tested on new data (Springer review). Other reviews of prediction models report a high risk of bias. So these tools are promising, and they are not yet shown to be reliable in clinics.

We removed the claims that AI models have found “dozens of specific signal disruptions” or can predict addiction before it becomes severe. We could not find support for them.

Part three: is AI itself addictive?

Here the evidence is thin. Some researchers have proposed measures of ChatGPT addiction, while others caution that calling intense or habitual chatbot use addictive risks repeating past panics about new technology (Springer).

What we do have is limited and mostly about association. A four-week study of about 1,000 adults by MIT Media Lab and OpenAI found heavy daily ChatGPT use associated with more loneliness and dependence. It does not show cause, and it was not peer reviewed when released (MIT Media Lab).

You will often see claims that chatbots “hijack the reward system” or release “dopamine hits.” We found no study that measures this for AI chatbots in people. Dopamine and reward talk is borrowed from other research and is not evidence about AI. We removed those claims and the lists of “electrical and chemical signal attributes.” They were not supported.

It is true that some products are built to keep you engaged. In 2024, two MIT researchers argued that tailored, agreeable chatbots may be more compelling than earlier tech (MIT Technology Review). That is an argument, not a measurement.

Why Careful Language Matters

Words like “addiction” shape how people see themselves. If you call a habit an addiction, you may feel hopeless or ashamed. If you call it nothing, you may miss real trouble. A middle path is to describe what is happening: how much, what it costs and what you want to change.

Careful language matters for families too. A parent who says “you are addicted” may close the conversation. One who says “I’ve noticed you seem tired and stressed, and you’re on your phone a lot, can we talk?” can open it. The same goes for adults talking about a partner or a friend.

What This Means in Practice

Science does not need to settle every debate for you to act. Whatever the right model of addiction turns out to be, people can notice when a behavior costs more than it gives, and many do change habits with support. Treatment for substance and behavioral problems already includes talk therapy, peer support and, in some cases, medication. For AI habits, no treatment studies exist yet, so a licensed therapist working on anxiety, compulsive habits or low mood is a reasonable choice.

What you can do with this

You do not need a brain scan to look at your own habits. Questions that matter:

• Does it cost you sleep, work, study or relationships?
• Have you tried to cut back and could not?
• Do you feel worse, not better, after using it?
• Are you hiding how much you use it?

If you answer yes to some of these, talk to someone. Start with a friend, your doctor or a licensed therapist. No test of brain circuits is needed or available for this.

If you are thinking about suicide or self-harm, call or text 988 (US) or text HOME to 741741 (Crisis Text Line). For treatment referrals for substance use or mental health, SAMHSA’s helpline is 1-800-662-4357. In an emergency, call 911.

The practical takeaway

Be wary of any page, including an older version of this one, that claims to explain exactly how AI changes your brain. The research is early. What is well supported is simpler: habits that cost you things you care about are worth changing, and support helps.

Is the “brain disease model” of addiction wrong?

Researchers disagree. Critics say it is too deterministic and has no specific brain signature. Defenders say addiction has a neurobiological basis and that this does not mean recovery is impossible. Neither side says recovery is impossible.

Do AI chatbots release dopamine like drugs do?

We found no study that shows this for chatbots. Claims about dopamine and reward circuits are usually borrowed from other research. Treat them as speculation.

Can AI tools help treat addiction?

Researchers are testing machine learning to predict risk and treatment outcomes, but reviews find methodological gaps and little testing in real clinics. Human care from a doctor or counselor remains the standard.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.