Is AI an Addiction? I Ran the Actual Test

Key Takeaways

• I predicted AI use would fail the clinical test for addiction. It does, but for a reason I didn’t expect.
• It scores loudest on salience and tolerance, and those are precisely the two criteria researchers have shown don’t track disorder at all.
• In a study of 4,256 people, those two components had no association with any psychopathological symptom. They measure engagement, not addiction.
• One pattern does clear the bar, and it isn’t the productivity one: companionship-shaped use, where the RCT evidence is genuinely bad.
• What’s left over after the test isn’t withdrawal. It’s atrophy, and the addiction framework has no slot for a harm that only shows up once you stop.

I opened it before I knew what I wanted to ask.

That’s the moment worth examining. Not the hours logged, not the tab that’s always there. The reach that precedes the question. Something in the sequence had inverted: the tool arrived before the thought it was supposed to serve.

The word people reach for when they notice this in themselves is addiction. I’ve used it. It’s in half the essays written about AI right now, usually as a rhetorical shrug, the same borrowed vocabulary that turns ordinary unease about AI into a diagnosis. But it isn’t a metaphor. It’s a clinical term with an operational definition, a measurement literature, and a bar that was set long before any of this existed.

So I thought I’d stop using it loosely and actually run the test.

My Prediction

Before the evidence: I expected this to come back dependence, not addiction. I expected the distinction to be real and the word to be borrowed, the way people say they’re “addicted to” running or spreadsheets, importing medical gravity into something that doesn’t earn it.

That’s roughly where it landed. But the interesting part turned out to be which criteria it passed, and what happened when I checked whether those criteria mean anything.

What the Word Actually Means

Behavioural addiction has a much higher bar than most people assume.

The DSM-5-TR recognises exactly one non-substance addictive disorder: gambling. Internet gaming disorder sits in Section III, the holding pen for conditions that need further study. That’s it. Decades of proposals, covering shopping, work, exercise, sex, social media and tanning, and the manual has admitted one. The bar is high on purpose, because the alternative is medicalising every intense human interest.

The framework most of those proposals are built on is Mark Griffiths’ components model (2005), which holds that all addictions display six features:

1. Salience: the behaviour dominates thinking and feeling.
2. Tolerance: you need more of it to get the same effect.
3. Mood modification: you do it to change how you feel.
4. Relapse: you try to cut down and revert.
5. Withdrawal: stopping produces genuine distress.
6. Conflict: it damages your relationships, work, or self.

Nearly every “are you addicted to X” scale in circulation descends from these six. So they’re the right thing to test against.

The Test

Scored honestly, against my own use:

Salience: yes. It’s in the background of most working thought. When I hit a wall, the reach is reflexive, and often precedes knowing what I’d ask. Clear pass.

Tolerance: yes. A year ago it was a better search engine. Now it drafts, argues, reviews, and runs multi-step work. The scope has escalated steadily, and going back to the old level would feel like a downgrade rather than a relief. Clear pass.

Mood modification: partly. There’s something real here: reaching for it when facing a blank page is partly about not sitting in the discomfort of not-yet-knowing. But the honest version is narrower than it sounds. It isn’t relief-seeking so much as friction-avoidance, and the thing being avoided is a work feeling, not a life feeling. Half a pass at most.

Relapse: no. Cutting back has never felt like willpower spent. When I’ve stopped for a stretch, resuming was a decision, not a collapse.

Withdrawal: no. Here’s where the frame starts to buckle. Without it I’m slower and the work is worse. But I’m not distressed. There’s no restlessness, no irritability, none of the aversive signature the criterion is describing. Something is missing, and it isn’t equanimity.

Conflict: no. Nothing has been damaged. No relationship strained, no work broken, nothing valued surrendered. This is the criterion that separates an intense interest from a disorder, and it’s a clean fail.

Two clear passes out of six, one partial. On most published scales, that lands somewhere unremarkable.

The Two That Don’t Count

Now the part I didn’t see coming.

The two criteria I passed loudest are the two the field has been quietly dismantling.

In 2023, Loïs Fournier and colleagues published Deconstructing the components model of addiction in Addictive Behaviors. Across 4,256 participants, they tested whether the six components actually cohere into one thing. They don’t. The model splits cleanly in two: salience and tolerance on one side, mood modification, relapse, withdrawal and conflict on the other. The one-factor model fit badly (RMSEA = 0.137); the split fit well (CFI = 0.986, RMSEA = 0.073).

Then the finding that matters. They ran network analyses linking each component to a battery of psychopathological measures: depression, anxiety, stress, social anxiety, and the DSM-5 cross-cutting symptom measure. The four core components were positively associated with twelve symptom scores. Salience and tolerance showed no association with any of them. Not weak. Null edges, in all four samples.

A 2025 replication in the Journal of Behavioral Addictions, Ciudad-Fernández and colleagues, 2,761 adolescents, went further. Salience and tolerance weren’t merely unassociated with distress. They were negatively associated with depression, anxiety and loneliness, and positively associated with life satisfaction and self-esteem. The authors’ conclusion is blunt: these items measure “engagement in” rather than “addiction to,” and folding them into addiction scales “contributes to overdiagnosis and pathologization of healthy intensive involvement.”

So the scorecard reads differently than it did a moment ago. The two criteria my AI use passes emphatically are the two that, in two large samples, predict better mental health rather than worse. What I was calling evidence of a problem is the psychometric signature of caring a lot about something.

Verdict: not an addiction. Not close. It fails every core criterion and passes only the discounted ones.

Where It Does Clear the Bar

That verdict has a real exception, and it isn’t the one people brace for.

MIT Media Lab and OpenAI ran a four-week randomised controlled trial with 981 participants and over 300,000 messages, testing text against voice modes and personal against non-personal conversation. The finding that held across every condition: participants who voluntarily used the chatbot more showed higher loneliness, higher emotional dependence, more problematic use, and less real-world socialising. Users with higher trust in and social attraction to the model showed the worst outcomes. The companion analysis of over three million conversations found the same signature and noted that a small number of users account for a disproportionate share of emotionally loaded exchanges.

That is a genuine core-criteria pattern: mood modification and conflict, the two that carry diagnostic weight. And note where it sits. Not in using AI to write code, draft a document, or think through an argument. In using it as company.

The line doesn’t fall between light and heavy use. It falls between what the tool is standing in for. Standing in for a reference book is unremarkable. Standing in for a person is where the evidence turns.

What Actually Survived

Which leaves the loose thread from the withdrawal criterion. Without the tool I’m worse, but I’m not distressed. The framework has no name for that, so it scores it as nothing.

It isn’t nothing.

At CHI 2025, Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 first-hand cases of using generative AI at work. Most reported reduced cognitive effort across the range of thinking the task required: 72% for knowledge, 79% for comprehension, 72% for analysis, 76% for synthesis. The sharper finding: higher confidence in the AI predicted less critical thinking; higher confidence in one’s own ability predicted more. The two confidences trade against each other. As trust in the tool rises, the effort you spend doubting it falls.

That’s not withdrawal. It’s atrophy, and the two differ in every way that matters for what you should do about it.

Withdrawal is acute. It hurts while it’s happening, it announces itself, and it resolves. The distress is the body recalibrating, and it ends. Atrophy is the inverse on all three counts. It’s painless while it’s happening. It never announces itself. And it doesn’t resolve on its own, because nothing is recalibrating; a capability is simply being used less and getting smaller.

Which explains why the addiction framework clears AI so easily. Every one of its six components is a contemporaneous harm detector. It looks for damage occurring while the behaviour is occurring. Conflict is happening now. Withdrawal distress is happening now. The whole instrument is calibrated to catch harms that hurt in the present tense.

The cost here is in the future tense. It’s a capability you will discover you no longer have, at some point you can’t schedule, under conditions you don’t control, which is the real version of the question about what’s left for us to do. There’s no criterion for that because until recently there was no behaviour that produced it at scale.

So “it isn’t an addiction” is not the reassurance it sounds like. The test came back clean because the test can’t see this.

Why the Wrong Frame Produces the Wrong Fix

This isn’t only a definitional quibble, because the two problems have opposite treatments.

Everything currently marketed for AI overuse is an addiction intervention: the detox, the abstinence week, the usage cap, the app that locks you out. Those tools do one thing: interrupt a compulsion and let a drive subside. They work on relapse and withdrawal. They’re aimed squarely at the criteria this fails.

Against atrophy they don’t merely underperform. They do nothing. A week away from the tool doesn’t rebuild the skill that transferred to it; it just gives you a week of doing less. Then you come back, relieved that you weren’t addicted after all, having tested for the wrong thing and passed.

The treatment for atrophy is the one thing a detox never asks for: deliberate practice on the specific part you’ve stopped doing. Draft the argument before asking for a draft. Form your own estimate before requesting one. Read the primary source rather than the summary of it. This is the effortful part of thinking, and it is the part that’s cheapest to hand over. Not as abstinence, but as load-bearing, in the ordinary course of the work, in the places where being wrong would actually cost you something.

The Microsoft data points at exactly where to spend it, too. The people who kept thinking critically weren’t the ones who distrusted AI. They were the ones with high confidence in their own judgment. That’s a capability you maintain by using it, and the only reliable way to keep trusting your judgment is to keep having occasion to check it.

The Uncomfortable Answer

I went in expecting to find that “addiction” was the wrong word, and I did. The instrument is clear: not an addiction, not a disorder, and the components that fire loudest are the ones that predict flourishing rather than distress.

I expected that to be reassuring. It isn’t.

An addiction, at least, is a problem that argues with you. It creates conflict, produces distress, and generates the friction that eventually makes you deal with it. This produces none of that. It’s frictionless in the present and expensive later, which is the hardest possible shape for a problem to have: the kind you never notice you’re not solving.

The question was never whether AI is an addiction. It’s whether you’d notice if it were something worse: something that passes every test we have, because we built the tests for a different injury.