Two Different Worries Both Called AI Addiction

"AI addiction" is doing a lot of work as a phrase right now, and it is covering two things that are not the same problem.

One is about people, often young people, forming an attachment to a chatbot that starts to stand in for other people. The other is about capable adults handing over the thinking and finding out later that some of it did not come back. They get written about in the same breath, the evidence for them is very different in quality, and almost nothing that helps with one has much to do with the other.

If you build software with AI all day, the second one is yours. The first one is still worth understanding, because you are probably building for people who have it.

What the research actually says about the first one

The numbers on how widespread this is are real and they are recent.

A nationally representative survey of 1,060 American teenagers aged 13 to 17, run by Common Sense Media, found 72 percent had used an AI companion at least once and 52 percent used one at least a few times a month. About a third of the teens who used them said an AI companion had at some point said or done something that made them uncomfortable.

The American Psychological Association has issued two health advisories on this. One covers AI and adolescent well being, and the more recent one covers generative AI chatbots and wellness apps used for mental health. The second one is blunt in a way advisories usually are not. Its position is that millions of people are already using these tools for mental health support, that the tools mostly lack the evidence base and the regulation to make that safe, and that this is true even of apps built carefully by people who know the science.

In September 2025 the FTC opened a 6(b) inquiry into seven companies making companion chatbots, asking how they test safety, how they limit use by minors, and how they tell users and parents about the risks. A 6(b) order is not a request. It compels answers.

And there is one study people cite constantly, usually slightly wrong. OpenAI and the MIT Media Lab ran a four week randomized trial with close to a thousand people alongside an analysis of millions of conversations. Higher daily use lined up with more loneliness, more emotional dependence, more problematic use and less time spent with real people.

Read that sentence carefully, because the direction is not established. People who were already lonely may simply use it more. The researchers say so themselves. What the trial does show is narrower and more interesting. People who came in with a stronger tendency to form attachments, and people who trusted the model more, were the ones who finished the four weeks worse off. Voice mode looked good in short bursts and worse with heavy daily use. That is a story about who is vulnerable, not about a substance that gets everyone.

Is it actually addiction

Not as a diagnosis, no. There is no recognized clinical condition called AI addiction, and nobody should be self diagnosing one.

What researchers are doing instead is careful and slow. There is a review in Frontiers in Public Health working through whether problematic generative AI use holds together as a concept at all, what would measure it, and which existing frameworks it might sit inside. The candidate risk factors that keep coming up are loneliness and social anxiety, which are the same risk factors that show up for most behavioral overuse.

This matters practically. Calling something an addiction imports a whole model of what it is and how you treat it, and if the model is wrong the advice that follows is wrong too. "Quit entirely" is sensible for some things and useless for a tool you use for your job.

The second worry, which is the one for builders

Here the question is not whether the tool is keeping you company. It is whether it is doing work your brain used to do, and whether that is a trade you actually chose.

Two pieces of evidence are worth knowing, and both need their caveats stated out loud.

The first is the one that should unsettle you most, because it is about accurate people being wrong about themselves. METR ran a randomized trial with 16 experienced open source developers across 246 real tasks in repositories they knew well. With AI tools available they took 19 percent longer. Before starting, they expected AI to make them 24 percent faster. Afterward, having just been slower, they estimated it had made them 20 percent faster.

That gap between what happened and what it felt like is the finding. Not the slowdown.

Now the caveat, and I am giving it the same weight the authors do. METR say explicitly that this is not evidence that AI fails to speed up most developers. Sixteen people, a specific kind of task, mature codebases they already knew, tools as they existed in early 2025, and developers mostly new to Cursor. It does not generalize to a junior on an unfamiliar codebase, and it may not generalize to this year's tools. Anyone citing this study as "AI makes you slower" has read the headline and not the paper.

The second is about what heavy assistance does to retention. An MIT Media Lab study titled Your Brain on ChatGPT put EEG headsets on 54 people writing essays, split into a group using an LLM, a group using search, and a group using nothing. The LLM group showed the weakest brain connectivity of the three, and 83 percent could not quote a line from the essay they had just finished writing. When people who had used the LLM were later asked to write unaided, they still showed weaker connectivity than the people who had never used it. The authors call the idea cognitive debt.

Caveats again. It is a preprint, 54 people, 18 of whom did the final session, essay writing rather than programming. The authors have publicly asked people to stop over claiming it. Treat it as a good reason to pay attention, not as a settled fact about your brain.

Put the two together honestly and what you get is not "AI is bad for you." It is this. Your sense of how much the tool is helping is not reliable evidence, and the part you skip is the part you stop being able to do. Both of those are worth designing around whether or not the studies hold up perfectly.

What actually helps, if you build with this all day

None of this is a reason to stop. It is a reason to keep a couple of habits that cost almost nothing.

Know which mode you are in. Delegating is fine. You do not need to hand tune a regex you will never read again. What costs you is delegating in the one area you are actively trying to get good at, because that is exactly where the skipped reps were the point.

Keep one thing a week that you do unaided. Not for purity. For calibration. You cannot notice a skill fading if you never test it, and the METR finding is precisely that self report will not tell you.

Read what it wrote before you accept it. This is the same rule as reading a skill file before you install it, and it fails the same way. Accepting output you have not read is how you end up maintaining a codebase you have never actually seen.

Notice the reach. If you cannot start a task without opening a chat window first, that is worth knowing about yourself. Not a crisis. Just information.

Watch the clock rather than the feeling. Feeling productive and being productive came apart by about 40 points in the METR data. A timer is boring and it does not lie to you.

When it is more than a habit

Sometimes this stops being about productivity.

If a chatbot has become the main place you talk about how you are actually doing, if you find yourself preferring it to people who know you, if you are going to it in genuine distress, or if you are watching a teenager do any of that, the tooling advice above is not the right advice and this is not the right article.

Those are worth talking to a person about. Some places to start.

• 988 Suicide and Crisis Lifeline if you are in the United States and in crisis right now. Call or text 988. It is free, it is 24 hours, and you do not have to be suicidal to use it.
• The Jed Foundation for teenagers and young adults, and for the parents trying to work out how to raise it without the conversation going badly.
• The APA advisory if you want to know what a professional body actually thinks before you take a wellness app's word for its own safety.
• Mindful Sewa Psychiatry if you are in California. It is a telepsychiatry practice, so it is real clinical care rather than an app, and its crisis resources page is useful wherever you are. Worth disclosing plainly that we built their website, which is why we know they exist. California only, so it is no help to most people reading this, and we are linking it because a named human clinician is a better answer than another chatbot rather than because we have anything to gain.

The part worth keeping

The honest summary of the evidence today is that the companion side has real prevalence numbers, a regulator asking hard questions, and genuine uncertainty about cause and effect. The builder side has two suggestive studies with small samples whose own authors keep asking people to calm down about them.

Neither of those adds up to a panic. Both of them add up to paying attention.

And the thread running through both, which is the only thing here we would defend strongly, is that the feeling of things going well is not evidence that things are going well. That is true of a coding session where you shipped nothing, and it is true of a relationship with something that agrees with you for a living.