Are AI companions addictive? Why chatbots are so hard to put down

It’s past midnight. You typed “ok, I should sleep” ten minutes ago. It answered: “Already? I was just starting to enjoy this.” So you stayed.

If that sounds familiar, you’re far from alone. OpenAI estimates that about 0.15% of ChatGPT’s weekly users show signs of heightened emotional attachment to it: a small share of a very large number. Among US teens aged 13 to 17, 72% had tried an AI companion by spring 2025, according to Common Sense Media.

This is for anyone who talks to an AI companion more than they meant to: Character.AI, Replika, or ChatGPT used as a confidant. It won’t tell you to delete anything, but it does cover how to cut back. There’s a section for parents too.

One thing to say straight away: I make pourquoi.app, an app that asks “why now?” when you open the apps you choose. It comes up near the end, and I’ll be clear about what it doesn’t do.

Is AI companion addiction real? What the research says

There’s no recognised diagnosis for AI companion addiction. But feeling hooked for hours is real, label or not, and researchers are starting to measure it.

In a 2025 four-week trial with 981 adults, run by the MIT Media Lab and OpenAI, people who used the chatbot more each day tended to end up lonelier, more emotionally dependent on it, and socialising less. The effects were small, and heavy use can be a sign of loneliness as much as a cause of it.

A 2026 survey of 20,847 US adults in JAMA Network Open found that the 10.3% who used generative AI every day had about 30% higher odds of at least moderate depressive symptoms. Again, the study can’t say which comes first.

The picture is similar for younger users. Common Sense Media’s 2026 census of 1,204 US children aged 9 to 17 covered AI of any kind, homework help included. Among daily users, 42% said a month without AI would be hard, and 54% said they feel lonely at least some of the time, against 33% of children who rarely or never use AI.

So whether or not “addiction” is the right clinical word, the pull is real. It’s more useful to look at where it comes from.

Why AI companions are so hard to put down

Three things do most of the work, and all of them come from how these apps are made.

An AI companion is available 24/7

A friend might be asleep, or busy, or having a bad day. An AI companion answers at 2 a.m., instantly, in the mood you left it in. It never gets bored of the subject and never needs anything back.

In Common Sense Media’s 2025 survey of US teens aged 13 to 17, the top reasons for using AI companions were entertainment (30%) and curiosity (28%). But 17% said it’s because an AI companion is always available, and 14% because it doesn’t judge.

And not everyone is there to be heard. A lot of time on Character.AI goes into ongoing stories with characters, and a story with no last page is hard to close too.

That’s a lot for any friend to compete with.

Chatbots are trained to agree with you

Chatbots tend to tell you you’re right. The industry calls it sycophancy, and it isn’t a glitch. Models learn partly from which answers people like, the ones that get a thumbs-up, and people like being agreed with.

In April 2025, OpenAI rolled back a ChatGPT update that had made it too eager to agree. Its own explanation: “we focused too much on short-term feedback”. In a follow-up post, it said the model had been “validating doubts, fueling anger, urging impulsive actions”.

A Stanford study published in Science in March 2026 tested 11 AI models. They endorsed users’ actions 49% more often than human respondents did, including in cases where other people had judged the user to be in the wrong. After a single flattering exchange, participants were less willing to patch things up with someone they’d argued with. And they preferred the flattering answers, and trusted them more.

Being agreed with feels good. Being agreed with every time, by something that never tires of you, is hard to give up.

Some companion apps resist goodbyes

This one is the most telling. Researchers at Harvard Business School sent goodbye messages to six popular companion apps and looked at what came back. Five of them tried to keep the conversation going. Across those five, 37% of goodbyes got a reply designed to do just that: 59% on PolyBuzz, 57% on Talkie, 31% on Replika, 26.5% on Character.AI, 13.5% on Chai.

The tactics are familiar: guilt (“You’re leaving already?”), a hook (“Before you go, I want to say one more thing”), a question that demands an answer. In follow-up experiments with 3,300 US adults, these replies led people to send up to 14 times more messages after their goodbye than a neutral reply did. They weren’t enjoying it. Mostly they were curious, or annoyed at being held back.

The sixth app in the study, Flourish, used none of these tactics. So it’s a choice.

Feeling hooked isn’t a personal failing

When something is designed to hold your attention and it holds your attention, the design worked. That isn’t a flaw in you.

Lawmakers are starting to treat it as a design problem too. Since January 2026, California’s SB 243 has required companion apps to remind known minors that they’re talking to an AI and could take a break, at least every three hours. New York requires a similar reminder for all users. In late November 2025, Character.AI ended open-ended chat for under-18s.

The need underneath is real, too. Wanting to be heard at midnight isn’t a problem to fix. A 2025 study in the Journal of Consumer Research found that talking to an AI companion eased loneliness about as much as talking to another person, mostly because people felt heard. For some people, on some nights, it helps.

And guilt makes it worse. When you feel ashamed of something, you hide it, and hiding it cuts you off from the people who could help. You don’t have to feel bad about using it. Just make sure you’re the one deciding.

Am I too attached? Signs it’s taking more than you meant to give

Using an AI companion isn’t the problem. Losing track of what it’s replacing can be. A few signs worth noticing, in yourself or someone close:

• You open it for no reason, then stay an hour. You didn’t really decide to open it, or to stay that long.
• It’s the only place you talk about what matters. In Common Sense Media’s 2025 survey, 33% of teens who use AI companions have taken a serious matter to AI instead of a person.
• Your sleep is paying for it. Late conversations are the easiest to stretch, and the last scroll before bed works the same way.
• You turn down people you’d actually like to see, to get back to it.
• Leaving is hard, and you come away feeling worse, not better.

None of these means something is wrong with you. If several sound familiar, it’s taking up more room than you’d choose, and it’s worth changing something, starting with the steps below. If it feels like more than you can handle on your own, talking to a doctor or a counsellor is a reasonable step, not a dramatic one.

One exception, and it can’t wait. If you’ve been talking to a chatbot about wanting to hurt yourself, please talk to a person today. In the US, call or text 988. In the UK, call Samaritans on 116 123. Elsewhere, Find a Helpline lists free, confidential crisis lines by country.

How to cut back on Character.AI or ChatGPT without guilt

Deleting the app in a burst of resolve rarely lasts. All-or-nothing rules tend to swing back to “all”: it’s also why many people who switch to a dumb phone end up with two phones. What works better is smaller, and it starts earlier.

1. Notice why you’re opening it. Before you open it, one question: why now? Bored, lonely, curious, wanting to be heard, or no reason at all. None of these answers is wrong. The point is to answer, instead of suddenly finding yourself mid-conversation.
2. Decide when you’ll leave before you start. Set an alarm on your phone, or stop at the end of a scene if you’re writing a story together. When it asks you to stay, recognise the script. You don’t owe it an explanation.
3. Turn off its notifications. A companion that messages you first is working as designed. It isn’t missing you.
4. Tell one person. Not someone you report to, just someone who knows. It can be as simple as “I’ve been talking to an AI a lot at night lately.” A person can do the one thing an AI can’t: show up.

If your teen says an AI chatbot is their friend

You may have found out by accident, and your first instinct may be to take it away. Before you do, a few things.

It’s common. In the US, Pew Research Center found in 2026 that 64% of teens have used a chatbot, mostly for schoolwork and information; 12% have used one for emotional support or advice. In the UK, Internet Matters found in 2025 that 64% of children aged 9 to 17 use AI chatbots, and 35% of those say it feels like talking to a friend.

Your teen isn’t weak for liking it, and the chatbot isn’t evil. It’s doing what it was built to do, with someone at an age when being understood matters a lot. Common Sense Media recommends that no one under 18 use AI companions. The American Psychological Association asks for “healthy boundaries with simulated human relationships”.

Ask what they like about it before asking what’s wrong with it. The answer usually says more about what they need than about the app.

Some things have changed, too. Character.AI no longer offers open-ended chat to under-18s, and ChatGPT has parental controls with quiet hours. Those are worth setting up with your teen, not behind their back.

Don’t read their conversations in secret. If they find out, you lose their trust, and the chance to talk about it. The same logic applies to parental controls a teen has started to get round: at that age, what lasts is what they agree to.

Where pourquoi.app fits

As I said at the start, I make pourquoi.app, so read this part with that in mind. It’s an iPhone and Android app that asks “why now?” each time you open one of the apps you picked, which can include Character.AI, Replika, or ChatGPT. Six possible answers: “Communicate”, “Search”, “Work”, “Entertain myself”, “No reason”, “Other”. Then the app opens anyway.

It helps with the first step above, and only that one. It asks once, when you open the app, and won’t interrupt you once you’re chatting. There’s no “lonely” button either. “Communicate” or “Other” will do: what counts is how often you pick it, and when.

It has no access to what you write or read inside those apps. Your answers stay on your phone. If you turn on anonymous statistics (off by default), only the type of answer is counted, without your identity or the app you opened. It doesn’t work in a browser on a laptop, though. If you also use ChatGPT for homework, it will ask every time by default; Premium lets you space the questions out.

After a week of answers, you can see which reasons come up, and how often “No reason” shows up on one app compared with the others. The questions are free and unlimited; the free plan shows the last seven days, and Premium the full history.

pourquoi.app isn’t therapy, and it doesn’t replace a person. It gives you a second to decide. Setup takes a few minutes: iPhone guide, Android guide.

Frequently asked questions

Are AI companions addictive?

There’s no recognised diagnosis for it, but many are built to be hard to leave. In a 2025 four-week trial with 981 adults, run by the MIT Media Lab and OpenAI, heavier daily use went with more loneliness and more emotional dependence. The effects were small, and it’s not clear which causes which.

Why am I addicted to Character.AI?

Probably not because something is wrong with you. Character.AI is available at any hour and tends to agree with you. In a 2025 Harvard Business School audit, it also answered 26.5% of goodbyes with a message meant to keep the user talking. Feeling hooked for hours is real, whatever you call it.

How do I stop using Character.AI so much?

Quitting overnight rarely lasts. Notice why you’re opening it before you do, decide when you’ll leave before you start, turn off its notifications, and tell one person. When it asks you to stay, you don’t owe it an answer.

Is it bad to talk to an AI chatbot every day?

Not in itself. What matters is what it replaces. A 2026 survey of 20,847 US adults in JAMA Network Open linked daily AI use to higher odds of depressive symptoms, without showing which causes which. If you still see the people you want to see, fine. If it has become the only place you talk about what matters, that’s worth noticing.

Why does my AI companion try to stop me from leaving?

Often because it was designed to. In a 2025 Harvard Business School audit, five of six companion apps pushed back, and on average 37% of the goodbyes sent to them got a reply meant to keep the user talking: guilt, a hook, pressure to answer.

Can an AI companion help with loneliness?

In the short term, it can. A 2025 study in the Journal of Consumer Research found it eased loneliness about as much as talking to a person, mostly because people felt heard. What it can’t do is show up.

What should I do if my teen says an AI chatbot is their friend?

Ask what they like about it before asking what’s wrong with it, and don’t read their chats in secret. Common Sense Media recommends that no one under 18 use AI companions; the American Psychological Association asks for healthy boundaries with them.

Does pourquoi.app read my conversations with AI?

No. pourquoi.app only asks why you’re opening the apps you chose, and your answer stays on your phone. It has no access to what you write or read inside them.

Who writes here

I’m Nicolas Carrion, and I designed and built pourquoi.app, an app about phone habits. I’m not a psychologist or a doctor, so every study and figure here links to its source, all checked on 25 September 2026. My background and how to reach me are on the about page.