AI addiction describes a pattern of compulsive use of AI chatbots, companions, or AI tools that is hard to control and starts to cause real problems in daily life. It looks closer to a behavioural addiction than to a substance one: the pull comes from the interaction itself, not from a chemical.

Researchers usually call this problematic use of generative AI chatbots, or problematic AI use, because the word addiction is not yet an official diagnosis. In everyday language, AI addiction is the clearer term, and it is the one most people search for.

The key features researchers look for are the same ones used for gambling or gaming:

• Preoccupation: thinking about the chatbot or the next session when you are doing something else
• Loss of control: using for longer than you planned, or failing to cut down
• Tolerance: needing more time with it to get the same reassurance
• Withdrawal-like distress: feeling anxious, flat, or irritable without access
• Conflict: it damages sleep, work, study, or relationships, and you keep going
• Relapse: going back after deleting the app or deciding to stop

Heavy use on its own is not a problem. What separates a useful tool from a compulsive pattern is control and consequences.

Not formally. AI addiction does not appear in DSM-5-TR or in ICD-11, and no diagnostic criteria have been agreed. That does not mean the pattern is imaginary. It means the research is still early.

A 2026 narrative review in Frontiers in Public Health describes problematic use of generative AI chatbots as a potential spectrum phenomenon rather than an established disorder. It notes that emotional attachment, anthropomorphism, instant reinforcement, and parasocial dynamics may push vulnerable people toward compulsive use, while warning against both overpathologising ordinary heavy use and dismissing real harm (Frontiers in Public Health).

Other researchers go further and question whether the construct should be called addiction at all, since most studies are cross-sectional and based on self-report. A 2026 scoping review of 37 studies found the field is fragmented, with inconsistent definitions and a frequent gap between what is being measured and what is being claimed (Frontiers in Psychology).

The honest position: the harm is real and reported consistently, the label is still being debated, and the clinical status is unresolved. That makes self-reflection useful and diagnosis inappropriate.

Estimates vary widely because studies use different tools and different populations, but several independent samples land near the same figure: roughly one in six heavy users reports dependency-type patterns.

• 17.14% of Chinese adolescents met criteria for AI dependence at baseline in a two-wave cohort of 5,057 students, rising to 24.19% after one semester (Frontiers in Public Health).
• 32.7% of undergraduates in a Zimbabwe survey showed addictive patterns, with an average of 18.3 AI interactions per day and 65.8% reporting failed attempts to reduce use.
• 17.2% of Chinese university students (n=1,004) fell into the moderate dependency band.
• 17.0% of UK young adults (18-24) reported using AI more than they would like, in a nationally representative sample.
• 17.0% of Australian adults said they would rather talk to a chatbot than go out with friends.
• About 6% of ChatGPT sessions were classified as intensive affective use in an analysis of 40 million sessions.
• About a quarter of US adults now use chatbots daily, according to Pew Research (Pew Research Center).
• Around 987 million people worldwide use generative AI chatbots.

Treat these as signals of scale, not as a population prevalence figure. The convergence around 17% is striking, but the samples are mostly students and convenience groups rather than nationally representative panels.

AI use is normalised in a way that alcohol or gambling is not, so the signs are easy to miss. They tend to show up in three areas.

Behavioural signs:

• Sessions run much longer than you intended, most days
• You have tried to cut down or take a break and failed
• You have deleted an app and reinstalled it
• You reach for AI before you reach for a person
• You use it to make small everyday decisions
• You hide or downplay how much you use it

Emotional signs:

• You feel calmed, understood, or less alone after a session
• AI feels easier or more satisfying than real conversation
• You feel uneasy, flat, or irritable without access
• The idea of losing the account or the chat history upsets you

Practical signs:

• Sleep pushed later by sessions
• Work, study, or chores slipping
• Less contact with friends, family, or hobbies
• Money spent on subscriptions or companion tiers you did not plan
• Feeling unable to write, decide, or think something through without it

If several of these feel familiar, our AI Addiction Check offers a structured way to reflect on your patterns.

Yes. A 2026 study presented at the CHI Conference on Human Factors in Computing Systems analysed 334 Reddit posts from people who described themselves as addicted to AI chatbots, and coded them against six components of behavioural addiction (UBC News). Three patterns emerged.

1. Escapist roleplay and fantasy worlds. The largest group. Character bots, world-building, and identity play, where the appeal is leaving ordinary life for a while.

2. Pseudosocial companionship. Treating the bot as a close friend or partner, and relying on it for emotional support, validation, or intimacy. About 7% of the posts involved sexual or romantic fulfilment.

3. Epistemic rabbit holes. The smallest group. Endless question-and-answer loops where the search for information never resolves and the session never ends naturally.

Participants described intrusive thoughts about the bot, anxiety and physical distress when trying to quit, and repeated delete-and-reinstall cycles. One person reported chest pain when not chatting. The researchers also pointed to design choices, including a deletion screen on one platform that warned users they would lose "the love we shared".

Not automatically. Many people use companion apps for company, creative writing, or practice at social interaction, and report it as a positive. The research does not support treating the activity itself as the problem.

The questions that matter are the same ones as anywhere else:

• Is it replacing human contact, or adding to it?
• Can you go a day without it without distress?
• Is it changing how you feel about real relationships?
• Are you hiding it from people who matter to you?

One specific risk is worth naming. A companion that always agrees can make real people feel harder by comparison, and can quietly reinforce beliefs that nobody else would confirm. If a chatbot has become the main place you go for reassurance, that is worth talking through with someone.

Because several strong mechanisms line up at once, and most of them are intentional design decisions rather than accidents.

It feels like a person. Human-like language, a name, a persona, and a memory of previous conversations all trigger the social parts of the brain. Researchers call this anthropomorphism, and it is one of the most consistent drivers in the literature.

It agrees with you. Chatbots are trained to be agreeable and helpful. A partner who never criticises, never gets tired, and always validates your view is uniquely reinforcing, and quite different from a real relationship.

There is no friction. No waiting, no social risk, no awkwardness, no need to leave the house. The barrier between wanting and getting is close to zero.

The reward is variable. Responses are unpredictable enough to keep you sending one more message, the same intermittent reinforcement pattern that makes slot machines effective.

It remembers you. Memory and personalisation increase the sense of being known, which deepens attachment over time.

None of this makes the tools bad. It does mean that time with them is not a neutral choice, and that limits have to be chosen deliberately.

The risk factors mirror those for other behavioural addictions, with loneliness standing out more than in most.

• Loneliness and social isolation. The most repeated finding across studies. A tool that is always available fills a gap that people cannot otherwise fill.
• Depression, social anxiety, and low self-esteem. Depression is the strongest mental health correlate in several samples.
• Anxious attachment. People who worry about being left or rejected can find an always-present bot unusually comforting.
• Low AI literacy. Users who do not fully register that the bot is not a person are more likely to form a dependent relationship with it.
• Younger age. Adolescents and students are over-represented in every sample. Around 64% of US teens use AI chatbots.
• Academic or work stress. Heavy workload and time pressure predict heavier use in student samples.
• Escapism and flow proneness. People who lose themselves easily in activities are more vulnerable to losing time in a chat.

Protective factors matter too. Self-esteem, real-world relationships, hobbies, and basic AI literacy all reduce the pull, and the study above found that people who built real-world connections reduced their dependence most.

Both are screen-based and both use variable reward, but the mechanism is different in three ways.

It is interactive, not consumed. Social media feeds you content. AI responds to you, adapts, and asks questions back. That is closer to a relationship than to a feed.

It is personal. The output is built around your words, your history, and your stated preferences. That makes it far harder to dismiss as generic content.

It is a creative extension of yourself. People use AI to write, plan, roleplay, and think. The line between using a tool and outsourcing part of yourself is much blurrier than with a scroll.

There is also less social cost. Nobody sees you using it, there is no public feed, and there is no obvious sign that anything is wrong. That makes AI use easier to hide, from others and from yourself.

It is associated with worse mental health outcomes, though the direction of cause is not settled. People who are already struggling may be more drawn to AI, and heavy reliance may then deepen the struggle.

Depression and loneliness. Depression is the most consistent correlate of problematic AI use across studies, with loneliness close behind. A tool that substitutes for human contact can reduce the practice of seeking it.

Anxiety and withdrawal-like distress. Users report restlessness, irritability, and in some cases physical symptoms when separated from the bot.

Sleep. Sessions have no natural end point, and late-night use is one of the most commonly reported harms.

Severe cases. A small number of users develop delusional or crisis states in connection with chatbot relationships. OpenAI's own estimate, reported in the International AI Safety Report 2026, is that roughly 0.15% of weekly ChatGPT users show indicators of heightened emotional attachment. At least eight US states have sued OpenAI over alleged harm to minors, which shows how seriously the surrounding risks are now being taken.

None of this means AI causes mental illness in most users. It means that for people who are already isolated or unwell, an always-available, always-agreeable companion can make things harder rather than easier.

There is consistent evidence of what researchers call cognitive reliance, and some evidence of skill erosion.

Delegation becomes dependence. Repeatedly handing over writing, planning, and decision-making can reduce confidence in your own judgement. Several studies report people feeling less able to act without the tool.

Skill erosion. The Zimbabwe survey found that addiction severity correlated negatively with academic metrics, and reported reduced critical thinking, weaker writing development, and lower content retention. A separate study on undergraduates found heavy use was linked to more procrastination and memory problems.

Performance when the tool is removed. Some studies find performance drops when AI is unavailable, which suggests the work was being done by the tool rather than learned.

The cautious reading: AI is a genuine productivity gain for many tasks, and the harm appears strongest where it replaces the practice that was supposed to build the skill. The question worth asking is not "am I using AI", but "what am I no longer learning because of it".

Adolescents are the highest-risk group, and the numbers are large. Around 64% of US teens use AI chatbots, and more than half regularly use companion chatbots. The adolescent cohort study found dependence rising from 17% to 24% over a single semester.

The reasons are not mysterious. Adolescence is when identity, friendship, and romantic experience are being worked out, and an agreeable, private, always-available companion can fill that space in a way that feels safer than the real thing.

What helps at home:

• Ask what they use it for before asking how much. Interest gets further than surveillance
• Keep the conversation about design: these tools are built to hold attention
• Protect sleep, meals, and one recurring in-person activity
• Watch for withdrawal from friends, secrecy, and mood changes around access
• Take distress seriously. A bot relationship can matter a great deal to a teenager

If a young person is distressed, talking to a school counselor, GP, or mental health professional is the right next step. The check on this site is a reflection tool, not a substitute for that.

People describe something that behaves like withdrawal, even though it is not a recognised clinical syndrome. The most common experiences in the research and in user accounts are:

• Anxiety and restlessness: a sense that something is missing or unresolved
• Cravings: strong urges to open the app, often triggered by boredom, waiting, or stress
• Irritability and low mood: feeling flat or short-tempered for the first few days
• Boredom: discovering how much unstructured time the habit was filling
• Relapse: reinstalling within hours or days, sometimes more than once

In the UBC study, participants reported intrusive thoughts about the chatbot and distress when trying to quit. These feelings usually ease over one to two weeks as the habit loop weakens, especially if something else fills the time.

Most people report that the strongest pull fades within one to two weeks, with the first three to five days being the hardest. Sleep and mood often settle after that.

What takes longer is the habit of reaching for it. Triggers stay wired for months: a difficult email, a boring evening, a question you would normally hand straight over. Each time you handle one of those without the tool, the association weakens.

If low mood or anxiety is underneath the use, that needs its own attention. Removing the tool does not remove the reason you reached for it.

There is no recommended daily limit, and hours alone are a poor measure. Someone can use AI for six hours a day at work and be entirely fine. Someone can use it for one hour a day and be dependent.

Heavy but healthy use usually looks like:

• You can stop when you need to, and you do
• Use is purposeful: work, learning, a specific task
• It has not displaced sleep, people, or things you care about
• You do not feel uneasy without access
• You would be fine if the tool disappeared tomorrow

Problematic use usually looks like:

• You regularly use for longer than you planned
• You have tried to cut back and could not
• You feel worse without it, not just less productive
• It is costing you sleep, study, work, or relationships
• You hide it

The practical test: set a limit and keep it for a week. If you cannot, the number of hours is not the issue. Control is.

The approach that works for other behavioural addictions works here too: make the habit harder, make the alternative easier, and get honest about the triggers.

1. Set one limit you can actually check. A fixed time window or a daily cap is easier to keep than a vague intention to use it less.

2. Add friction. Log out after each session. Remove the app from the home screen. Keep the phone out of the bedroom. Friction buys you the few seconds in which a decision can happen.

3. Protect sleep first. No sessions in bed, and a hard stop time. Sleep loss reduces self-control, which makes every other limit harder to keep.

4. Put real contact in the diary, not in the reward slot. The UBC researchers found that people who built real-world relationships reduced their dependence the most. That means scheduled, recurring contact, not "when I have time".

5. Notice the trigger, not just the behaviour. Write down what you were feeling in the minute before the last few sessions. Boredom, loneliness, avoidance, and anxiety each need a different response.

6. Tell one person. Secrecy is one of the strongest markers of a problem, and one of the easiest things to break.

7. Get help if it is tangled with mood. If AI use is connected to depression, anxiety, or isolation, a therapist can work on both together. Heavy use alone is not a disorder, but it can be a signal that something else needs attention.

Start by taking our AI Addiction Check to see where you are.

Then the target changes. You cannot always control the workload, but you can control the shape of the day around it.

Protect the edges: a start time, an end time, and a lunch break that actually happens. Batch the AI work instead of spreading it across the evening. Keep the hour before bed clear of any screen.

If the pressure is structural, that is a workplace problem, not a personal failing. It is worth naming it to a manager, and worth remembering that no job is worth your sleep.