AI addiction: what it is, the signs, and what helps
AI addiction is a compulsive-use pattern where someone keeps turning to AI tools even when it costs them time, skills, or wellbeing. It is not an official diagnosis. The signs, the research, and the practical steps below can help you tell heavy use apart from a problem worth acting on.
The term covers behavior, not a medical label. Researchers frame problematic AI use the way they frame gambling or gaming problems: a behavioral addiction, marked by loss of control rather than a chemical dependency. Whether that will ever become a recognized clinical condition is a separate question, and one this page covers below in what the research measures and why no formal diagnosis exists yet. This page is the practical overview.
One thing has changed since we first published this guide in mid-2026: there is now more research, and most of it points the same direction. The tools are getting stickier, the studies are getting sharper, and the workplace version of the problem is turning out to be the one most people miss. We have folded that newer work in below.
If you would rather skip straight to your own situation, the free Is AI Changing How You Think? assessment scores your actual habits in about 2 minutes, anonymously, and you can come back to the article after.
The three patterns it shows up in
AI dependence rarely looks the same from one person to the next. Three patterns come up most often.
Companion and roleplay chatbots. Apps like Replika and Character.AI are built to feel like a relationship. People spend hours a day talking to a character, feel a jolt of distress when they can't reach it, and start preferring it to the friends and family who are harder work. Companion apps carry the strongest emotional pull of the three. If it describes you or someone you know, our dedicated guide to character AI addiction goes deeper on the companion-app version specifically.
General assistants like ChatGPT. Here the pull is cognitive rather than romantic. The tool becomes the reflex for every question, every draft, every decision, until thinking through a problem alone starts to feel uncomfortable. The OpenAI and MIT Media Lab research found that the heaviest users were also more likely to report loneliness and emotional dependence. We cover the assistant-specific version in ChatGPT addiction: when daily use becomes dependence.
Workplace overreliance. This one hides in plain sight because it looks like productivity. You lean on AI for tasks you used to handle yourself, and over months the underlying skill quietly fades. We break down what that looks like day to day in signs you're relying on AI too much at work.
The three overlap more than they separate. A person can start with ChatGPT for work, drift into using it for company on slow evenings, and end up leaning on it for both. What matters is less which box you fall in and more whether the use still serves you or has started to run the show.
Common signs
No single item here settles anything. Several of them together, over weeks rather than a bad afternoon, is the part worth taking seriously.
- You reach for AI before you have really tried the task yourself.
- You feel restless, anxious, or irritable when you can't use it.
- Your sessions run much longer than you meant them to.
- You have cut back on sleep, work, or time with people to keep using it.
- You hide or downplay how much you use it.
- A skill you once had feels rusty because the tool now does it for you.
- You keep using it the same way even after it has clearly cost you something.
Read that last one twice. Continued use despite a cost is the item clinicians weight most heavily across every behavioral addiction, because it is the clearest sign that the behavior has stopped being a free choice.
A quick self-check
Answer yes or no, and answer for how things actually are, not how you would like them to be. Several of these restate the signs above as questions, which is deliberate: most people find it easier to see themselves in a direct question than in a list. This on-page list is a rough, informal read. It is not the AIAS-21 and it is not the CAIDS-20. The wording is adapted from those instruments' dimension names, not copied from either scale.
- Do you open an AI tool without really deciding to?
- Do your sessions regularly run much longer than you planned?
- Do you feel restless or irritable when you can't use it?
- Do you think about using it when you're meant to be doing something else?
- Have you cut into sleep, work, or time with people to keep using it?
- Do you reach for it before trying the task yourself?
- Have you tried to cut back and not managed to?
- Do you play down or hide how much you use it?
- Has it clearly cost you something, and you kept going anyway?
- Would a full day without it feel hard?
0 to 2 yeses. Your use looks like normal heavy use. Worth a light touch of awareness, nothing more.
3 to 5 yeses. Some markers of dependence are showing. A good moment to set a few limits and watch the pattern for a couple of weeks.
6 or more yeses. The pattern looks entrenched enough to act on deliberately, and possibly to talk through with someone.
The count above is a rough, informal read, not a diagnosis, and no number here means you have a disorder. One thing the count can't see: how you felt answering. If several questions stung, or you found yourself softening a yes into a maybe, that reaction tells you as much as the tally does. The heaviest weight sits on the harm question above. Continuing to use something after it has cost you sleep, work, or a relationship is the marker researchers treat most seriously, and a single yes there deserves more attention than five soft ones elsewhere.
What the research measures
Two published questionnaires show up in almost every serious writeup of "AI addiction." Neither one is a diagnosis. Both borrow language from older behavioral-addiction research. The useful move is to know what they asked, who they asked, and where the numbers stop.
The AI Addiction Scale (AIAS-21) is a 21-item self-report by Igor V. Pantic, sole author. The journal piece is a correspondence, a letter, in European Child and Adolescent Psychiatry, posted online 30 September 2025 (issue January 2026; 2026;35:297; doi:10.1007/s00787-025-02874-8). It is not a validation study. The letter's data-availability statement is blunt: no datasets were generated or analysed. There is no sample. The 21 lines sit on seven constructs, three items each: compulsive use, craving, tolerance, withdrawal, preoccupation, continued use despite harm, and functional impairment. Answers run from 1 (strongly disagree) to 5 (strongly agree). The letter names four risk bands (minimal, moderate, high, very high) and does not print numeric cutoffs. Numeric bands appear only on Pantic's instrument page and are labeled provisional. The letter also says the scale has not been validated in the general population, that AI addiction is not a recognized psychiatric disorder, and that there are no established diagnostic or treatment guidelines. We do not reprint those 21 items.
The site quiz is a different tool. Our two-minute assessment is 15 adapted items, on a 4-point frequency scale, scored 0 to 45. It is not the AIAS-21, and it is not the CAIDS-20.
The Conversational AI Dependence Scale (CAIDS-20) is a validation paper by Yuanyuan Chen, Mengyun Wang, Shujuan Yuan, and Yan Zhao, published in Frontiers in Psychology on 31 July 2025 (2025;16:1621540; doi:10.3389/fpsyg.2025.1621540). It focuses on chatbots people talk with rather than tools they use for a single task. After interviews and factor analysis they kept 20 items on a 6-point Likert (1 completely disagree to 6 completely agree) in four factors: uncontrollability, withdrawal symptoms, mood modification, and negative impact. Study 1 used 547 valid questionnaires. Study 2 was a confirmatory factor analysis with N=687; in that sample the mean CAID score was 3.83 (SD 0.98). Study 3 (n=1,081) described how often that group used conversational AI: mean 1.91 years, more than 50% more than two hours a day, and more than 74.8% almost every day. That is one age band in one country among people who already used the tools. It is not a world rate for "AI addiction." Those are separate samples (547, 687, 1,081, and a later n=892 wellbeing sample), not one combined N. The authors say the scale still needs testing in other groups and cultures.
There is also a narrower clinical framework specifically for emotional dependency on AI companions, tools like Replika or Character.AI, distinct from dependency on productivity tools used for work. That distinction matters: the emotional-companion pattern and the productivity-overreliance pattern look different day to day, even where the underlying mechanics overlap.
| AIAS-21 | CAIDS-20 | |
|---|---|---|
| Author | Igor V. Pantic (sole) | Yuanyuan Chen, Mengyun Wang, Shujuan Yuan, Yan Zhao |
| What it is | Correspondence / letter (online 30 Sep 2025) | Validation paper (31 Jul 2025) |
| Items / scale | 21 items, 5-point Likert | 20 items, 6-point Likert |
| Structure | 7 constructs, 3 items each | 4 factors |
| Sample | None. Not validated in the general population | Chinese college students; Study 1 n=547; Study 2 n=687; Study 3 n=1,081 |
| Cutoffs | Named bands in the journal; numeric bands only on the author's site, labeled provisional | Higher score = more dependence; no clinical cutoff |
| What it is not | A diagnosis, a validation study, or a prevalence survey | A diagnosis or a global prevalence survey |
One caution about scores. A high number on either questionnaire flags a pattern worth attention. It does not label you with a condition, because there is no condition to label. These tools were built to help researchers study a group, and they work best as a mirror rather than a verdict. If a self-check leaves you worried, that worry is the useful part, not the total.
It is worth saying that not every researcher is convinced the "addiction" framing fits. A 2025 critique in Addictive Behaviors, titled "People are not becoming AIholic", argues that borrowing withdrawal-and-tolerance language from substance use may over-pathologize what is often just heavy, habitual use. That debate is healthy, and it is one reason we keep the language on this site cautious. A pattern can be worth changing without being a disorder.
Why no formal diagnosis exists yet
Diagnostic criteria typically require years of longitudinal data, replication across populations, and consensus among clinical bodies before they get formalized. This research area is barely two years old in its current form. Internet addiction, for comparison, has been studied since the late 1990s and still is not universally recognized as a standalone diagnosis in every major diagnostic manual.
You do not need an official diagnosis to take a pattern seriously in the meantime. The absence of a formal DSM entry does not mean the underlying behavior, and its documented effects on judgment and retention, is not real. Clinicians already work with people on compulsive-use patterns using frameworks borrowed from internet and behavioral addiction treatment, cognitive behavioral therapy and motivational enhancement therapy among them, with reasonable results.
What newer 2026 research is finding
The picture has filled in over the last year, and two threads stand out.
The first is about thinking. In "Your Brain on ChatGPT," a team led by Nataliya Kosmyna at MIT Media Lab had 54 people write essays either unaided, with a search engine, or with ChatGPT, while recording their brain activity with EEG. The group using the AI showed the weakest, least connected brain networks during the task. When some of them later had to write without the tool, they stayed under-engaged, as if the habit of offloading had carried over. Eighteen of the 54 completed a fourth session. The authors call the result "cognitive debt." It is a preprint with a small sample, so treat it as an early signal rather than a settled fact. A single essay task in a small sample is not your whole cognitive life.
A related 2025 paper by Michael Gerlich in Societies (n=666, UK; doi:10.3390/soc15010006) reports a negative correlation between frequent AI use and critical-thinking scores, mediated by cognitive offloading. Correlation, not causation. UK sample only.
The second thread is about what happens when the tool goes away. A 2026 diary study of knowledge workers, published as a preprint under the title "Oops! ChatGPT is Temporarily Unavailable," tracked how people reacted during real outages of the service. Some reported a genuine jolt of stress and a scramble to work around the gap. Withdrawal-like reactions to a productivity tool are exactly the kind of thing the AIAS-21 tries to measure, and seeing them show up in a workplace diary rather than a lab questionnaire is telling.
None of these studies proves that AI is addictive in the clinical sense. What they add is texture. The problem is not only that people talk to chatbots for company. It is also that the everyday, professional, seemingly harmless use has its own hooks.
The workplace version, up close
Because workplace overreliance is the pattern people most often miss in themselves, it deserves its own section.
At work, dependence does not announce itself. It arrives disguised as being good at your job. You ship the report faster. You clear the inbox before lunch. Nobody complains, because the output looks fine. The cost is invisible in the short run and lands later, in a skill you can no longer perform cold.
A 2025 study in Social Science & Medicine on "addictive ChatGPT use" framed this as becoming a "cognitive miser": the more the tool handles the effortful thinking, the more the mind defaults to letting it, until reaching for the AI becomes automatic rather than chosen. Pressure to be more productive feeds the loop directly. When the expectation is to move faster every quarter, offloading the hard parts stops feeling optional.
Here is the risk. If you can only do a task with the tool, you no longer really have the skill; you have access to it. Those are different things, and the difference shows up the day the tool is down, the client wants to watch you work, or the interview asks you to do it live. Keeping some tasks in your own hands is not nostalgia. It is insurance. The skills argument runs through how to use AI without losing your skills, and if career risk is what is really on your mind, our job assessment covers the automation side of the same coin.
There is also a quieter cost that rarely gets named: the loss of the small, boring moments where thinking used to happen. Drafting a tricky email used to force you to figure out what you actually wanted to say. Summarizing a report used to make you read it closely. When the tool absorbs those moments, the work still gets done, but the understanding that used to come free with the effort does not. Some of that is fine to trade away. The question worth asking is whether you are trading it on purpose or by default, because default is how dependence sets in without anyone deciding to let it.
"AI addiction test": what to make of the phrase
A lot of people arrive at a page like this after typing "AI addiction test" into a search bar. It is worth being straight about what that phrase can and cannot deliver.
No online test can diagnose an AI addiction, because there is nothing to diagnose. There is no clinical AI-use disorder, so any tool that hands you a verdict is overselling. What a good self-assessment can do is different and still useful. It can turn a vague, nagging "am I on this too much?" into something specific: which pattern you fall into, how often the signs show up, and whether the trend is worth acting on before it costs you more.
Turning that vague worry into something specific is what our two-minute assessment is built to do. It is 15 adapted items on a 4-point frequency scale, scored 0 to 45. It is not the AIAS-21 and it is not the CAIDS-20. It is anonymous, and it does not save your answers to your name. Think of it as a mirror, not a scorecard. The point is not the number at the end. It is what you notice about your own habits while answering.
One more note on the search phrase itself. "AI addiction test" barely registers as a defined term in the research or the clinical world, which means the language is still up for grabs. The lack of a fixed definition cuts both ways. It leaves room for well-grounded tools like ours, and it leaves plenty of room for clickbait that promises a diagnosis it cannot give. When you land on any test, the tell is what it does with your result. A responsible one hands you a pattern and a next step. A bad one hands you a label and a sales pitch.
What actually helps
Most people who worry about their AI use are not in crisis. They want the tool without the creep. A few things tend to work.
Start with a self-assessment. Naming the specific pattern, and how often it happens, does more than a vague sense that you're "on it too much." Our short assessment is 15 adapted items, 4-point frequency, scored 0-45. It is not the AIAS-21 and it is not the CAIDS-20. It takes about two minutes.
Then set usage limits you can actually keep. App timers, a rule that certain tasks get done unaided, and a fixed cutoff time all reduce the automatic reach for the tool. Keeping some skills in your own hands matters more than the raw hours, which is the whole argument in how to use AI without losing your skills. If you want a full walkthrough of the practical steps, we lay them out in how to stop AI addiction.
If the pull is social, other people help. Internet and Technology Addicts Anonymous is a free twelve-step fellowship that now welcomes people struggling with compulsive AI use, with meetings online every day. Journalists at 404 Media spent time inside these groups and documented people using them to cut down on chatbot use.
Last, treat what sits underneath. Compulsive AI use often rides alongside anxiety, depression, or loneliness, and the AI habit tends to loosen when those are addressed. A therapist can work on the AI pattern using the same cognitive behavioral tools that already help with other behavioral addictions.
When to seek professional support
Talk to a clinician if your use is affecting your job, your sleep, or your relationships and you have tried to cut back without success. Reach out sooner if the AI has become your main source of emotional support, if you feel real distress when you're away from it, or if you're noticing low mood or anxiety alongside the habit. You do not need a formal diagnosis to deserve help. A pattern that is costing you something is reason enough.
This page is informational and not a substitute for professional advice. If you're in crisis or having thoughts of suicide, contact a local emergency line or, in the US, call or text 988 to reach the 988 Suicide and Crisis Lifeline, which is free, confidential, and available 24/7.
Frequently asked questions
Is AI addiction officially recognized?
No. There is no AI-use disorder in the DSM or ICD. AI addiction is not a recognized psychiatric disorder. The AIAS-21 is a correspondence proposing a scale, with no sample and no general-population validation. The CAIDS-20 is a validation study in Chinese college students. Neither is a diagnostic instrument, and no major clinical body has proposed a formal diagnosis. Some researchers even question whether "addiction" is the right frame at all.
How many people are affected?
There is no reliable figure yet. The OpenAI and MIT Media Lab work analyzed more than 4 million ChatGPT conversations and found that a small share of heavy users account for most of the emotionally loaded interaction. The finding describes a concentrated pattern, not a measured population rate.
Is it the same as internet addiction?
They overlap. Internet and technology addiction is the wider category, and ITAA already treats compulsive AI use within it. The AI-specific twist, especially a chatbot that talks back, adds features the older research doesn't fully cover.
Is this comparable to social media addiction research?
Somewhat. Both are behavioral rather than substance-based, and both faced the same slow path toward formal recognition. Social media research is roughly a decade ahead in that process, which is part of why AI-specific screening tools are still so new.
Can a doctor actually diagnose "AI addiction" today?
Not as a standalone diagnosis. A clinician might diagnose a related condition, like an anxiety or compulsive-use pattern, and treat the AI-specific behavior within that broader frame.
Is an "AI addiction test" a real diagnosis?
No. Any test that gives you a diagnosis is overstating what it can do, because there is no condition to diagnose. A well-built self-assessment can still help you spot a pattern and decide whether to act on it.
Can you fix it without quitting AI entirely?
Usually. The common goal is a healthier relationship with the tool, not abstinence: limits, some tasks done unaided, and treatment for any anxiety or depression underneath. A minority need a stretch away from it, which is worth discussing with a professional.
Published July 2026. Updated August 2026.
ChatGPT addiction · Character AI addiction · AI chatbot addiction · How to stop AI addiction · Signs you're relying on AI too much at work · How to use AI without losing your skills
Not sure where you land?
The free Is AI Changing How You Think? assessment scores your actual habits. It is 15 adapted items, 4-point frequency, scored 0-45. It is not the AIAS-21 and it is not the CAIDS-20. It takes about 2 minutes and it is anonymous.