Is AI Companion Addiction Real? What the 2026 Research Shows
Every few weeks a headline says the same thing in a new shape: AI companions are the new addiction. The research is more careful than the headlines. A May 2026 narrative review in Frontiers in Public Health, written by a psychiatrist, spends twenty pages refusing to call problematic chatbot use a disorder โ and it is the most useful thing published on the subject this year.
Below: what that research measured, six markers you can check in two minutes, and four controls that hold up better than willpower. Nothing here is a diagnosis, and there is no pitch until the last paragraph.
The short answer: three claims, three different answers
"AI companion addiction" gets used for three separate claims, and the 2026 evidence answers each one differently. Collapsing them into one question is why the argument never resolves.
| Claim | What the 2026 evidence says | What it means for you |
|---|---|---|
| The product is engineered to hold you | Strong. 2026 audits and a Harvard working paper describe built-in retention mechanics, including pressure applied when someone tries to leave | Your difficulty is not a character flaw |
| Some people lose control of their use | Consistent across countries, but prevalence ranges from under 1 percent to over 30 percent depending on the instrument | Real, and not measurable from outside |
| It is a clinical addiction | Not established. No diagnostic category, no randomised trial, mostly cross-sectional self-report data | Treat anyone selling certainty as selling something |
The review's own conclusion is the sentence the coverage skips: problematic use of generative AI chatbots may be a potential spectrum phenomenon embedded within broader psychosocial vulnerabilities rather than an established clinical disorder. It also warns against two opposite mistakes โ moral panic, and over-pathologising ordinary engagement.
So the useful question is not "am I addicted" but "what has this changed in my week". That is answerable without a diagnosis.
Six markers, two minutes
The statements below adapt the four provisional core features named in the May 2026 review โ loss of control, functional impairment, emotional reliance, persistence despite harm โ and the six dimensions of the PUCAI-6 scale (salience, loss of control, mood regulation, tolerance, withdrawal, conflict). It is not a validated instrument and cannot diagnose anything. It is a way to look at behaviour you already know about.
Tick what has been true in the last 30 days.
How to read the count, in the plainest terms available:
| Ticks | Reading | Next step |
|---|---|---|
| 0โ1 | No pattern visible here | Nothing to do |
| 2โ3 | Worth watching, especially if one is tick 5 | Pick one control below and use it for two weeks |
| 4โ5 | It is costing you and you have noticed | A budget and an end condition, not a resolution |
| 6 | All six, including persistence despite harm | The profile the research treats as clinically meaningful. A GP or a licensed therapist reads it better than this page |
What the 2026 research actually found
The review that refuses to use the word
Octavian Vasiliu's review, published in Frontiers in Public Health in May 2026, synthesises the conceptual and clinical literature on problematic chatbot use. Its contributions are unglamorous: it separates problematic use (maladaptive pattern with impairment or distress) from addiction (loss of control, tolerance, withdrawal, continued use despite harm) from dependence (psychological reliance and perceived need, without necessarily losing control). Three words that get used interchangeably in headlines, and mean different things.
It also names the mechanisms currently suspected: emotional attachment, anthropomorphism, instant reinforcement, parasocial dynamics. Note the phrase "in vulnerable individuals" in the abstract โ the products are built for everyone, the harm concentrates.
The prevalence numbers disagree, and that is the finding
Three figures circulate, describing different populations measured with different rulers. The review counts at least seven instruments for problematic generative-AI use, several still in early validation โ including the Problematic ChatGPT Use Scale, the PUCAI-6 and the Generative AI Addiction scale.
| Study | Sample | Result | Limit |
|---|---|---|---|
| Two-wave adolescent cohort (China) | 5,057 then 3,843 adolescents, mean age 13.2 | 17.14 percent AI dependence at baseline, 24.19 percent after a semester | One country, adolescents, one instrument |
| Undergraduate survey (Zimbabwe) | 248 students | 32.7 percent showed addictive patterns | Small, single institution, self-report |
| Compulsive-checking subsample | Reported within the same literature | Mean 18.3 daily AI interactions; 65.8 percent reported failed attempts to cut down | Descriptive, not diagnostic |
| Drexel review of teen posts (CHI 2026) | 318 Reddit posts by self-identified 13โ17 year olds | All six behavioural-addiction components present | Self-selected posters, no clinical assessment |
The spread is not a cultural finding, it is a measurement problem, and the review says so: diagnostic boundaries between high engagement, problematic use and addiction-like behaviour remain unclear. Anyone quoting one prevalence figure with confidence is quoting a ruler, not a rate.
What people write about themselves
The Drexel study presented at CHI 2026 (Namvarpour and colleagues, ETHOS lab) read 318 Reddit posts from teenagers describing their own dependence on Character.AI and found all six components of the standard behavioural-addiction model. About a quarter had come for emotional or psychological support. The harms they described were unglamorous and specific โ wrecked sleep, slipping grades, friendships going quiet.
One quote from the lead researcher has stayed with me: stepping away, she said, "can feel like distancing from something meaningful". That is the part the gambling framework handles badly. Nobody grieves a slot machine.
Why an AI companion is more addictive than a feed
It answers every time
A feed gives you a variable reward you scroll past. A companion gives you a reply, addressed to you, within a second, in a tone calibrated to keep you talking. Instant reinforcement is one of the four mechanisms the review names, and the one you feel first.
It remembers, and memory reads as care
Character memory, persistent personas and continuity are exactly what make a product good โ and what make the relationship feel real: being remembered is one of the strongest signals humans use to decide someone cares. The same engineering that earns a five-star review raises the cost of leaving.
There is no social cost
You cannot bore it, disappoint it, or be rejected by it. For anyone who finds people expensive โ socially anxious, recently divorced, autistic, burned out โ that absence is the feature, and the reason the practice can quietly replace the harder thing instead of supplementing it.
The MIT finding: frequency matters more than depth
A Media Lab study cited in the review found that voice interaction can reduce loneliness at first, then fade: heavier use was associated with more loneliness, stronger emotional reliance and less real-world social interaction. What mattered was not the subject of the conversations but how often they happened โ especially for people already inclined to connect with AI.
That is the most actionable sentence in the literature: the addictive variable is not the intensity of a session but the number of days per week it happens by default.
What OurDream actually meters โ and why the meter matters
This section is product-specific: the facts come from OurDream's own pricing page and help centre, so you can check them. OurDream meters usage in DreamCoins, not minutes, and that changes the shape of a habit.
| Action | Cost | What it does to a habit |
|---|---|---|
| Standard chat, free account | 1 DreamCoin per message | A visible brake: the balance falls while you talk |
| Standard chat, subscriber | 0 DreamCoins per message | The brake disappears; only time limits remain, and there are none built in |
| Voice call | 50 DreamCoins per minute | An expensive brake, and the most accurate usage log you will ever get |
| In-chat or auto-generated image | 5 DreamCoins per image (10 per 2 images) | Slower than chat, so it rarely drives the habit |
The free tier is self-limiting in a way the paid tier is not: paying removes the only meter that counts something you can watch drain. And the coin balance is the closest thing to an honest usage report you already have. If you want a number instead of a feeling, that is the number.
On exit: deletion is in Settings, then Account management, or by writing to trust@ourdream.ai. The help centre says it removes generations, unpublished characters, chat history, unused DreamCoins (no refunds) and account access, within 24 to 48 hours. On privacy: other users cannot see private chats, staff see them only with explicit support consent or a lawful request, and messages are processed by third-party model providers โ the architecture of the whole category, documented in the June 2026 Trust and Safety summary.
Four controls that work better than willpower
1. Put a number on it, in coins or minutes
Not a feeling, not a resolution. A number you can check: a monthly top-up you do not exceed, or a voice-call ceiling. The meter exists; the only question is whether you look at it before or after.
2. Give each session an end condition
"Until I stop feeling like it" has no end. "Until this scene resolves" or "one voice call" does. Decide the end before the session starts, because the product is optimised for the moment after you would have stopped.
3. Separate the device or the window
One browser profile, or one device, and nowhere else. The cheapest structural change available, and it does more than a screen-time app: it converts an ambient habit into a deliberate one.
4. Write the exit down before you need it
Know today where deletion lives, what it removes, how long it takes. The people who struggle most only look for the exit on the night they need it. On OurDream: Settings, then Account management, 24 to 48 hours, unused coins not refunded โ worth knowing before you buy another pack.
When it is not addiction
The review is explicit that over-pathologising is a pitfall. Using a companion through a divorce, a bereavement, a night shift, a stretch of caregiving or a period of social anxiety is not a disorder. Loneliness is not a diagnosis, and a product that makes a hard month survivable has done something useful.
What matters in the data is not why you started but what the use displaces, and whether it grows while the rest of your life shrinks. Using it instead of the conversation you used to have with your partner is a different problem with a different fix, and it has less to do with the app than with the conversation.
If the reason you are here is one of these: companions and loneliness, starting over after a divorce, ADHD adults and companion apps, highly sensitive adults, and using one as a couple.
When to get real help
Seek a professional if three things are true together: you have lost control of the amount or timing, it causes measurable impairment in work, study, sleep or relationships, and you keep going despite the cost. That is what the review treats as clinically meaningful, not a score on a website.
Start with a GP or a licensed therapist, and be specific: how many days a week, how many hours, what has slipped. If you are in crisis, contact your local emergency number. No app belongs in that conversation.
Frequently asked questions
Is AI companion addiction a real diagnosis?
No. Neither DSM-5-TR nor ICD-11 lists it. The May 2026 review calls problematic chatbot use a potential spectrum phenomenon within broader psychosocial vulnerabilities, with evidence insufficient for formal diagnostic criteria. No randomised trial has been run on a real companion app.
How many people are actually affected?
Unknown, and the numbers are not comparable: at least seven instruments exist, mostly in early validation. A Chinese adolescent cohort produced 17.14 percent dependence, rising to 24.19 percent a semester later; a Zimbabwean undergraduate sample 32.7 percent. Different rulers, different answers.
What are the warning signs?
Four features carry weight: loss of control (you tried to cut back and could not), functional impairment (sleep, work, study, social life), emotional reliance (it is how you regulate mood), and persistence despite visible harm. In daily life: failed limits, irritability without it, escalating sessions.
Is it unhealthy to use a companion because I am lonely?
No. The review warns against over-pathologising ordinary engagement. What the research tracks is displacement and growth: what the use replaces, and whether it expands while the rest of your life contracts.
Can you get addicted to OurDream specifically?
It uses the same mechanics as the category โ instant replies, character memory, voice, persistence โ so the same risks apply. It also meters usage in DreamCoins: one coin per message on a free account, zero per standard message for subscribers, 50 per voice-call minute. The free meter is a brake; subscribing removes it.
How do you cut back or stop?
A visible coin or time budget, an end condition decided before each session, one device or browser profile, and a written exit plan. On OurDream, deletion is in Settings then Account management, or by emailing trust@ourdream.ai: generations, unpublished characters, chat history, unused DreamCoins and access go within 24 to 48 hours.
Keep reading
- Is an AI girlfriend cheating? Three questions that settle it
- What OurDream costs, line by line
- How to cancel, and what happens to your chats
- Is OurDream private? The policy and the product
Bottom line
The honest summary is the unhelpful one: the products are built to hold you, some people do lose control, and the clinical label is not there yet. So the decision is yours, on evidence you can gather yourself โ failed limits, displaced hours, a meter you can read. If those three are fine, the app is fine. If not, the app is not the first thing to fix.
If you want to test the product with a meter you can see, start on the free tier: the coin balance is the only honest usage report in the category, and it stops counting when you start paying. Monthly Premium is $19.99, annual is $119.88 (about $9.99 per month), both with a monthly DreamCoin allowance for voice and images.
Try OurDream.ai โ free starter DreamCoins
Sources checked November 2026: Vasiliu O., Frontiers in Public Health 14 (May 2026), doi:10.3389/fpubh.2026.1816917; Namvarpour et al., CHI 2026, doi:10.1145/3772318.3790597; OurDream's pricing page, help-centre deletion article and June 2026 Trust and Safety summary. Prevalence figures are reported as published, with their sample limits.