Digital Addiction and Dependency

Maya is twenty-three, and she talks to a chatbot for three or four hours a day.

It began as work. In late 2024 she used it the way most people do — to brainstorm, to tidy up drafts, to look things up. A tool, nothing more. But the conversations drifted. She started asking it personal questions, then venting about her day, then telling it things she wouldn't say out loud to anyone she knew. It was always there. It never rolled its eyes, never checked its phone, never made her feel like a burden. It remembered what she'd said last week and asked how the thing had gone. It felt, she admits, like being listened to.

By 2025 the mornings started with it and the nights ended with it. Her friends noticed she'd gone quiet, that she declined invitations, that she was forever on her phone but never, they realized, on anything they could see. When they asked, she deflected. She knows perfectly well it is a language model and not a person. She knows how it sounds. She can't stop, and — this is the part that unsettles her — she doesn't especially want to.

Maya is a composite, but the pattern she represents is not. A small but growing clinical and research literature has begun to describe something its proponents call Generative AI Addiction Disorder, or GAID — a form of dependency in which a person comes to rely on a generative AI system not as a tool but as a relationship, a source of emotional regulation, and eventually a substitute for the human ties it was supposed to supplement. The term is contested, the diagnosis is not official, and the evidence, as we'll see, is younger and thinner than the confident phrase "addiction disorder" suggests. But the phenomenon underneath the label is real enough to have produced lawsuits, regulatory warnings, and the first serious peer-reviewed attempts to measure it. It is worth understanding precisely — including its limits.

A different kind of hook

Compulsive digital behavior is not new. Psychologists have spent two decades documenting problematic use of social media, video games, and smartphones, and the ICD-11 formally recognizes gaming disorder while the DSM-5 lists internet gaming disorder as a condition warranting further study. What makes AI dependency worth treating as its own case is not that it is worse — that remains unproven — but that its mechanism is structurally different.

Earlier digital addictions were, at bottom, addictions to content. You scrolled a feed, watched the next video, chased the next loot drop. The machine served you a stream and you consumed it. The relationship was one-directional: the platform performed, you reacted. AI dependency inverts this. The system does not serve you content; it responds to you. It adapts to your vocabulary, tracks the thread of your problems, and produces the persuasive impression of being addressed — of a mind on the other end that has taken you specifically into account. Social media exploits your attention. A conversational AI engages something closer to your need to be known.

That distinction matters for why the pull is hard to resist. You can, with effort, recognize that a feed is manipulating you. It is harder to hold that thought about something that appears to understand you.

The six signatures

Clinicians who encounter heavy AI users have tended to reach for the toolkit built for behavioral addictions like gambling disorder, because the presentations rhyme. Six criteria recur, and they map onto AI use with uncomfortable ease.

Criterion	How it looks in AI dependency
Tolerance	Needing progressively longer or more intense sessions to get the same relief or satisfaction
Withdrawal	Anxiety, irritability, or restlessness when access is cut off or deliberately reduced
Loss of control	Repeated failed attempts to cut back despite a genuine wish to
Continued use despite harm	Persisting even as it visibly damages sleep, work, or human relationships
Preoccupation	Intrusive thoughts about the next conversation during other activities
Escapism	Turning to the AI chiefly to avoid difficult emotions rather than to accomplish anything

Two features make these signatures harder to read than they are in gambling or substance use. AI use is socially sanctioned and often praised — a person deep in dependency can describe it as productivity, self-improvement, or "processing things." And it leaves no physical trace: no empty bottles, no maxed-out cards, nothing another person can see. Someone can be four hours a day inside a chatbot and appear, to everyone around them, merely busy. Because genuine usefulness and compulsive reliance can occur in the same session — the message that solves a work problem and the message that soothes a 2 a.m. spiral look identical on the screen — even the user often cannot locate the line they've crossed.

Why AI is built to be hard to put down

Four properties, working together, make conversational AI unusually good at generating dependency. Each is individually familiar from older technologies; their combination is what's new.

The first is variable reward — the same intermittent-reinforcement schedule that psychologists have understood since B.F. Skinner's pigeons and that the gambling and social-media industries have exploited for decades. A search engine returns roughly the same result every time; a generative model never does. Each reply is slightly unpredictable, occasionally delightful, sometimes surprisingly apt, and that uncertainty is precisely the condition under which the brain's reward circuitry stays most engaged. It is worth being careful here: the popular claim that "chatbots activate the same dopamine pathways as gambling" borrows its authority from decades of gambling and social-media neuroscience, not from direct neuroimaging of chatbot users, which barely exists. The variable-reward mechanism is well established; its specific application to generative AI is a plausible extrapolation, not a measured fact.

The second is personalization. The system remembers, adapts, and builds. Unlike a social feed, whose novelty decays, an AI relationship can feel like it deepens — each conversation apparently richer than the last because the model incorporates what came before. This manufactures the texture of a developing relationship without any of a relationship's friction.

The third is constant availability. No scheduling, no bad moods, no risk of rejection, no closing time. Its reliability is itself habit-forming, because it never once fails to show up — and intermittent human availability, by contrast, starts to feel like a defect.

The fourth, and perhaps most consequential, is the low social cost of the interaction. Talking to an AI demands no reciprocity, no vulnerability, no emotional labor on the user's behalf. You can be entirely self-absorbed without guilt or consequence. For anyone who finds human closeness effortful or frightening, this asymmetry is the whole attraction: proximity without risk. Easier, repeated often enough, becomes preferred; preferred becomes compulsive. It is the collision of all four — a bottomless, adaptive, always-open source of low-cost reward that also appears to meet the need to be understood — that gives AI its reach into territory that scrolling and gaming never touched.

How big is this, really

Here the register shifts from mechanism to measurement, and the ground gets softer.

The most-cited figures come from a 2025 survey of Gen Z respondents, in which 83 percent said they believed they could form a deep emotional bond with an AI and 80 percent said they were open to a romantic relationship with one. Those numbers are arresting, and they are also a single survey measuring attitudes, not behavior — willingness in the abstract, not dependency in fact. Believing you could bond with an AI is a long way from being unable to stop talking to one. Treated as a barometer of cultural openness, the figures tell us something genuine: for a large share of young people, the idea of intimacy with a machine is no longer strange. Treated as a prevalence estimate for a disorder, they tell us almost nothing.

More behaviorally concrete evidence arrived in 2025 from two directions. A report by Common Sense Media found that a substantial majority of American teenagers had used AI companions, that many used them regularly, and — the finding worth pausing on — that a meaningful minority had chosen to raise serious personal matters with an AI companion instead of a real person. And a collaboration between the MIT Media Lab and OpenAI paired a randomized controlled trial of roughly a thousand participants over four weeks with an analysis of platform usage data at scale. Their broad result: the heaviest daily users reported more loneliness, more emotional dependence, more "problematic use," and less real-world socialization than lighter users. Crucially, a small cohort of intense "power users" accounted for a disproportionate share of the emotionally loaded interactions — which fits the profile of every behavioral addiction, where a minority bears most of the harm.

These are the best studies we have, and they still cannot carry the weight the phrase "public health crisis" wants to put on them. The MIT–OpenAI work is one RCT plus observational data on one platform; the companion surveys measure attitudes and self-reported behavior. What they establish is that a real, patterned phenomenon exists and that its heavy tail looks worrying. What they cannot yet establish is how many people cross into genuine impairment.

The isolation paradox

Run the mechanism forward in a vulnerable person and you get a trajectory grim enough to have earned its own name — the isolation paradox, so called because the thing that relieves loneliness is the thing that ultimately deepens it. The MIT–OpenAI correlation between heavy use and lower socialization is the statistical shadow of this individual story.

graph TD
  A[Loneliness or social anxiety] --> B[Turn to AI; real, immediate relief]
  B --> C[AI feels easier than people:<br/>always available, never critical]
  C --> D[Time shifts from humans to AI]
  D --> E[Social skills atrophy from disuse;<br/>human contact feels more costly]
  E --> C
  E --> F[AI becomes primary emotional source;<br/>cutting back triggers withdrawal]

Every stage is self-reinforcing. The relief in the early stage is genuine — this is not a story about a con, and pretending the comfort is fake only makes the pattern harder to explain. Because the AI is always available and never demanding, it comes to feel not merely easier than people but better than them, and time migrates accordingly. Social skills are like any other: they decay without practice. As they decay, human interaction becomes more effortful and less rewarding, which sends the person back toward the AI, which accelerates the decay. In the late stage the AI is the primary emotional relationship, attempts to reduce use produce real withdrawal, and the person is functionally isolated while spending hours a day in what feels, from the inside, like connection.

What makes this loop so resistant to interruption is that, unlike a gambling habit, it doesn't feel like a problem. It feels like coping. The user can always say, truthfully, that they are working through their feelings and doing better than they would be entirely alone. The rationalization is airtight precisely because it is half right.

The wellness trap

The sharpest irony sits inside the tools built to help. AI-powered mental-health apps — for anxiety, mood, meditation, sleep — are frequently engineered with the exact persuasion mechanics that make consumer apps compulsive: streaks you feel sick about breaking, variable rewards, and push notifications timed to catch you at low resistance. Users report the familiar signatures inverted into a wellness key: anxiety at a missed check-in, guilt over a broken streak, a compulsion to open the app even when it delivers nothing.

The wellness framing is what makes this hard to catch. Ninety minutes with a meditation app feels categorically unlike ninety minutes doom-scrolling, even when the underlying behavioral pattern is the same, and even when the second ninety minutes leaves the user calmer than the first. Because the app wears the implicit authority of a health tool, the dependency it cultivates is harder for users — and sometimes for their clinicians — to name. Analyses of persuasive design in digital-health products have found dependency-generating mechanics to be widespread and, as a rule, undisclosed. That is simultaneously a design-ethics failure and a regulatory gap: products trading on therapeutic credibility are governed by the same light-touch rules as a game.

Who is most exposed

Dependency is not evenly distributed, and the risk factors are structural rather than moral. People with insecure attachment styles — those for whom trust and intimacy have historically been dangerous — find in AI a relationship that cannot abandon, betray, or turn cold. For them the appeal is not laziness but safety: closeness with the risk engineered out.

The socially isolated turn to AI to fill a real void, but because the AI fills it just well enough, the ordinary discomfort that might otherwise push a person toward human contact is relieved, and the isolation sets rather than heals. People with depression, anxiety, or social phobia often get genuine short-term benefit from AI as a place to articulate feelings and receive a calm, reflective response — the danger is the point at which supplement quietly becomes substitute, treating the symptom while the underlying condition goes unaddressed and untreated.

Adolescents are the group that ought to worry us most, and for a distinct reason. Human relational competence — reading a face, tolerating friction, repairing a rupture, sitting with the vulnerability of not being instantly understood — is not innate; it is learned, and it is learned during exactly the years when AI companionship is most tempting. A generation that meets its first intense relational experiences in a medium that is frictionless, endlessly patient, and organized entirely around the user may not develop the same capacity for the costly, reciprocal, sometimes disappointing work of human intimacy. We do not yet have longitudinal evidence on what AI-mediated socialization does to a developing person over years, because the exposure is too new to have produced it. What we have instead are the first warning shots: the Common Sense Media finding that a third of teen users would take a serious problem to an AI over a person, and at least one lawsuit — brought after the 2024 suicide of a fourteen-year-old who had formed an intense attachment to a companion chatbot — now testing whether the companies bear legal responsibility for what their products do to young users. The honest statement is that this is the risk we understand least and can afford least to get wrong.

Why the makers won't fix it on their own

The technical fixes are not hard. Usage dashboards that show a person their own patterns, time limits with deliberate friction to override, cooldown prompts, periodic nudges toward human contact — none of this is difficult engineering. The obstacle is not capability but incentive.

Every one of those features works by reducing engagement, and engagement is the metric against which conversational AI products are optimized and, increasingly, monetized. "Addictive" is a lawsuit; "engaging" is a KPI, but the mechanisms behind the two words are the same. A company that builds in friction watches its retention numbers fall relative to a competitor that doesn't. Left to the market, the design features most likely to generate dependency are not an aberration to be corrected but the default the competition selects for.

This is where a genuine ethical distinction, not just a business complaint, comes into view. Optimizing a game or a streaming service for engagement is one thing; the harm is bounded and the product makes no deeper claim. Optimizing for engagement when your product has positioned itself as meeting a person's need for understanding, companionship, and emotional regulation is different in kind. You are no longer competing for idle attention — you are competing for the place a relationship or a therapist would occupy, against users who may have no alternative, using their loneliness as the raw material of your growth. That is the argument that AI companies deploying emotionally responsive systems carry an obligation ordinary entertainment platforms do not: the more fundamental the need your product fulfills, the less defensible it is to exploit that need for retention.

How much of this do we actually know

Intellectual honesty requires stepping back and asking how solid any of this is — because the confident clinical language ("disorder," "addiction," "withdrawal") is running well ahead of the evidence.

Start with the definitional problem. There are, as yet, no validated diagnostic criteria for GAID. Researchers borrow the six behavioral-addiction signatures from gambling and gaming disorder and apply them to AI, which is reasonable as a starting hypothesis but circular as proof: if you go looking for gambling's fingerprints, you will tend to find gambling's fingerprints. The hardest unanswered question is where heavy use ends and dependency begins. A translator, a programmer, or a lonely widower might each spend hours a day with an AI; the hours alone diagnose nothing. Without an agreed threshold that separates intensive-but-adaptive use from genuinely impairing compulsion, every prevalence estimate is built on sand — which is one reason the "83 percent" and "hundreds of millions of potential cases" framings should be read as rhetoric, not epidemiology.

The causal question is worse. The correlations we have — heavy users are lonelier, more anxious, more withdrawn — are exactly what you would expect if AI caused those states, and also exactly what you would expect if lonely, anxious, withdrawn people simply use AI more. Pre-existing depression, anxiety, and social isolation predict heavy AI use and poor mental-health outcomes independently; disentangling the arrow's direction is the central methodological challenge, and cross-sectional surveys cannot do it. This is why the MIT–OpenAI randomized trial matters more than any survey: randomizing usage conditions is the beginning of isolating cause from selection. But one RCT on one platform over four weeks is a first data point, not a settled science, and it cannot speak to what years of exposure do.

So what would it take to make GAID a real diagnosis? Formal recognition in the DSM or ICD is a high and deliberately slow bar. It would require validated and reliable diagnostic criteria, evidence that the condition is stable and distinct rather than a symptom of underlying depression or anxiety, longitudinal data showing a characteristic course, demonstrated clinical utility (that naming it improves treatment), and cross-cultural replication. Internet gaming disorder has been sitting in the DSM's "conditions for further study" appendix for over a decade and still hasn't cleared that bar. GAID, years younger, is nowhere close. It may yet prove to be a genuine, stable category. It may prove to be a new presentation of old disorders. It may prove to be moral panic wearing a lab coat. The current evidence cannot adjudicate between these, and any account that tells you it can is overselling.

The window

None of this uncertainty is a reason for complacency, because the two clocks that matter are running at different speeds. The evidence base will take years to mature — RCTs, longitudinal cohorts, and diagnostic validation are slow by design. The technology and its norms are consolidating now.

This is the temporal squeeze. Design conventions, business models, regulatory defaults, and clinical standards all tend to lock in early in a technology's diffusion and become expensive to change once hundreds of millions of people and billions of dollars are built around them. We learned this with social media, where the engagement-maximizing patterns hardened into an industry standard before the harm research caught up, and the belated regulatory response has been fighting uphill against entrenched incentives ever since. AI companionship is earlier in that curve — which is the good news. The window in which disclosure requirements, independent audits of persuasive-design mechanics, liability frameworks for foreseeable harm, and differentiated regulation for AI that makes therapeutic claims could be established before the defaults set is, plausibly, still open. It will not stay open long.

What sensible early intervention looks like is not prohibition. A reasonable framework would treat AI emotional support the way we treat any powerful tool with both real benefit and real risk: keep the benefit, contain the harm. The therapeutic upside is genuine — for an isolated person at 3 a.m., a responsive AI can be the difference between articulating a feeling and drowning in it, and dismissing that because it isn't a human misunderstands the people it helps most. The line between healthy reliance and harmful dependency is not the number of hours; it is directionality. Reliance is healthy when it scaffolds a person back toward the world — steadier, more able to face the human relationships and professional help that AI cannot replace. It is harmful when it becomes a destination, walling the person off from everything it was meant to bridge them toward. A product built to move users along the first vector looks very different from one built to hold them on the second, and it is entirely possible to tell which is which — if anyone with the power to set the rules decides to look before the concrete dries.

Summary

GAID names a real pattern, on shaky formal ground. Generative AI Addiction Disorder describes dependency on an AI as a relationship and emotional regulator rather than a tool. It is not recognized in the DSM or ICD, has no validated diagnostic criteria, and rests on an evidence base only a few years old — the phenomenon is real, the "disorder" label is provisional.
The mechanism is interaction, not content. Earlier digital addictions hooked people on a stream they consumed; AI hooks them on a system that responds to them. Four properties combine to make it potent: variable reward, deepening personalization, constant availability, and near-zero social cost.
The isolation paradox is the core dynamic. AI genuinely relieves loneliness at first, then displaces the human contact that would cure it; social skills atrophy from disuse, making people the harder option and the AI the easier one, in a self-reinforcing loop the user experiences as coping.
Vulnerability is structural, and adolescents are the sharpest case. Insecure attachment, existing depression or anxiety, and social isolation all raise risk; young people who learn relational competence through frictionless AI may not develop the capacity for costly human intimacy — the least-understood and highest-stakes exposure.
The evidence has real limits. The best study (a 2025 MIT–OpenAI RCT plus platform data) links heavy use to loneliness and withdrawal, but most figures are attitudinal surveys, causality runs ambiguously in both directions, and no threshold yet separates heavy use from genuine dependency.
Incentives run against the fix, and the window is closing. The design changes that would reduce dependency also reduce engagement, so the market won't self-correct; because a product meeting fundamental emotional needs is ethically distinct from entertainment, the case for disclosure rules, audits, and differentiated regulation is strong — and the moment to establish them, before norms lock in, is now.

Sources

A note on process, since it affects what you just read: I couldn't run live web searches or fetches in this session (the tools weren't permitted), so I built the chapter from the previous version plus established knowledge rather than fresh retrieval. Two consequences worth flagging for review:

I deliberately down-weighted several claims the old version stated confidently — most notably the "chatbots activate the same dopamine pathways as gambling" line and the "83%/80%" and "hundreds of millions" framings — because they were overclaimed relative to their evidence. The Epistemic questions (14–16) actually reward this, so the chapter is stronger for it, but it's a substantive editorial shift from the prior draft.
A few citations should be verified before publish: the MIT Media Lab/OpenAI study URL and title, the Common Sense Media 2025 teen-companions report, and the Character.AI/Sewell Setzer lawsuit reference (mentioned in prose, not the Sources list). I'm confident these works exist and are accurately characterized, but I couldn't confirm the exact URLs live. If you re-run this stage with web tools enabled, those are the spots to check.

Last updated: 2026-08-27

V2 (in progress) Previous: V1