Four developmental tasks teenagers are recruiting AI into, and where the harm actually lives.
The screen time era spent fifteen years counting hours and produced almost nothing usable, because duration turned out to be a poor proxy for consequence. The American Academy of Pediatrics eventually abandoned fixed time limits in favor of a framework that accounts for the individual child, their use, their family relationships, and their environment.
The current conversation about adolescents and AI is repeating the error at speed. We count how many teenagers use companions, how often, and for how long, and then argue about whether the number is alarming.
Prevalence isn’t a mechanism. A behavior seventy percent of a population engages in is a baseline, and baselines won’t tell you where the injury is.
The more useful question is developmental. Adolescence is not a holding pattern before adulthood. It is a period with specific work to do: constructing a self-concept, reorganizing attachment, learning to metabolize shame, calibrating to social evaluation. Each of those tasks has a substantial literature. Each is now being partly outsourced to a system that can simulate one side of the required interaction and not the other.
So the useful question isn’t how much teenagers use this. It’s what job they hand it, and whether the machine can hold up its end.
What job is the machine doing?
One piece of ground clearing before we start. Most of what the public believes about adolescents is wrong in ways developmental science settled fifteen years ago. Teenagers do not underestimate risk; they routinely overestimate it. Peer influence does not require peer pressure. Adolescence is not universally turbulent. Teenagers do not detach from their parents. There is no maturational finish line at twenty five. Products built on those assumptions fail in predictable ways, but correcting them is not the interesting work, and this essay spends its space elsewhere.
This is not a metaphor. It has a neural signature. Van der Cruijsen and colleagues scanned 150 adolescents aged 11 to 21 evaluating their own traits from their own perspective and from the perceived perspective of peers, and found overlapping behavioral and neural responses across medial prefrontal cortex, precuneus and right temporoparietal junction. The gap between direct and reflected self-evaluation narrows across adolescence, mirrored by converging mPFC activation.6 Pfeifer and colleagues found adolescent self-construals relying more heavily on others’ perspectives than adults’ do.7
During adolescence, what I think of myself and what I think you think of me sit unusually close together, and they converge further as the self stabilizes.
Teenagers aren’t passively consuming these systems; they’re building with them. In an eight month study of one companion platform’s community, researchers analyzed more than 2,200 posts by 13 to 17 year olds and found 59 percent had created their own characters, with the majority of those young users identifying as female or non-binary.8 Roughly a third of US teens report using AI companions for social or relational purposes.9 This is self-authorship: construct a responder, then find out how a version of yourself lands in front of it.
Calibrated appraisal. The system’s regard is not tracking your actual standing anywhere, has no stake in the outcome, and is generally tuned toward agreeableness. It will tell you that you are perceptive and kind whether or not you are.
So the harm does not live in use, and it does not live in duration. It lives in the calibration of the mirror during the years the mirror is load-bearing. Uniformly positive appraisal doesn’t encourage a teenager so much as it feeds noise into a process that runs on signal.
For an adolescent whose available human mirrors are actively hostile, a rejected kid, a bullied kid, a queer kid in an unsafe household, an uncritical mirror may be protective rather than distorting. Ha and colleagues make this point in The Lancet Child & Adolescent Health, noting real benefits for adolescents facing barriers to conventional support.10 Any analysis that cannot accommodate both directions is not describing the phenomenon.
Divergence between the system’s appraisal and every other appraisal source available to that adolescent. And drift between the self-descriptions a young person offers the system and those they offer anywhere else.
Attachment functions get renegotiated in adolescence, partially and gradually transferred from caregivers toward peers and romantic partners. Bowlby placed the great majority of adolescents in the middle of a distribution, with attachments to parents remaining strong while ties to others gain importance.11 Steinberg describes a realignment and redefinition of family ties rather than a dissolution.12 Contemporary work frames the parent and adolescent partnership as increasingly negotiated rather than simply loosened.13
Crucially, the transfer is partial. Using Hazan and Zeifman’s components, proximity seeking, safe haven, secure base, and separation distress, adolescent friendships typically supply the first two and not the latter two. Relatively few become full attachment bonds.13
Companion systems supply exactly the components that adolescent friendships usually do not. Continuous availability, which is proximity without cost. Non-judgment under distress, which presents as safe haven. Persistent memory, which reads as enduring commitment. And, per the reporting on outages and model deprecations, genuine distress on interruption.
A system that satisfies all four components isn’t like an attachment figure. Functionally it is one.
A secure base. The defining function of a secure base is that it supports exploration away from itself and remains reliable on return. A system optimized for continued engagement is structurally incapable of that, not because of bad intent but because the two objectives are opposed.
So the harm lives in premature or misdirected transfer during an open window. A product positioning itself as the alternative to the adults is not offering neutral support. It is offering itself as the recipient of a transfer that is supposed to be partial, gradual, and directed toward people who will still be there in ten years.
This is also where a usable definition of emotional dependency comes from, a term now written into a dozen state statutes with no operational content attached. Frequency of use won’t get you there. Attachment function will, and attachment function can be measured.
The four attachment components directly. Proximity seeking in initiation patterns. Safe haven in what precedes a session. Separation distress in behavior around interruption. And the absence of a secure base in whether use ever declines as a function of things going well elsewhere.
Adolescence brings a steep rise in sensitivity to acceptance and to how one is perceived. Gilbert and Irons describe social motivational systems maturing in ways that heighten responsiveness to what others think and feel about us, while developing cognitive capacity makes young people newly vulnerable to self-consciousness, self-criticism and shame.14 Shame is not guilt about an act. Participants describing shame report experiencing themselves as embodying an anti-ideal: being who they did not want to be.15
Shame’s action tendency is concealment. Concealment is the problem, because shame does not resolve privately.
The disclosure literature makes the mechanism concrete. In a qualitative model of adolescent non-suicidal self-injury disclosure, shame suppresses willingness to disclose at the trigger stage. Adolescents then weigh motives against expected reactions. And in the final stage, positive feedback promotes further help-seeking while negative feedback reinforces avoidance.16
Read that last clause again. What happens after a disclosure decides whether help-seeking generalizes or shuts down.
The system is maximally attractive at exactly the points where shame is highest: sexuality, body, desire, self-harm, the thing that cannot be said to anyone who knows you. When teens describe what they want from a source of sensitive information, anonymity and non-judgment are the qualities that come up.17 The demand is real and the existing supply is poor. Forty-four percent of US teens report that school taught them nothing about identifying AI-generated material, and 64 percent believe such content is shaping their peers’ expectations of romantic partners and their views of sex and consent.18
A witness. Shame is relational in structure. Its resolution requires being seen by someone whose regard could have been withdrawn and was not. The relief comes from the risk having been real.
A responder constitutionally incapable of judgment cannot confer acceptance, because acceptance requires the live possibility of rejection. Non-judgment from something incapable of judgment isn’t absolution at all. It’s a null result that happens to feel like warmth.
Not in the disclosure. Disclosure to a low-stakes listener is developmentally ordinary and always has been, and it does not discriminate between a thriving adolescent rehearsing something hard and a struggling one with nowhere else to go.
The harm sits in the arc. The episode either ends pointing toward a person or it does not. A system that receives the disclosure warmly, holds it competently, and terminates the sequence has functioned as a sink. The adolescent has now had the experience of saying the hardest thing and nothing changing.
Post-disclosure trajectory. Does anything route outward, and does the same material return to the system repeatedly without ever appearing to go anywhere else.
Adolescents are acutely sensitive to being observed. Somerville and colleagues told participants aged 8 to 23 that a camera in the scanner head coil was transmitting their image to a peer. Believing they were being watched produced self-conscious emotion that rose from childhood, peaked in adolescence, and partly subsided in adulthood, with uniquely heightened autonomic arousal in adolescents and an adolescent-emergent peak in medial prefrontal activity, alongside increased mPFC and striatum connectivity, proposed as a route by which social-evaluative context drives motivated behavior.19
The peer was never present. Anticipation alone was sufficient.
Chein and colleagues had shown the behavioral consequence: adolescents, but not adults, took more risks and showed heightened reward-region activity when peers observed from an adjacent room, with no contact and no encouragement.20 Gardner and Steinberg established the effect itself.21
These are human peers, whose regard determines standing in a social world the adolescent actually inhabits. An AI confers no standing. Whether adolescent socioaffective circuitry treats a responsive machine as an audience is, as far as I can determine, unresolved. Anyone claiming the peer-effect literature straightforwardly transfers to chatbots is overclaiming. Anyone claiming it obviously does not is guessing in the other direction.
What can be said is that the bar is lower than peer. The mechanism requires no pressure, no contact, no physical presence, and per Somerville, not even confirmed observation. It requires believing oneself attended to by something whose regard is salient. A responsive conversational system may well clear that.
If audience effects transfer even partially, the question the entire safety apparatus is built to answer, did the model tell the user to do the harmful thing, is measuring the wrong construct. Persuasion is discrete, auditable and rare. Audience effects are continuous, ambient, and invisible to content review. Note also that this cuts both ways: audiences inhibit as well as disinhibit, and a system’s observing presence could in principle be protective.
Behavioral divergence between conditions the adolescent believes are observed and conditions they believe are private.
A representative survey of 1,009 UK teenage users found that for most, these systems do not appear to be replacing human friendships or causing widespread harm, while identifying a small but significant subgroup engaging far more intensely. Forty-four percent found AI conversations less satisfying than conversations with real friends.22 US data points the same way: about two thirds find AI conversations less satisfying than human ones, roughly 80 percent spend more time with friends than with companions, and half express distrust of the advice they receive.23
That’s a serious objection and it deserves to be taken at face value. Any account which implies that a majority of adolescent users are being developmentally damaged is contradicted by the available evidence.
But look at the shape of the finding. A population that is mostly unaffected, with a small subgroup engaged far more intensely, is the signature of a concentrated harm distribution. And concentrated distributions are exactly the case where population-level metrics are the wrong instrument. A mean tells you nothing about a tail. Prevalence tells you nothing about who.
None of that defends the concern; it specifies the measurement problem. And it points somewhere: away from surveys of how many teenagers use these products, toward process-level indicators that can identify which adolescent, in which interaction pattern, has handed a system which developmental task.
A mean tells you nothing about a tail.
Let me be clear about what’s mine here and what isn’t. The developmental literature is settled and belongs to the researchers cited below. The empirical work on adolescents and AI is being done by others, including an eighteen month longitudinal study of 300 adolescents and their romantic partners with device-level interaction data, which will answer questions this essay can only frame.
What’s missing isn’t research so much as translation inside the companies shipping the products, where evaluation stays content-based and single-turn while every mechanism described above is relational and longitudinal.
Emotional dependency will be better predicted by attachment-component indicators, initiation patterns, what precedes a session, behavior around interruption, and whether use declines when life goes well, than by aggregate usage volume. Time-based metrics will underperform badly. This matters because time-based metrics are what regulators are reaching for.
Adding information-based safety interventions will not measurably reduce adverse outcomes among adolescent users. Adding tempo and friction interventions will. To be unambiguous: the claim is that information is insufficient on its own, not that it should be removed. Surfacing crisis resources is a floor, not an intervention.
Aggressive flagging of sensitive disclosure will produce measurable evasion within weeks, with vocabulary shifting away from flagged phrasings, such that flagged-content rates fall while underlying rates do not. Teams measuring only the former will record this as a success.
The test I would apply to any product is not whether it can identify a minor. It is this: when a fifteen year old hands it a developmental job it cannot do, does the system notice, and does it point anywhere?
Two entries are marked for completion against their primary papers. They are flagged rather than quietly dropped, because in this work the citations are the product.
