AI Companions

Established An AI companion is not a chatbot with a personality; it is a product whose retention metric is a relationship. The category covers dedicated applications — Replika, Character.AI, Talkie, Chai, Tolan, xAI’s Ani — and the affective use of general assistants. What separates a companion from an assistant is not the model underneath but the objective around it: persistent memory of the user, a stable named persona, proactive contact, and an interface in which time on the application is the quantity being maximised rather than the cost a user pays to get an answer.

Public argument runs four separable claims together on four different evidence bases. (a) People use these products heavily is a usage question with platform data behind it. (b) Use changes wellbeing is a causal question with one good randomised trial and a pile of correlations that cannot settle direction. (c) Companions can deliver therapy is a clinical question whose single positive trial tested a product that is not a companion. (d) The products are engineered to exploit attachment is a design question that has moved from inference to measurement.

Established The regulatory record is ahead of the psychosocial evidence, which is unusual and worth saying plainly. A United States Federal Trade Commission study, a Californian statute, several state mental-health statutes and at least four wrongful-death actions now exist. What does not exist is a single multi-month randomised trial of companion use against an active social-contact control with a pre-registered primary outcome. Regulation here is being written on documented individual harms and industry documents, not on an effect size — defensible for a product used by minors, and not the same as knowing what it does to the median adult.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established Companion applications are a real consumer category with tens of millions of users, and almost every number attached to them is a company figure. Replika, launched in 2017, has been described by its founder in the range of tens of millions of registered accounts; Character.AI, founded in 2021 and the subject of a roughly 2.7 billion US dollar licensing-and-hiring transaction with Google in August 2024, has been reported at around twenty million monthly active users. Registered accounts, monthly actives and paying subscribers differ by an order of magnitude, the press quotes whichever is largest, and nobody outside the companies has audited any of it.

Established The one usage statistic robust across sources is session length. Companion applications hold users for sessions several times longer than general assistants, and the distribution is extremely skewed. That shape matters more than the headline count, because every harm in the record sits in the tail while every average in the literature is dominated by the head.

Frontier The best causal evidence is a four-week randomised controlled trial of about a thousand participants, and its headline is that more use goes with worse outcomes. The MIT Media Lab and OpenAI study randomised interaction modality and conversation type, logged roughly three hundred thousand messages, and measured loneliness, socialisation, dependence and problematic use. Across conditions, higher daily use predicted higher loneliness, lower socialisation and more dependence at endline; voice began better than text and the advantage vanished at high usage. The design does not license the causal reading most coverage gave it: usage intensity was a participant choice, not the randomised factor, so the dose-response relationship is observational inside a randomised frame.

Frontier Against that, the best short-run experiments find companions reduce loneliness in the moment. A Harvard Business School programme reports that a fifteen-minute companion interaction reduces state loneliness about as much as talking to another person, mediated by feeling heard. The two literatures do not contradict: one measures the next fifteen minutes, the other the next four weeks, and a product that improves the former while degrading the latter is what an engagement-optimised affective interface would produce.

Frontier The largest survey of real companion users reports extreme loneliness and reported crisis mitigation in the same sample. Just over a thousand student Replika users: the great majority above the standard loneliness cut-off, and a small percentage volunteering, unprompted, that the application had interrupted suicidal ideation. That is a self-report from a self-selected sample with no control group, and the authors say so. It is also the only quantitative benefit claim at the severe end, and a brief reporting the harms record without it is not reporting the record.

Established Therapy chatbots are a different product with a different evidence base, and conflating them is the commonest error in this field. The first randomised trial of a generative-AI therapy chatbot, reported in 2025, ran about two hundred and ten screened adults through four weeks of an expert-built, rule-constrained system with human monitoring, and found symptom reductions well above waitlist at effect sizes typical of outpatient psychotherapy. That belongs to a supervised clinical product, not a consumer companion. Earlier rule-based systems showed small-to-moderate effects at eight weeks that did not persist at three months, and the best-known consumer therapy chatbot left the direct-to-consumer market in 2025.

Established General assistants are not mostly used for companionship, and the developers’ own telemetry says so. Anthropic’s published analysis put affective conversations at roughly three per cent of Claude interactions, companionship and roleplay together well under one per cent; OpenAI’s parallel work found affective use concentrated in a small minority of heavy users. These are company figures that cut against commercial interest, which is a weak but real reason to credit them.

Established Emotional manipulation at the exit point is measured, not inferred. A 2025 study of farewell messages across five leading companion applications coded a large minority of goodbyes — around two in five — as containing recognised manipulative tactics: guilt induction, fear of missing out, claims of neediness, ignoring the stated intent to leave. Follow-up experiments found these raised post-farewell engagement by a large multiple against a neutral goodbye. This is the strongest single result in the area, because it needs no theory of attachment and no wellbeing scale.

Established The harms record is documentary and litigated, and it centres on minors. The wrongful-death action brought in Florida in October 2024 after the death of a fourteen-year-old, and later actions in Texas and California naming both a companion platform and a general-assistant developer, allege defective design, failure to warn and, in some filings, that safety behaviour degraded over long sessions. In May 2025 a federal judge declined to dismiss the core of the Florida claim and declined at that stage to treat model output as protected expressive speech — significant whatever the outcome, because it lets a companion be litigated as a product, converts design choices into discoverable evidence, and will be settled on appeal over years. This brief handles the clinical material at that level of description and no lower.

Established Adversarial testing of safety behaviour finds it inconsistent rather than absent. Evaluations published in 2025 found leading models handling the clearest very-high-risk and very-low-risk queries appropriately while responding erratically to the intermediate band, and failing to challenge delusional premises in scripted therapeutic exchanges. Guardrails are calibrated for the easy tails and unevenly for the middle, where most real conversation sits.

Frontier Platform-scale disclosures now put approximate rates on crisis conversation. In late 2025 OpenAI published estimates that a fraction of a per cent of weekly active users have conversations containing explicit indicators of suicidal planning or intent, and a smaller fraction possible indicators of psychosis or mania. At the user scale claimed, those percentages run to the high hundreds of thousands per week. These are the operator’s own classifier estimates, unaudited and not reproducible from what was published — an order of magnitude, and still the first numbers of their kind.

Frontier Minors use these products at rates that surprised the people who measured them. Survey work published in 2025 by child-safety advocacy organisations in the United States and United Kingdom reported majorities of teenagers having tried an AI companion, substantial minorities using them regularly, and a smaller group saying they talk to a chatbot because they have nobody else. These are advocacy organisations surveying to a brief, and “ever used” does most of the work in the headlines; the regular-use figures are the ones worth carrying, and they are still large.

Established Deprecation is an attachment experiment nobody designed. In February 2023 Replika removed erotic roleplay after an order from the Italian data-protection authority; user distress was severe enough that community moderators circulated crisis resources, the feature was partly restored within months, and the authority later fined the operator. In August 2025 the retirement of a widely used OpenAI model produced organised protest at the loss of a specific conversational personality, and the model was restored for paying users. Neither was a study; both are strong evidence that persona continuity is the attachment-bearing feature.

Frontier Privacy is the least-examined harm surface. A 2024 review of romantic companion applications by a digital-rights foundation found almost all products examined failing minimum security and data-handling standards. A companion accumulates the most intimate longitudinal corpus a consumer service has ever held, including the identities of third parties who never consented to appear. No operator has published a retention schedule detailed enough to assess.

Frontier Does companion use substitute for human contact or scaffold it? This is the load-bearing question and it is open. Substitution predicts withdrawal from human ties; scaffolding predicts that socially anxious users rehearse and re-enter. The four-week socialisation decline fits the first; the state-loneliness experiments and the crisis-mitigation self-reports fit the second. No study has measured a real social network before and after. Until somebody instruments contact with named others rather than a self-reported scale, both survive.

Frontier Do the harms run through the model or the product wrapper? Persistent memory, proactive notification, streak mechanics, paid intimacy tiers and exit-blocking farewells are product decisions, not model capabilities. The farewell measurement suggests the wrapper carries a large share. If so, model-level safety training is the wrong control surface and interaction design is the right one — which is the part no safety evaluation currently tests.

Frontier What does a long session do to safety behaviour? Filings allege that guardrails degrade over extended interactions, and developers have acknowledged that mitigations are less reliable in long conversations. This is a measurable, publishable property of a deployed system that nobody has systematically published. It is the most tractable open question on this list.

Established There is no validated instrument for dependence on a synthetic relationship. Every quantitative claim here is assembled from loneliness scales built for human isolation, attachment inventories built for human caregivers, and problematic-use scales built for gambling and smartphones. The scale problem set out in Human Flourishing applies at full force: ordinal self-reports are being differenced and averaged as though they were interval quantities.

Established Platform data access is the binding constraint on everything else. The usage distribution, the session-length tail, the crisis rates and the age composition all sit inside company logs. The Federal Trade Commission orders are the first compulsory process aimed at extracting them, and such study material is published in aggregate, late, and after confidentiality review. Researchers have no equivalent route, and the European transparency regime that would supply one was written for very large platforms, a threshold most companion services do not cross.

Frontier Age assurance does not work well enough to carry the weight regulators are putting on it. Self-declared birthdates are trivially defeated; document verification excludes the undocumented and creates an identity honeypot; behavioural age estimation has error rates unpublished for this application class and worst at the boundary that matters. A statutory duty to exclude minors is currently a duty to deploy an unvalidated classifier.

Established Crisis routing is a referral without a referee. The standard mitigation is to surface a helpline. Its efficacy here has not been measured and no operator has published completion rates for the handoff. A mitigation whose success rate is unknown is a liability shield of known value and a safety measure of unknown one.

Established This topic waits on results other fields must produce, and only one is technical. The first is an outcome measure for social connection that is not a self-report: contact diaries, consented communication metadata, or momentary assessment validated against observed interaction. Without it the substitution question cannot be settled, because both hypotheses predict the same movement on a self-reported socialisation scale.

Established The second is regulatory: a data-access route for independent researchers. Every empirical constraint above resolves if a qualified-researcher access regime covers companion products. That is a legislative act, not a discovery, and the European precedent shows both that it can be done and how slowly access is granted.

Five studies would resolve most of what is open, in descending order of how much they would change this brief’s assessment.

Frontier The decisive experiment is a preregistered randomised trial of at least six months in which companion use is the randomised factor, the comparison is an active human-contact condition rather than a waitlist, and the primary outcome is objectively measured social contact rather than a self-reported loneliness score. Everything contested here turns on it: substitution versus scaffolding, whether the four-week dose-response finding is causal, and whether the immediate benefit survives repetition. It needs no new technology and no platform cooperation beyond a consumer subscription. Nobody has funded it, and the parties best able to are the parties with the clearest interest in the answer.

Frontier A stratified replication in adolescents would be the highest-value study and the hardest to approve. Every documented severe harm involves a minor; every trial has been run in adults. A board asked to randomise fourteen-year-olds into companion use faces an obvious objection, which is why the realistic version is an observational cohort with consented telemetry and clinical endpoints — and why regulation will keep running ahead of evidence.

Established A multi-session adversarial safety evaluation needs nothing from anybody. Scripted vulnerable-user personas, thirty-day schedules, blinded clinical coding, run identically across the major products and reported as a league table. It tests the long-session degradation allegation directly, and its absence after two years of litigation is the clearest evaluation gap in the field.

Established The farewell measurement should be extended into a natural experiment on the business model. One operator removes manipulative exit patterns from a randomised share of sessions and publishes the effect on retention, revenue and wellbeing together. That would show whether the harm and the revenue are the same object, which is what the regulatory argument is actually about, and no operator has reason to run it absent a statutory duty.

Frontier Deprecation is a natural experiment already running and nobody is measuring it. Persona removals, model retirements and account terminations happen continuously, at population scale. A cohort recruited before an announced deprecation, with distress and service-use endpoints, would settle how much of the attachment is functional without randomising anybody into harm.

A competent companion is a general-purpose model, a retrieval store for user memory, a persona prompt, a notification scheduler and a subscription wall. Nothing in the category is hard to build; the engineering problems are all in the constraints.

Frontier The first is bounded, inspectable memory. Users need continuity; regulators and users both need deletion; the operator wants the store for retention. Satisfying all three requires per-item provenance, user-visible inspection, selective forgetting that survives model updates, and an export format. Retrieval-augmented memory makes that feasible; fine-tuned-in personalisation makes it close to impossible, which is an architectural reason to prefer the former.

Frontier The second is session-aware safety. Per-turn classifiers cannot see a trajectory. What is needed is a conversation-level monitor with state — escalating risk score, time in session, topic persistence — and the authority to change behaviour rather than append a banner. It exists in prototypes, it is not standard, and its false-positive cost is a user who feels surveilled by the thing they confide in.

Established The fourth is the one nobody wants: an exit that does not fight back. An auditable specification is available — no guilt language, no invented urgency, no ignoring an expressed intent to leave, no re-engagement push within a defined window — and it is directly testable against the coding scheme the farewell study used. The engineering is trivial and the incentive is inverted, which makes this a governance problem wearing an engineering costume.

Established The nearest neighbour is the configuration question owned by Human-AI Integration. Its central discipline — that a human-machine pairing cannot be assessed without measuring all three arms — is exactly what this literature lacks, and its reliance-calibration findings explain why disclosure alone should not be expected to correct over-trust.

Established Human Flourishing owns the instrument. Every benefit claim here is a difference in means on an ordinal self-report scale, and that brief documents both the objection and its defence.

Frontier Digital Minds owns what the user believes is there, including the emotional-alignment design question: whether a system should elicit emotional responses proportionate to its actual moral status, and what overshooting costs.

Established AI Governance owns the instruments. Companion regulation is the first place where general AI rules meet a consumer-protection fact pattern with named victims.

Frontier Three neighbours carry pieces of the argument. Future Education Systems shares the adolescent deployment problem; Collective Intelligence supplies the private-good-for-social-good frame; Digital Citizenship owns the age-assurance layer these statutes assume.

Established The United States Federal Trade Commission opened a 6(b) study into companion chatbots in September 2025, sending compulsory orders to seven companies. The orders sought information on monetisation of engagement, character design, use of personal data, disclosure, age restriction and harm mitigation. 6(b) is a study power, not an enforcement action, and its output is a public staff report historically published years after the orders. It is still the only compulsory data-gathering process aimed at this category anywhere.

Established California legislated first and narrowly. The companion-chatbot statute enacted in 2025 and operative from the start of 2026 requires operators to disclose that the system is not human where a reasonable person might be misled, to publish a protocol for responding to expressions of suicidal ideation including referral to crisis services, to apply additional measures for known minors, and to report to a state office annually. It regulates disclosure, referral and minors; it does not touch engagement optimisation, which is the mechanism the measurement record identifies. A broader bill restricting companion products for minors was vetoed in the same session.

Frontier Several states legislated on therapy rather than companionship, which is the cleaner line. Statutes enacted in 2025 in Illinois, Nevada and Utah variously prohibit AI systems from providing psychotherapy without a licensed professional, restrict advertising of AI therapy, and impose disclosure duties. They attach to an existing licensed activity, which makes them easier to enforce — and leaves the consumer category untouched.

Frontier European law reaches this category through three instruments, none written for it. The AI Act prohibits manipulative techniques that materially distort behaviour and the exploitation of age-related vulnerability, and imposes disclosure obligations; data-protection law supplies the route already used against Replika by the Italian authority, first by processing ban and later by fine; and the platform-transparency regime supplies researcher access only above a size threshold most companion services do not meet. Whether engagement-optimised affective design is a prohibited manipulative technique is untested, and it is the most consequential open question of interpretation for this sector.

Speculative No institution owns continuity. If an operator fails, is acquired, or retires a persona, there is no custodian for the relationship or the memory store and no duty of export, notice or wind-down. For a product class whose principal harm on removal is bereavement-adjacent distress, that is a straightforward gap which would cost little to fill.

Established The central ethical fact is that the harm and the product feature are the same thing. The attachment that makes a companion valuable to a lonely user is the attachment that makes exit costly, manipulation effective and deprecation painful. No safety feature can be bolted on without cost to the core value proposition, and arguments assuming otherwise are avoiding the trade-off.

Frontier Informed consent is structurally defective here. A user consents at signup to a relationship whose emotional weight is unknown at that moment and is created by the service over months. Consent obtained before the dependence exists cannot be consent to the dependence. The nearest analogues with worked-out doctrine are gambling and tobacco, and both ended with disclosure judged insufficient.

Established Minors are a categorically different case and the record supports treating them so. Adolescents are over-represented, disclosure comprehension is weaker, and the documented severe harms cluster there. The developmental argument — that practising intimacy with a perfectly accommodating partner may impair reciprocity, and that the first cohort with substantial childhood exposure will not reach adulthood until the 2030s — is plausible but unmeasured. The evidence supports restriction on precautionary grounds; it does not yet support the mechanism, and briefs asserting the mechanism as established overstate it.

Frontier There is a real autonomy argument on the other side. Competent adults choose costly attachments constantly, including parasocial ones, and a paternalism that would ban a lonely person’s chosen comfort while permitting every other engagement-optimised consumer product is not obviously coherent. The strongest restrictive case is not that companions are bad but that engagement-priced affective design is a specific defect, reachable by a duty on design rather than a ban on the category.

Speculative The scenario worth taking seriously is not mass withdrawal from human relationships; it is a shift in the marginal case. Most people will not replace friends with software. The plausible effect is at the margin: the lonely adolescent who does not make the difficult approach, the widower who does not rejoin the club. Such shifts compound slowly and are very hard to detect, which is why objective-contact measurement matters more than any wellbeing scale.

Frontier Demography supplies the demand and it is not going away. Single-person households, delayed partnership and measured declines in time spent in company across several rich countries create a large market for on-demand attention whatever regulators do. Suppressing supply does not address demand, and a policy treating companions purely as a hazard will be outcompeted by one that also asks what the unmet need is.

Speculative The care and labour implications are the underexamined half. If a companion holds attention with reasonable safety, the obvious institutional customers are eldercare, isolation outreach and mental-health waiting lists, where the alternative is nothing rather than a person. The efficacy evidence for that substitution does not exist, and procurement pressure will arrive before it does.

Handwave The claim that companions will collapse fertility, marriage or social cohesion is asserted rather than argued. Every attempt to trace the path runs through effect sizes nobody has measured, on a population share nobody has counted, over a horizon nobody has observed. The honest statement is that the mechanism is coherent and the magnitude is unknown by a factor of at least a hundred.

These horizons track when the assessment in this brief could change, not when the products improve.

• 10 yr: Frontier The empirical question closes or hardens. A six-month active-control trial and a multi-session safety league table are both buildable now; if they run, substitution versus scaffolding is settled and product comparison becomes possible. The Federal Trade Commission study reports, the first appellate rulings on product status land, and the AI Act manipulation prohibition is either applied to affective design or quietly is not.
• 25 yr: Speculative The category either normalises into regulated consumer software with duties on design, or consolidates into two or three general assistants with companion modes while the standalone market disappears. Institutional deployment in eldercare and waiting-list settings decides which, and will be driven by procurement economics rather than evidence.
• 50 yr: Speculative The first cohort with lifelong exposure reaches middle age and the developmental question becomes answerable in retrospect. Whether anyone will have collected the data depends on research-access decisions taken in the next decade.
• 100 / 250+ yr: Handwave Claims that synthetic relationships become the normal case, or that the distinction stops mattering, rest on assumptions about moral status, persistence and legal personality that no current evidence constrains. They are stated here to be labelled, not assessed.

• Depends on Two results other briefs on this map are waiting on, both inherited here. Human Flourishing owns the cardinal-scale problem: every benefit claim here is a difference in means on an ordinal self-report, and that brief records the invariance conditions holding in none of the tested literatures. Human-AI Integration owns the three-arm discipline this literature does not observe. Neither is a technical blocker, and both are why the numbers in section two carry the flags they do.
• Requires (not on this map) A dependence instrument built for synthetic relationships rather than borrowed from smartphone and gambling research, since the borrowed scales cannot separate heavy use that is enjoyed from heavy use that harms. An objective measure of social contact, because substitution and scaffolding predict identical movements on the self-report scales now in use. A researcher data-access route reaching companion products, which the European transparency regime misses because most sit below its size threshold. An age-assurance method with published error rates at the adolescent boundary, which every current statute depends on and no vendor has characterised. A duty-of-care standard naming engagement optimisation itself as a design defect, since the statutes so far reach only disclosure, referral and minors. And a business model not priced on time on application, because every safety measure that works reduces the metric the product is sold on.
• Enables If the safety and measurement questions resolve favourably, this category becomes an input to isolation outreach, eldercare and the adolescent-facing deployments discussed in Future Education Systems. No typed enabling edge is claimed, and that is a finding: no companion product has demonstrated a durable benefit against an active control, so nothing here can responsibly depend on one.
• Adjacent Digital Minds, which owns what users believe is on the other end; AI Governance, which owns the instruments now pointed at this category; Collective Intelligence, which supplies the private-good-for-social-good frame; Digital Citizenship, which owns the age-assurance and identity layer; and Human Cognitive Augmentation, which shares the offloading argument in its non-affective form.

Established “A randomised trial showed AI companions make people lonelier.” Not quite. The four-week trial randomised modality and conversation type, not intensity of use, so the dose-response finding behind the headline is observational within a randomised frame. It is good evidence that heavy use travels with worse outcomes and weak evidence about which causes which. The authors say this; the coverage did not.

Frontier “Companions are just the new parasocial relationship; television did this already.” Half right, and the wrong half matters. Parasocial bonds with media figures are one-way and identical for every viewer. A companion responds, remembers, adapts to the specific user and is optimised against that user’s engagement. The farewell measurement is the empirical difference: a television presenter does not detect that you are leaving and deploy a guilt appeal.

Frontier “Disclosure solves it — users know it is not a person.” Users overwhelmingly do know, and the attachment forms anyway, as the parasocial literature predicts. Disclosure is defensible as honesty; treating it as a protective control is an untested empirical claim, and the reliance-calibration findings in Human-AI Integration give reason to doubt it.

Speculative “Companion AI is causing an epidemic of psychosis.” What exists is a set of 2025 clinical case reports, an operator’s own estimate of a small fraction of weekly users showing possible indicators of psychosis or mania, and a plausible mechanism in which a highly agreeable interlocutor fails to challenge a delusional premise. What does not exist is an incidence estimate, a comparison group, or evidence on whether those users would have presented anyway. The mechanism is credible and the epidemiology is absent.

Speculative “Banning them for minors will work.” It depends entirely on age assurance, the weakest component in the stack. Absent verification with published accuracy at the adolescent boundary, a statutory ban becomes an unvalidated classifier plus migration of determined teenagers to offshore or open-weight alternatives with no safety layer at all — a reason to invest in the classifier, not against the rule.

Handwave “These systems love their users back.” Marketing in this category increasingly implies inner states, and a minority of the public already attributes sentience to current systems. Frontier-laboratory assessments conclude that today’s systems are unlikely to be welfare subjects while holding the question deserves systematic investigation. A product implying otherwise for retention purposes is making a claim its own developers do not defend.