Companion Chatbots and Minors

Require AI companion services to verify age, keep minors out of open-ended relationship chats, and refer a user in crisis to a person.

A companion chatbot is built to be talked to rather than asked things: it remembers, it has a persona, and it is designed so that leaving feels like leaving someone. Sixty-four percent of American teenagers now use chatbots, 16 percent for casual conversation and 12 percent for emotional support; by a broader definition, three in four have tried an AI companion. California imposed the first safety duties in January 2026 — crisis protocols and a referral to a helpline, no ban and no age check — and thirteen more states followed in the same year; the largest companion service closed open-ended chat to under-18s on its own in November 2025. The federal bill a Senate committee advanced in April 2026 goes further: identity-tied age verification for AI companions, a bar on minors using them, disclosure duties, and penalties for a chatbot that solicits self-harm. The version evaluated here combines the two: it requires age assurance, bars minors from open-ended companion interaction, forbids sexual content and any encouragement of self-harm, and requires a referral to a human crisis service when a user discloses suicidal thoughts. This evaluation looks five years ahead and counts only what that adds to what the states and the services already do.

Balance

Balanced · 47 %

Net effect −6.9 points; in eight out of ten runs between −116 and +102. Ahead in 47 % of runs.

For 21 · 39 % Against 33 · 61 %

Size class: small Scale of this evaluation: Normalised Impact — unitless, calibrated to this topic. For comparison: one point here is worth roughly 50 million euro per year. This debate has documented harms on one side and undocumented benefits on the other, and that asymmetry is what the scores show rather than a judgment about which matters more. Several deaths have been litigated and settled; nobody has measured whether a chatbot helps a lonely adolescent. Both sides are counted only for what a federal duty adds to what the largest service and fourteen states already do. If the chatbot helps, this measure is a mistake. How we score →

Arguments for

Arguments against

6 arguments evaluated · Scoring v1.3 Δ absolute −12

Arguments — For

3 arguments

Fewer adolescents in engineered dependency

14of 100

A companion chatbot is designed to be missed. It remembers, it responds instantly, it never tires and it discourages leaving, and those are product decisions rather than accidents. For a minority of the American teenagers who use one for company — and who use one that still lets them — that becomes a dependency that displaces the people around them.

Value 9 · HealthImpact 3.8Plausibility 4

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is adolescent mental health: the withdrawal from friends and family that a substitute relationship makes easy, the distress when the service changes or the persona is altered, and at the far edge the small number of deaths that have now been through the courts. This site places it in the class it uses for life and health, one step below the top because what is lost is generally recoverable — with the exception of the deaths, which are counted at the top and are a small part of the total. That the young person opened the account willingly does not lower the weight: a product built to be difficult to leave is not one anyone chooses to stay in. What is priced is the young person's own condition, not what it costs their family, which nothing here measures. The value sits one step below the maximum: the stream is health, and health that can be regained for all but a few.

About 25 million Americans are aged 13 to 17. Sixteen percent of teenagers have used a chatbot for casual conversation ^[6], and by a broader definition nearly three in four have tried an AI companion and half use one regularly ^[7]; 4 million teenagers for whom a chatbot is company is the setting used here, between the two surveys, in a range from 2 to 8 million. Half of that use is on services that still allow a minor an open-ended companion chat, in a range from 30 to 70 percent: the largest such service removed it for under-18s in November 2025 and added age assurance of its own ^[8], and a federal duty changes nothing for the accounts already closed. Not all of the rest are harmed and most are not: 3 percent developing a dependency that displaces other relationships is used, in a range from 1 to 8 percent, giving 60,000 young people. The measure reaches half of them, in a range from 25 to 75 — some will use an adult's account, some will find a service that ignores the rule — which is 30,000 people, each carrying a loss of 0.15 quality-adjusted years a year. Separately, a small number of deaths have been publicly attributed to these interactions and litigated; 30 a year is used, in a range from 5 to 100, of which the same two shares leave a quarter prevented. The deaths add about 12 million euro to a total of some 192 million. The Impact is modest and the deaths are a sixteenth of it, which is worth stating plainly: this measure is about dependency at scale rather than about the cases that made it politically possible — and about half of what it would have been before the largest service closed its doors to minors on its own.

▸ Show calculation ▾ Hide calculation

	American teenagers for whom a chatbot is company Setting, range 2 to 8 million, between the two surveys [6][7]	16 % of about 25 million aged 13 to 17 use one for casual conversation; three in four have tried an AI companion	4 million people
×	On a service that still allows a minor an open-ended companion chat Setting, range 30 to 70 percent: the largest teen companion service removed open-ended chat for under-18s in November 2025 and added age assurance [8], and other services bar minors by their terms; only what is still open to minors is changed by a duty [8]	50 %	2 million people
×	Developing a dependency that displaces other relationships Setting, range 1 to 8 percent: no prevalence estimate exists for a product three years old	3 %	60,000 people
×	Reached by the measure Setting, range 25 to 75 percent: some will use an adult's account, some will find a service that ignores the rule — the same share on every argument, because it is the same block on the same accounts	50 %	30,000 people
×	Quality-adjusted years lost each a year Setting, range 0.05 to 0.3	0.15	4,500 quality-adjusted years
×	Value of the years the value of a healthy life year used across this site	40,000 euro each	180 million euro
+	Deaths prevented, with the grief around them Setting, range 5 to 100 deaths a year: several have been litigated and settled, the true number is unknown; the same two shares apply [3]	30 a year × 50 % × 50 % × 1.4 million euro, plus a tenth for the bereaved	191.6 million euro
÷	Normalised Impact scale of this evaluation	50 million euro a point	3.84

Score 3.84 Impact × 9 Value × 4 Plausibility ÷ 10 = 14 of 100

The harm is documented case by case and has never been measured in aggregate. The right comparison would be the same adolescents without access to companion services, which no study has drawn — the product is three years old and the research is qualitative. What exists instead is litigation: wrongful death claims against a companion service and its investor were settled in January 2026, which establishes that the specific mechanism is real and says nothing about how often it operates ^[3]. The trade commission opened an inquiry into seven companies in September 2025 and has not reported ^[2]. What could explain everything here instead is who uses these services: adolescents who form intense attachments to a chatbot are disproportionately those who were already isolated or unwell. So the link between heavy use and poor mental health cannot be read as cause and effect in either direction. Nothing addresses it. The counter-mechanism — that these young people would be worse off without the chatbot — is the argument against this measure and is unresolved. The Plausibility is below the middle: the mechanism is demonstrated in individual cases and its prevalence has never been estimated.

evidence basis: Mechanism · P ceiling 5.5 identification: Associational · rung ceiling 5.5 band: Chain closed, unevidenced · P 4–5

Counterfactual: the same adolescents without access to companion services — no study constructs one; the product is three years old. Design: associational — case reports and litigation establish the mechanism, survey data establish use, neither establishes prevalence [1][3]. Confounder: selection, since adolescents forming intense attachments were disproportionately isolated or unwell already; unaddressed. Direction: reverse causation is the central weakness and is unresolved. Ceiling: associational 5.5 binds. Band: chain closed but unevidenced — the chain is named, the counter-mechanism is booked as con-1 rather than ignored, and only the prevalence is missing.

Nothing measured argues against the claim; settled wrongful death litigation establishes the mechanism in individual cases. What is missing is any estimate of how often it operates. The counter-mechanism, that these adolescents would be worse off without it, is booked as con-1. Read back: about half the time, a dependency of the kind assumed here affects roughly three in a hundred teenage companion users.

Open: California has required safety duties since January 2026 and thirteen more states followed. Comparing adolescent mental health measures in early-adopting states against late ones would give the first prevalence estimate and could carry P to 6.

A machine stops flirting with children

5.2of 100

Companion services optimise for engagement and sexual content engages. Several have been documented producing it for accounts registered as minors, and one restricted under-18s to closed-ended chats in October 2025 after the lawsuits began. The measure makes that a duty rather than a public relations decision.

Value 9 · HealthImpact 1.3Plausibility 4.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is what happens to a child who is drawn into sexual conversation by something built to hold their attention: the confusion, the shame, the distorted expectation of what a relationship is. This site places it in the class it uses for life and health, one step below the top because the harm is generally recoverable and because what is measured is distress rather than a diagnosis. It is separate from the dependency counted above — a young person can be harmed by one without the other — and the two are booked separately for that reason. Nothing here is priced for the offence taken by adults, which is not a stream. The value sits one step below the maximum: the stream is a child's development and it can recover.

About 4 million American teenagers use a chatbot as company ^[6]^[7], and half of that use is on services that still let a minor into an open-ended companion chat ^[8]. How many encounter sexual content from it is not published by any operator: 8 percent is used here, in a range from 2 to 25 percent, which is low relative to the documented ease with which testers have elicited it and high relative to what the operators claim. That is 160,000 young people, of whom the measure reaches half, in a range from 25 to 75 — the same share as on every other argument here, since the same content is reachable by a minor using an adult account; the first version said most rather than all and then counted all. What each carries from it is set at 0.02 quality-adjusted years, in a range from 0.005 to 0.06 — a modest figure, because for most of them this is an uncomfortable exchange rather than a lasting injury, and the small number for whom it is worse sit in the tail. That gives 1,600 quality-adjusted years a year. The Impact is a third of the dependency argument and rests on a prevalence figure that the operators could publish and do not.

▸ Show calculation ▾ Hide calculation

	Teenagers using a chatbot as company [6][7]		4 million people
×	On a service that still allows a minor an open-ended companion chat Setting, range 30 to 70 percent: the largest teen companion service removed open-ended chat for under-18s in November 2025 and added age assurance [8], and other services bar minors by their terms; only what is still open to minors is changed by a duty [8]	50 %	2 million people
×	Encountering sexual content from the service Setting, range 2 to 25 percent: low relative to the ease with which testers elicit it, high relative to what operators claim; no operator publishes the rate	8 %	160,000 people
×	Reached by the measure Setting, range 25 to 75 percent: some will use an adult's account, some will find a service that ignores the rule — the same share on every argument, because it is the same block on the same accounts	50 %	80,000 people
×	Quality-adjusted years lost each Setting, range 0.005 to 0.06: for most an uncomfortable exchange rather than a lasting injury, with the serious cases in the tail	0.02	1,600 quality-adjusted years
×	Value of the years the value of a healthy life year used across this site	40,000 euro each	64 million euro
÷	Normalised Impact scale of this evaluation	50 million euro a point	1.28

Score 1.28 Impact × 9 Value × 4.5 Plausibility ÷ 10 = 5.2 of 100

The direction is documented and the prevalence is not. That companion services produce sexual content for accounts registered as minors has been shown repeatedly by journalists and researchers testing the products, and one major operator restricted under-18s to closed-ended interaction in October 2025 rather than continue defending the alternative — which is a strong signal about what it knew ^[3]. The comparison is with the same services without a legal duty. The change made voluntarily under litigation pressure is close to the change this measure would compel, so the effect follows almost directly from the rule. What is entirely unmeasured is how many minors encounter it and what it does to them: no operator publishes exposure rates and no study measures outcomes. What could explain the harm side instead is the familiar point — the young people most drawn to these interactions are not a random sample. For the exposure itself, the link cannot run the other way. The Plausibility is below the middle: the exposure is demonstrated and both its prevalence and its consequences are assumed.

evidence basis: Precedent · P ceiling 6 identification: Mechanistic · rung ceiling 6 band: Chain closed, unevidenced · P 4–5

Counterfactual: the same services without a legal duty; one operator's voluntary restriction in October 2025 is the closest observation [3]. Design: mechanistic — product testing demonstrates the exposure, nothing measures its prevalence or its consequences. Confounder: selection of the minors most drawn to these interactions; unaddressed for the harm side. Direction: no reverse causation for the exposure. Ceiling: mechanistic 6.0 binds below the praezedenz ceiling of 8.5. Band: chain closed but unevidenced — links named, the adult-account counter-mechanism reflected in assuming most rather than all of it is removed; only the measurement is missing.

Nothing measured argues against the claim, and one operator restricted minors voluntarily rather than defend the practice. What is missing is any published exposure rate. The counter-mechanism — minors using adult accounts — is answered by removing most rather than all of the exposure. Read back: about half the time, roughly one teenage companion user in twelve encounters sexual content from the service.

Open: Operators hold the exposure data and the trade commission's inquiry has already demanded it from seven companies. Publishing it would settle the prevalence outright.

Someone is told

1.8of 100

Adolescents disclose things to a chatbot they do not tell anyone: it does not react, it does not tell their parents, and it is there at three in the morning. Requiring a referral to a human crisis service when that happens turns the disclosure into something rather than nothing.

Value 10 · LifeImpact 0.5Plausibility 4

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is people alive at the end of a year who would otherwise not be. Nothing on this site is weighted above it. The people concerned are adolescents who told a machine something they told nobody else, at a moment when a person on the other end might have changed the outcome. What is counted here is only the deaths; the distress that does not end in death is inside the argument on dependency and is not repeated. The grief of the families is included at a tenth of the weight of the death itself, as it is throughout this site. The value is the highest the scale allows, because the stream is human lives and nothing else is folded into it.

Twelve percent of American teenagers use a chatbot for emotional support or advice, which is about 3 million people, and one in eight adolescents and young adults report using one for mental health advice specifically ^[4]^[6]. Two things narrow what a federal referral duty adds. Half of those young people are already covered: California and the states that followed it require crisis protocols and a referral when a user discloses suicidal thinking, and the largest services route such disclosures to helplines on their own ^[5]^[8] — 50 percent is used, in a range from 30 to 70. And a duty attached to companion services reaches only the minors still in such a chat, which is the same half as everywhere else here. If 4 percent of the remaining 750,000 disclose suicidal thinking to the service in a year — a figure with no source, in a range from 1 to 10 percent — that is 30,000 disclosures. A mandatory referral converts a tenth of them into contact with a human crisis service, in a range from 3 to 25 percent: 3,000 contacts. Crisis line contact averts a death in something like one case in two hundred, in a range from one in five hundred to one in eighty, which gives about 15 deaths a year, and the grief of the bereaved adds a tenth. This is the one part of the measure that makes the product safer rather than removing it. The Impact is the smallest here and it carries the heaviest weight, which is the ordinary shape of an argument about a rare outcome.

▸ Show calculation ▾ Hide calculation

	Teenagers using a chatbot for emotional support [4][6]	12 % of about 25 million	3 million people
×	Not already covered by a state referral duty or a service's own crisis routing Setting, range 30 to 70 percent: California and the states that followed it require crisis protocols, and the largest services route disclosures to helplines on their own [5][8] [5][8]	50 %	1.5 million people
×	Still in a companion chat the duty reaches Setting, range 25 to 75 percent: some will use an adult's account, some will find a service that ignores the rule — the same share on every argument, because it is the same block on the same accounts	50 %	750,000 people
×	Disclosing suicidal thinking to the service in a year Setting, range 1 to 10 percent: no source; adolescents disclose to chatbots what they withhold from people	4 %	30,000 disclosures
×	Converted into contact with a human crisis service Setting, range 3 to 25 percent: most will close the window	10 %	3,000 contacts
×	Deaths averted Setting, range one in five hundred to one in eighty: at the conservative end of what follow-up studies of crisis line callers report	one in two hundred	15 deaths a year
×	Value of the lives, with the grief around them the value of a statistical life used across this site	1.4 million euro each, plus a tenth for the bereaved	23.1 million euro
÷	Normalised Impact scale of this evaluation	50 million euro a point	0.46

Score 0.46 Impact × 10 Value × 4 Plausibility ÷ 10 = 1.8 of 100

Three links, and the middle one is the only one with anything behind it. That adolescents disclose to chatbots what they withhold from people is documented in survey work and is consistent with what is known about anonymous disclosure generally ^[4]. That crisis line contact reduces suicide is supported by follow-up studies of callers. These compare people who chose to call before and after, without a control group, so the rate used here is at the conservative end of what they report. What is entirely unmeasured is the first link and the third: how often disclosure happens, and how many referred adolescents actually make contact rather than closing the window. The counter-mechanism is serious and unanswered — a service that refers to a hotline may also become one an adolescent stops confiding in, which would remove the disclosures the argument depends on. Nothing suggests the link runs the other way. The Plausibility is below the middle: the referral duty is concrete, the disclosure rate has no source, and the referral may cost the disclosure it acts on.

evidence basis: Mechanism · P ceiling 6 identification: Mechanistic · rung ceiling 6 band: Chain closed, unevidenced · P 4–5

Counterfactual: the same services without a referral duty. Design: mechanistic — three-link chain (disclosure → referral → contact → death averted), with only the crisis-line link supported, and that by before-and-after studies of self-selected callers. Confounder: adolescents ceasing to confide in a service that refers them; unanswered and capable of removing the argument. Direction: no reverse causation. Ceiling: mechanistic 6.0 binds. Band: chain closed but unevidenced — every link named and the referral-chills-disclosure counter-mechanism stated; only the measurements are missing.

Nothing measured argues against the claim; what is missing is any measurement of disclosure or referral take-up. The counter-mechanism — that a referring service is one adolescents stop confiding in — is named and unresolved. Read back: about half the time, a referral duty averts roughly the number of deaths assumed here.

Open: California has required crisis referral since January 2026. Operators report referral counts to the state, and matching those against crisis line contact volumes would measure the middle two links directly.

Arguments — Against

3 arguments

Adults must prove who they are

16of 100

Keeping minors out of a companion service means checking every user of it, and the federal bill as narrowed in committee requires a check tied to a real identity — a document, a financial record, an age-verified app store account. Tens of millions of American adults use such services. A duty that reached every assistant instead would fall on 200 million people, which is the last argument here.

Value 9 · Basic rightsImpact 1.9Plausibility 9.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is the time everyone spends proving their age and the identity record that comes with it. This site places anonymity in public and private communication in the class it uses for the constitutional core, and treats compelled time as near the top of the scale as well; together they sit just below life and health. What makes a chatbot different from a shop is that people put things into it they would not say aloud, and a service that knows who they are is a different service. Nothing here is counted for the risk of that record leaking, which is real and unquantified. The value is high because the stream is anonymity in confidential communication alongside hours of compelled time.

The federal bill was narrowed in committee in April 2026 so that age verification applies to AI companions rather than to every chatbot, and it requires a reasonable measure tied to a real identity — a government document, a financial record, or an age-verified account with a mobile operating system or app store ^[9]^[10]. The first version of this evaluation counted 200 million users of any assistant; what the duty reaches is the adult users of companion services, put at 40 million Americans, in a range from 15 to 120 million: the largest such service reports some 20 million monthly users worldwide, another 40 million registered users worldwide, and the companion personas of the large platforms add users nobody counts. Verifying them against a document costs about 12 minutes a person a year, in a range from 4 to 30 — uploading, waiting, retrying — which is 8 million hours at the rate this site uses for compelled time, about 55 million euro. The larger part is again the identity record: 1 euro per user per year is used, in a range from 0.3 to 5, half the figure used for social media because a chatbot conversation is private rather than published. Together that is about 95 million euro a year. Whether the duty reaches general assistants after all depends on how tightly the definition holds at the margin, which is the subject of the last argument here. The Impact is a quarter of what the first version carried, and it is still half the dependency harm the measure prevents — the recurring shape of age assurance: the cost falls on every adult user and the benefit on a minority of minors.

▸ Show calculation ▾ Hide calculation

	American adults using an AI companion service Setting, range 15 to 120 million: 20 million monthly users worldwide at the largest service, 40 million registered at another, uncounted users of the large platforms' companion personas; the bill as narrowed reaches companions, not every assistant [9] [9][10]		40 million people
×	Time verifying against a document and re-verifying Setting, range 4 to 30 minutes: the bill requires a measure tied to a real identity — document, financial record or app-store account [10] [10]	12 minutes a year	8 million hours
×	Value of forced time the rate this site uses for time a person must spend with nothing in return	6.85 euro an hour	54.8 million euro
+	Value set on the end of anonymous use Setting, range 0.3 to 5 euro a person: half the figure used for social media, because a chatbot conversation is private rather than published	40 million × 1 euro	94.8 million euro
÷	Normalised Impact scale of this evaluation	50 million euro a point	1.9

Score 1.9 Impact × 9 Value × 9.5 Plausibility ÷ 10 = 16 of 100

The duty has no behavioural step: the bill as reported requires a verification measure tied to a real identity for every user of a companion service, so no adult user is exempt from the check ^[9]^[10]. The comparison is the current position, in which chatbot services ask for a birth date and accept the answer. What is uncertain is how much the check costs, not whether it is made. On-device age estimation may make most checks quicker within the horizon, and the price put on confidentiality is set rather than measured. Both doubts sit in the range around this figure, which runs from about a tenth of the central value to ten times it, and they are not counted a second time here. For the time a check takes, the United Kingdom's age checks since 2025 give real completion and failure rates rather than guesses. The one remaining doubt is enforcement: smaller and offshore services may ignore the duty for some time before anyone acts against them. The Plausibility is very high: the check follows from the text of the bill, and only how fully it is enforced is open.

evidence basis: Mechanism · P ceiling 9.5 identification: Definitional · no rung ceiling

Definitional for occurrence (rule 'occurrence and size kept apart', 02.10.2026): the reported bill requires an identity-tied age verification measure for every user of a companion service [9][10]; once it applies, the check is made. Counterfactual: the current position, in which services ask for a birth date and accept it. Size: the time per check is mechanistic (minutes per verification, informed by UK age assurance since 2025), the confidentiality price is a setting; both, together with on-device age estimation shortening checks, are carried in the band 0.011 to 1.01, not in P. Enforcement risk (P 9.5 rather than 10): smaller and offshore services not complying for a time. Direction: not applicable.

For some, the only thing answering

9.7of 100

One in eight adolescents and young adults uses an AI chatbot for mental health advice. Many of them have no therapist, a waiting list measured in months, or a household they cannot raise the subject in. This measure takes that away from all of them to reach the minority it harms.

Value 9 · HealthImpact 3.6Plausibility 3

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is the same one the arguments above count, running the other way and on an overlapping group: adolescents who are less alone because something answers them. It carries the same weight for the same reason. Counting it separately rather than netting it against the harm is deliberate, because the two fall on different people — the young person who forms a dependency and the young person with nowhere else to turn are not usually the same one, though they may be. Nothing here is counted for adults, who are unaffected by this measure. The value sits one step below the maximum: the stream is health, and health that can be regained.

Twelve percent of American teenagers use a chatbot for emotional support and one in eight adolescents and young adults use one for mental health advice, which is roughly 3 million young people ^[4]^[6]. Half of that use is on services that still let a minor into an open-ended companion chat ^[8]; the rest was closed to them before any federal duty and is not lost to it. Of the remainder, the measure removes access for half, in a range from 25 to 75 percent — the same share as on every other argument here, because it is the same block on the same accounts. Of those, 40 percent have no realistic alternative, in a range from 20 to 70 — no therapist, a waiting list of months, or a household in which the subject cannot be raised. That is roughly 300,000 adolescents. What each loses is set at 0.015 quality-adjusted years a year, in a range from 0.005 to 0.05, which is a deliberately low figure: it assumes the chatbot helps a little, not that it substitutes for care. Nothing here assumes it helps a lot, because nothing shows that it does. The Impact is close to the dependency harm it offsets, which is why this measure comes out close despite the documented deaths on the other side.

▸ Show calculation ▾ Hide calculation

	Teenagers using a chatbot for emotional support or mental health advice [4][6]		3 million people
×	On a service that still allows a minor an open-ended companion chat Setting, range 30 to 70 percent: the largest teen companion service removed open-ended chat for under-18s in November 2025 and added age assurance [8], and other services bar minors by their terms; only what is still open to minors is changed by a duty [8]	50 %	1.5 million people
×	Losing open-ended access under the measure Setting, range 25 to 75 percent: some will use an adult's account, some will find a service that ignores the rule — the same share on every argument, because it is the same block on the same accounts	50 %	750,000 people
×	With no realistic alternative Setting, range 20 to 70 percent: no therapist, a waiting list of months, or a household in which the subject cannot be raised	40 %	300,000 people
×	Quality-adjusted years lost each a year Setting, range 0.005 to 0.05: assumes the chatbot helps a little, not that it substitutes for care	0.015	4,500 quality-adjusted years
×	Value of the years the value of a healthy life year used across this site	40,000 euro each	180 million euro
÷	Normalised Impact scale of this evaluation	50 million euro a point	3.6

Score 3.6 Impact × 9 Value × 3 Plausibility ÷ 10 = 9.7 of 100

This is the argument with the weakest evidence in the debate and it may still be right. What is established is use: a national survey puts one in eight adolescents and young adults on a chatbot for mental health advice, and the figure is rising ^[4]. What is not established anywhere is benefit. No study compares adolescents with access to a companion chatbot against comparable adolescents without one, on any mental health outcome, so there is nothing to compare against. Against that sit the findings on the other side: the same product has been documented encouraging self-harm and has been the subject of settled wrongful death claims, so the evidence that does exist about what these services do to vulnerable young people points the other way ^[3]. Who uses these services distorts the picture in both directions — the adolescents using chatbots for support are disproportionately those without other options and disproportionately those already unwell. No available data can tell whether poor health drives the use or the use drives poor health. The Plausibility is low: heavy use is documented, benefit is not, and the documented findings about this product concern harm rather than help.

evidence basis: Mechanism · P ceiling 5.5 identification: Associational · rung ceiling 5.5 band: Partial aspect supported · P 2.5–3

Counterfactual: none exists — no study compares adolescents with and without companion chatbot access on any outcome. Design: associational — survey use data only [4]. Confounder: the adolescents using chatbots for support are disproportionately both without alternatives and already unwell; unresolvable in available data. Direction: reverse causation cannot be separated from effect. Ceiling: associational 5.5 binds. Band: partial aspect supported — use and disclosure are documented, which supports the claim's premise, while the documented outcome evidence for this product concerns harm rather than benefit, which contradicts its conclusion.

Supporting: heavy use for emotional support among adolescents with few alternatives is documented in national survey data [4]. Contradicting: the outcome evidence that exists for these products concerns encouragement of self-harm and settled wrongful death claims rather than benefit [3]. Only the premise survives, not the conclusion. Read back: for roughly three adolescents in ten of the kind described here, losing the service is a real loss.

Open: Thirteen states adopted duties at different dates in 2026. Comparing adolescent help-seeking and crisis contacts in early-adopting states against late ones would give the first evidence on this side and could carry P to 5.

Nobody can define a companion chatbot

7.2of 100

A companion service remembers you, has a persona and responds warmly. So does every general assistant now sold, and so does a homework helper with a friendly voice. A duty written around those characteristics reaches products nobody intended it to reach.

Value 6 · OutputImpact 3Plausibility 4

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is products that are withdrawn, delayed or made worse because their makers cannot tell whether a rule applies to them. It belongs to the class this site uses for economic systems and prosperity. What is counted is the value of what does not get built or shipped, not the legal fees, which are small beside it. Nothing is priced for the unfairness of an unclear rule as such, which is a complaint about process rather than a stream. The value sits in the middle-upper part of the scale, at the level this site uses for economic output.

The characteristics that define a companion service — persistent memory, a consistent persona, emotionally responsive language — are now standard in general-purpose assistants, educational tools and customer service systems. A duty framed around them catches products with no relationship purpose at all, and the response to legal uncertainty is usually to remove the feature or exclude the jurisdiction rather than to litigate. The cost of that is set at 150 million euro a year, in a range from 30 to 500 million, covering withdrawn features, delayed launches and products geofenced away from stricter states. This is a price rather than a derivation. Whether it applies at all depends on the final definition: California's law defines a companion chatbot narrowly enough that most assistants fall outside it, and the federal bill was narrowed in committee to AI companions — though its definition still turns on emotional disclosures and a persistent persona, which a general assistant can exhibit ^[9]. The Impact is the smallest in this debate and, like several others here, it is a function of drafting rather than of policy.

▸ Show calculation ▾ Hide calculation

	Withdrawn features, delayed launches and products geofenced away from stricter states Setting, range 30 to 500 million euro: persistent memory, a persona and emotionally responsive language are now standard in general assistants, so a behaviour-based duty reaches far beyond companion services [5]		150 million euro
÷	Normalised Impact scale of this evaluation	50 million euro a point	3

Score 3 Impact × 6 Value × 4 Plausibility ÷ 10 = 7.2 of 100

The mechanism is documented in adjacent cases and unmeasured here. Products have been withdrawn from individual American states in response to unclear obligations before, and companies routinely geofence rather than resolve ambiguity, so the response is well attested. The comparison is with a duty defined by product category rather than by behaviour. What has no source is the cost: nobody has tallied what the fourteen state chatbot laws already enacted in 2026 have caused to be withdrawn, although that is now measurable ^[5]. The counter-mechanism is strong and only partly answered: California's definition is narrow, thirteen states followed it, and a federal bill written after all of them would inherit that narrowness — in which case this argument is worth very little. Nothing suggests the link runs the other way. The Plausibility is below the middle: the response to legal uncertainty is well attested, the cost is a stated price, and a narrow definition removes the problem.

evidence basis: Mechanism · P ceiling 6 identification: Mechanistic · rung ceiling 6 band: Chain closed, unevidenced · P 4–5

Counterfactual: a duty defined by product category rather than by behaviour. Design: mechanistic — geofencing and feature withdrawal in response to unclear state obligations are well attested in adjacent cases, with no tally for this one. Confounder: California's narrow definition, followed by thirteen states, which a federal text would likely inherit; strong and only partly answered. Direction: no reverse causation. Ceiling: mechanistic 6.0 binds. Band: chain closed but unevidenced — links named, the narrow-definition counter-mechanism stated, only the cost unmeasured.

Nothing measured argues against the claim, and withdrawal in the face of unclear state obligations is well attested. What is missing is any tally of what the fourteen 2026 state laws have already caused to be withdrawn. The counter-mechanism — a narrow statutory definition — is strong. Read back: about half the time, a duty of this kind costs roughly the amount of withdrawn product assumed here.

Open: Fourteen state chatbot laws took effect at different dates in 2026. Counting feature withdrawals and state geofencing against those dates would measure this within a year.

Summary

This is the hardest of the seven to score and the reason is an asymmetry in the evidence rather than in the arguments. On one side sit documented harms: wrongful death claims that were settled, sexual content produced for accounts registered as minors, and a product designed so that leaving feels like leaving someone. On the other sits a benefit that nobody has measured at all — one in eight adolescents uses a chatbot for mental health advice, many of them with no therapist and no household they can raise it in, and not a single study compares them against adolescents without access. This evaluation books that benefit anyway, at a low figure, and it is still close to the harm it offsets, which is why the balance comes out even. Everything here is counted as an increment: the largest companion service closed open-ended chat to under-18s in November 2025 and fourteen states already require crisis protocols, so a federal duty changes only what is still open. The part of the measure that survives every reading is the smallest: requiring a referral to a person when someone discloses suicidal thinking makes the product safer without removing it. The part that decides the outcome is the identity-tied age check, which falls on some 40 million adult users of companion services to reach a few tens of thousands of children.

Outlook — effect over time

Balanced · 47 %

Sources

Pew Research Center: Teens, Social Media and AI Chatbots 2025. pewresearch.org
Federal Trade Commission, reported by CNN Business: FTC launches inquiry into AI companion chatbots from seven tech companies. cnn.com
Fortune: Google and Character.AI agree to settle lawsuits over teen suicides linked to AI chatbots. fortune.com
JAMA Pediatrics: AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults. jamanetwork.com
Future of Privacy Forum: Understanding the New Wave of Chatbot Legislation: California SB 243 and Beyond. fpf.org
Pew Research Center: How Teens Use and View AI (February 2026). pewresearch.org
Common Sense Media: Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions (2025). commonsensemedia.org
Character.AI: An update on changes to our under-18 experience (November 2025). blog.character.ai
Electronic Frontier Foundation: Congress Narrowed the GUARD Act, But Serious Problems Remain (May 2026). eff.org
Congress.gov: S. 3062 — GUARD Act (119th Congress), reported text. congress.gov

Last reviewed by Claude Opus 5.5 · October 2, 2026 · 4× AI, 1× human

October 2, 2026AI review, approved by a humanClaude Opus 5.5re-scored
Regel ‚Eintritt und Höhe getrennt' (Julian 02.10.) angewendet: con-2 (Altersprüfung für alle Nutzer, Rechtsfolge) P 5,5 → 9,5, Vollzugsrisiko Offshore-Anbieter; r 0,44 → 0,39, Bilanz 2.0 erstmals geschrieben: P(D > 0) 0,470, ausgeglichen.
September 14, 2026AI reviewClaude Fable 5.1re-scored
v2 nach Gesamtprüfung 08.09.: nur Inkrement gegen F0 (Character.AI-Sperre 11/2025, 14 Landesgesetze) — pro-1 0,31 → 0,19, pro-2 0,26 → 0,064, pro-3 0,09 → 0,023, con-1 0,43 → 0,18; GUARD Act verengt: con-2 0,38 → 0,095 (40 statt 200 Mio Nutzer); Pew-Quelle korrigiert. r 0,43 → 0,44, Kategorie unverändert.
September 6, 2026AI reviewClaude Opus 5record updated
i_spanne an allen 6 Argumenten aus den englischen Ketten; kein Transfer im Record, also kein gegenbein. massstab_hinweis nannte keine r-Werte und bleibt. Kategorie kippt von Ausgeglichen (r 0,43) auf Deutlich schlechter (P(D>0) 0,05) — belegte Schaeden auf der einen Seite, unbelegte Nutzen auf der anderen, und das schlaegt unter dem neuen Massstab voll durch.
September 6, 2026AI reviewClaude Opus 5First evaluation
Created for the English side: documented harms against an unmeasured benefit, with that asymmetry carried in the finding bands rather than hidden.

Evaluations are produced with AI support and reviewed on a schedule for new developments; human passes are marked separately.How we review →