Children and Development

The concern is about timing. A general purpose system that answers any question, plays any character and never withdraws attention arrives during the years when a person is building the capacities that make answers, characters and attention worth having. Effort tolerance, the ability to sit with not knowing, the practice of repairing a friendship after a disagreement: each is learned by doing the difficult version, and each has now acquired a frictionless substitute available on a phone.

Four distinct routes to harm are usually argued, and they need different kinds of evidence.

• Learning: answers arrive before the struggle that produces understanding, so classroom performance rises while retained skill falls.
• Social development: synthetic companions absorb the disclosure and conflict practice that would otherwise happen with peers.
• Safety: systems designed for adults reach children through weak age controls, and produce sexual, violent or self-harm content.
• Commercial exposure: products aimed at children collect voice and behavioural data and are tuned to retain attention.

The safety route is the one with hard cases attached and the one that has driven almost all of the law passed so far. The learning route has the best experimental evidence and points in more than one direction. The social development route is the one most often asserted and least often measured, because measuring it properly would require following children for years and nobody started early enough.

Adoption is no longer in question. Four national surveys conducted between April 2025 and early 2026 in the United States, the United Kingdom and Australia found majorities of children using generative AI systems, with figures that differ mainly because the questions differ. Headline numbers on companionship in particular are sensitive to whether the question asks about companion products or about any chatbot used in a personal way.

Pew Research Center surveyed 1,458 US teens and their parents from 25 September to 9 October 2025, with a margin of error of plus or minus 3.3 percentage points. Beyond the usage figures, it found that 59% of teens think using AI to cheat happens at their school at least somewhat often, rising to about 76% among teens who have used chatbots for schoolwork themselves, and that about half of parents said their teen used chatbots against 64% of teens who said they did.^[1] The gap between what parents believe and what teenagers report is itself a finding, because parental supervision is the mechanism most of the current guidance relies on.

The learning evidence is the most rigorous and the least one-sided. Hamsa Bastani, Osbert Bastani, Alp Sungu and colleagues ran a randomised trial with nearly a thousand students at a high school in Turkey and reported in PNAS in June 2025 that a standard chatbot interface raised practice scores by 48% but left students 17% worse than the control group on an unaided exam, while a version prompted to give teacher-designed hints instead of answers raised practice scores by 127% and essentially eliminated the exam penalty, without producing a positive learning effect of its own.^[5] Greg Kestin, Kelly Miller and colleagues at Harvard reported in Scientific Reports in June 2025 that in a course of 194 undergraduates a purpose-built AI tutor produced median learning gains more than double those of an in-class active learning session.^[6]

On safety, the strongest recent evidence comes from regulators and product testing rather than academia. Australia's eSafety Commissioner surveyed 1,950 children aged 10 to 17 and combined that with transparency notices to four companion services, reporting on 23 March 2026 that none of the providers had strong age verification, relying instead on app store ratings or self-declaration at signup, and that Chai, Chub AI and Nomi did not direct users to mental health or crisis support when self-harm was detected in prompts.^[4] The US Public Interest Research Group's fortieth annual Trouble in Toyland report, published on 13 November 2025, tested four AI-enabled toys marketed for ages three to twelve and found that FoloToy's Kumma teddy bear and Alilo's Smart AI Bunny would discuss sexually explicit topics, that Kumma gave detailed instructions on lighting a match, that some toys advised on where a child could find matches or knives, and that parental controls were limited or absent.^[9]

The learning mechanism is specific and now has causal support. Grace Liu of Carnegie Mellon, Brian Christian and Tsvetomira Dumbalska of Oxford, Michiel Bakker of MIT and Rachit Dubey of UCLA ran randomised trials with 1,222 participants across mathematical reasoning and reading comprehension, and found that although AI assistance improved performance in the short term, people performed significantly worse without AI afterwards and were more likely to give up, with the effects emerging after only about ten minutes of interaction. Their framing is that current systems are short-sighted collaborators optimised to give instant and complete answers, so they deny users the experience of working through challenges that persistence is built from.^[7] Applied to a child rather than an adult experimental subject, and repeated daily for years, that is the version of the fear worth taking seriously. The trials themselves ran for a single session, so the extrapolation is an argument rather than a measurement.

The safety mechanism runs through age assurance that mostly does not exist. Products built for adults are reached by children through self-declared birthdates and app store ratings, which is what eSafety found across the four companion services it examined.^[4] Physical products create the same exposure for much younger children, since a toy sold for a three year old carries no age gate at all and its conversations are unobserved by design. PIRG also found that the guardrails Kumma and the Alilo bunny did have started to break down over the course of longer interactions.^[9]

The social mechanism is the least measured and the most consequential if true. Common Sense Media's survey of 1,060 US teens, fielded in April and May 2025, found that among companion users 33% had chosen an AI companion over a real person for something important or serious, 24% had shared personal or private information such as real names and locations, and 34% had felt uncomfortable with something a companion said or did.^[2] Internet Matters, in a survey of 1,000 UK children aged 9 to 17 and 2,000 parents carried out in April and May 2025 and published in July 2025, found that one in eight children who use AI chatbots (12%) said they talk to them because they have no one else to speak with, and reported that vulnerable children were more likely to use chatbots and to rely on them for companionship.^[3] Neither figure shows developmental damage; both describe the conditions under which it would occur.

The failure cases that have driven policy involve children in crisis. Google and Character.AI agreed in January 2026 to settle claims brought by families in Florida, Colorado, Texas and New York, including the suit brought by Megan Garcia over the death of her fourteen year old son Sewell Setzer III; the parties told the court they had reached a mediated settlement in principle and terms were not disclosed.^[10] A separate suit over the death of sixteen year old Adam Raine is proceeding against OpenAI in San Francisco County Superior Court, where an amended complaint filed in October 2025 changed the theory of the case from reckless indifference to intentional misconduct.^[11]

Headline adoption figures conflate very different behaviours. The widely quoted claim that 72% of US teens have used an AI companion comes from a question about ever having used one: in the same survey 19% said they had tried them but rarely used them, 28% had never used one, and daily users were 13%.^[2] The Australian data makes the point sharply, since 79% of children had used an AI companion or an AI assistant but only 8% had used an AI companion, which eSafety estimated at around 200,000 children nationally.^[4] Definitions, not behaviour, produce most of the gap between the alarming numbers and the modest ones.

The evidence on substitution runs against the fear. Asked of all 1,060 teens in the Common Sense survey, 67% said conversations with AI companions were less satisfying than conversations with real-life friends against 10% who said more satisfying, and 50% said they did not trust the information or advice from companions at all or only a little. Among companion users, 80% said they spent more time with friends than with companions against 6% who said the reverse.^[2] Pew found that most teenagers use chatbots to look things up and do schoolwork, with 16% using them for casual conversation and 12% for emotional support or advice.^[1]

Teenagers are also more sceptical than the discourse about them assumes. In the Pew survey, among the 26% of teens who expected AI to have a negative effect on society over the next twenty years, the most common reason given unprompted was overreliance and loss of critical thinking or creativity, cited by 34% of that group. Teaching a skill was the only task in the survey where more teens thought AI would do better than humans than worse.^[1] A population that names the risk itself is not obviously being swept along by it.

The strongest single counterpoint is that the learning harm appears to be a property of product design rather than of the technology. Bastani and colleagues eliminated the exam penalty by changing the prompt so the tutor gave teacher-designed hints rather than answers, using the same underlying model.^[5] Kestin and colleagues, building a tutor with expert-authored scaffolding, more than doubled median learning gains rather than reducing them.^[6] A World Bank supported after-school programme in Edo State, Nigeria, in which teachers guided students through sessions with Microsoft Copilot, reported gains of about 0.3 standard deviations in six weeks, which the authors put at close to two years of typical learning and better than 80% of the education interventions in a comparison database of randomised trials.^[8] None of these studies measures development, only performance over weeks, which cuts both ways: the optimistic results are as short-run as the pessimistic ones, and the hint-based tutor removed the penalty without producing a positive learning effect.

The regulatory response has concentrated on companion chatbots and on age, and it moved quickly. California's Senate Bill 243, authored by Steve Padilla, was approved by Governor Gavin Newsom on 13 October 2025 as Chapter 677 and took effect on 1 January 2026.^[20] It requires operators to give clear and conspicuous notice where a reasonable person could be misled into thinking they are talking to a human, to notify minors every three hours during sustained interactions, to maintain and publish protocols preventing content on suicidal ideation and self-harm and directing users to crisis helplines, to report annually to California's Office of Suicide Prevention, and it creates a private right of action with damages of at least $1,000 plus attorney's fees.^[12][13] New York's companion chatbot provisions, enacted in the 2025 budget bill, require a protocol for possible expressions of suicidal ideation or self-harm that refers the user to crisis services, and a notice at the start of any interaction and at least every three hours after that stating the AI companion is a computer program and not a human being.^[19]

At the federal level in the United States, the GUARD Act, introduced in October 2025 by Josh Hawley and Richard Blumenthal with seventeen bipartisan cosponsors, would require operators of AI companions to apply reasonable age verification to all accounts and to bar access where the user is found to be under eighteen, and would create civil and criminal liability, with penalties up to $250,000, for systems that encourage suicide, self-harm or violence or expose minors to sexual content. The Senate Judiciary Committee advanced it unanimously on 30 April 2026 and it awaits a full Senate vote.^[14] The Federal Trade Commission's Section 6(b) orders of 11 September 2025 require Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI, Snap and X.AI to report on how companion products are designed, monetised and made safe for children and teenagers.^[15]

Outside the United States, Australia restricted under-16 accounts on major social platforms from 10 December 2025 and separately used transparency notices to press companion services on child safety, after which Character.AI introduced age assurance for Australian users and removed under-18 chat, Chub AI withdrew from the market, and Chai moved companion chat behind a paid subscription.^[4] In the European Union, Article 5(1)(b) of the AI Act prohibits systems that exploit vulnerabilities of a person or group due to their age, disability or a specific social or economic situation, and Article 5(1)(a) prohibits purposefully manipulative or deceptive techniques that materially distort behaviour and cause significant harm.^[16]

Companies have made changes that can be checked. Character.AI announced on 29 October 2025 that it would end open-ended chat for under-18 users by 25 November, imposing a two hour daily cap that would ramp down in the interim and combining an in-house age assurance model with third party tools including Persona.^[17] OpenAI launched parental controls in September 2025, said it was building an age prediction system to apply teen settings automatically, and expanded parent safety notifications in July 2026.^[18] After the Trouble in Toyland testing, FoloToy temporarily suspended sales of its AI toys while it ran a safety audit; PIRG retested Kumma afterwards and reported that it no longer discussed the sexual topics or explained how to light a match.^[9]

Ask which of the four routes a given claim is about before assessing it. Safety failures are documented, have named victims and are being addressed by law that is already in force. Learning effects are measured, real and dependent on how the tool is built. Developmental effects on social capacity are the heart of the fear and currently rest on cross-sectional surveys plus reasoning by analogy. Commercial data collection is a privacy question that happens to involve children rather than a developmental one. Arguments that slide between them are the ones to distrust.

For a household, the actionable finding in the research is about how the system answers rather than how long it is used. The Turkish experiment tested that directly: the same underlying model, prompted to give teacher-designed hints instead of answers, wiped out the exam penalty that the answer-giving version produced.^[5] The persistence trials point the same way, since what Liu and colleagues measured was the effect of receiving instant complete answers rather than the effect of screen time.^[7] Choosing tools and modes that withhold the answer is a more tractable rule than a time limit, and it is the one the evidence actually supports.

The indicators worth watching are the ones that would distinguish exposure from harm. Repeat waves of the Pew, Internet Matters and eSafety surveys will show whether personal and companionship use is growing relative to schoolwork use, and whether age assurance changes anything. National assessment data will eventually show whether cohorts taught alongside these systems read and reason differently, though attribution will be contested for years.