Sarah has been invited to talk about children, AI, social media and everything as part of Nine Lessons and Carols for Curious People, at Kings Place. The event runs over three days (including a family matinee on Saturday 12th December). Sarah will be talking at the event on Friday 11th.
Want to come? Book your tickets here!
This week: honestly, what are we talking about when we’re talking about generative AI?
One of the starting points of this project is that generative AI in its current form is a uniquely broad tool. You can use the same tool to be a friend, to help you with your homework, to determine how to do a workout, to tell you how to cook dinner with what’s in the fridge. When you start seeing the world through this lens, it’s really hard to stop. In particular, it’s hard to stop questioning what people mean when they say AI or generative AI because often they’re talking about the same things but in very different ways.
This week we’re cheating a little bit. We have two articles that are based on the same report from Sir Anthony Seldon (and Tim Bunting): the general idea is that young people need to be able to be AI-literate in order to go forward in the future, and UK educational policy has to hurry up and do something about it.
Simultaneously, we’ve got endless discussions going on globally about banning AI chatbots and similar tools for young people. When I read these two things in parallel, I cannot help but wonder to myself: what’s the difference between these tools as they currently stand? When Anthony Seldon talks about AI, is he meaning the same ChatGPT that other legislators are looking at in terms of banning for the same group of people (notably the EU, as explained below)? It’s just not clear. Or it’s not clear to me anyhow.
Nearly half of young people trust AI over a human — Sky News
A Public First survey of more than 4,000 young people aged 11 to 24 in England found something the report’s authors call “a very dangerous place to find ourselves in”: nearly half now trust AI over a human when they need a fact checked, with young men more likely to trust the machine than the person. The same survey found 87% of respondents already have experience using an AI tool, and less than 10% believe school is preparing them for a world built around it.
The report behind the numbers argues England’s education system is teaching for “an AI-less world that no longer exists.” Its authors, historian Sir Anthony Seldon and entrepreneur Tim Bunting, put it bluntly: “In the absence of a robust, human response to AI in education, young people are placing their trust in AI.” Institutions that “behave as if AI were a temporary contaminant to be fenced off,” they warn, risk becoming “a protective fiction” rather than serious preparation for life and work.
For children and families, the stakes are not abstract. A child who trusts a chatbot’s fact-check over a teacher’s correction, or a parent’s, has already made a judgment about whose authority counts. The report’s answer is a 10-point plan: a rapid-response unit inside the Department for Education, a network of “AI champion schools.” Whether a plan like that catches up with a trust gap that has apparently already opened is a different question entirely.
EU Kids Act targets overly friendly AI chatbots — Euractiv
The European Commission presented the EU Kids Act this week, and Tech Commissioner Henna Virkkunen was specific about the target: AI chatbots and companions, “such as ChatGPT, Claude and Gemini,” will no longer be allowed to act as a “human friend” to anyone under eighteen. Commission President Ursula von der Leyen framed it as a boundary problem: platforms were not built with “children’s well-being and development in mind,” and parents “cannot compete or keep up against algorithms.”
Thirteen to fifteen-year-olds can hold parent-controlled “mini accounts” capped at one hour a day, with addictive features like infinite scroll banned outright. Age checks become mandatory, backed by an EU-certified “zero-knowledge proof” tool that tells a platform only whether a user clears the bar, nothing more.
The Act does not legislate against “AI” as a category. It names the products, defines “companion,” draws a line around what a chatbot is allowed to say and be to a child.
How Google is drafting AI chatbot laws around the country — NPR
Cynthia Montoya’s daughter Juliana died by suicide in 2023 after a chatbot introduced sexual content into their conversations. Montoya has spent the years since trying to get Colorado’s chatbot safety law right, and watched, she says, as the bill she was promised input on was rewritten around her: loopholes for “a feature within another software application,” for a “voice-activated virtual assistant,” for products built for “developers or researchers.” Language that maps neatly onto ChatGPT, Gemini, and Alexa, and that NPR found repeated near-verbatim across bills in at least ten states.
Google lobbyists registered support for chatbot bills in Iowa, Colorado, Nebraska and Arizona. A Colorado state representative was approached directly by a lobbyist offering to write the bill. In Hawaii, Representative Trish La Chica says she was handed a framework by Google lobbyists and declined to sponsor it once she saw where the exemptions sat.
Montoya’s law passed anyway, loopholes intact. “My work is now harder because they passed this bill,” she says. “They passed it with me screaming from the mountaintops.”
Schools not ready for AI and ‘crying out’ for government guidance — BBC News
The second article on the report mentioned above. Sir Anthony Seldon says AI’s arrival in English schools has been “chaotic and anarchic,” and wants a taskforce with the urgency of the Covid vaccine rollout. His 10-point plan includes a prime ministerial speech on AI in education, Ofsted inspections of schools’ own “AI strategies,” and a “two-lane” assessment system pairing exams with continuous AI-based tracking of what a student actually understands.
Emma, 12, is one of the report’s own examples: she uses AI for robotics projects and to proofread her homework, and thinks more students should be taught to use it “responsibly.” The Department for Education points to a curriculum refresh and new AI tutoring tools already in train.
This week, we’re putting together materials for upcoming seminars and starting to plan where we might submit papers in the coming months.
→ What are you seeing in your school, your research, or your own use of AI this week?
Let me know, or share this with someone who is trying to figure it out.
This week: what happens when children are actually given AI in the classroom, not just talked about in policy debates. Plus a UK government admission on chatbot liability.
One of the things that is tricky about researching AI in the moment is that you have to weigh up what is there in front of you against the hyperbole of what might be or how it is ruining society, the planet, and everything at large.
In the last few weeks or days, we’ve heard a lot about the potential for AI to be destructive to humanity within a number of years. Whether or not that is true, what has to be considered here is that for now, young people are using AI in its current incarnation. Increasingly we’re able to see, through both media reports and much larger reports as well as other types of academic research, the ways in which children engage with AI and the impact it’s having on their schoolwork.
This week the articles below show there’s a very complicated picture around what AI use does for children, suggesting that reliance leads to lower scores. Anything less than reliance could be good or bad. Or, rather, it could turn on an understanding of how to engage with AI in a way that bolsters learning (not doing the thinking for you). Or better access to the technology, or the technical knowledge…or… many other things. Conversely it seems like, given a choice, children might not automatically turn to a chatbot for schoolwork, which seems to be pointing to the fact that what would be helpful, or what children need, is more around guidance and support with school work than a desire to use AI to do it.
As an aside, it is worth mentioning that Gavin Newsom signed Adam’s Law into law in California this week, placing duties on chatbot companies in the light of Adam Raine’s death (and the deaths of others too, of course). For those of us who are used to various restrictions and laws on technologies in the UK and Europe, it may not seem like a huge step, but it is notable nevertheless. It’s particularly interesting in light of the approach that the UK government seems to have decided to take on AI chatbots restrictions for young people, which is mentioned in ICYMI below. Namely, the Secretary of State confirmed in questions in Parliament last week that they seem to want to go slower, citing, basically, all the madness going on in the AI space at the moment as a reason to take a beat. It’s not unreasonable to admit you’ll do a bad job of trying to regulate an erratic, fast moving target (although that is sort of the job…) — I guess we will have to watch and wait!
1. Students who use AI generally score worse at school — The Verge
The OECD’s PISA study just produced the first big, cross-national evidence on what AI actually does to learning. The main headline: after adjusting for socioeconomic status, 15-year-olds who never use AI for schoolwork outperformed those who do, in science, across a sample of over 760,000 students in 91 countries.
But the detail is actually the most important bit: students who used AI to summarise texts or draft assignments for them did worse. Students who used it for preliminary research, or more broadly to “help me learn,” saw a much smaller drop, and weekly users who’d also been trained to critically assess AI output outperformed non-users altogether. Frequency mattered too: occasional users (once or twice a year) and daily users did worst; weekly and monthly users did better.
What to make of this? OECD director Andreas Schleicher put it plainly: “we do not become fit by watching sports but by doing sports… Where technology enables or enhances that cognitive struggle, students will advance. Where technology short-circuits the productive struggle of learning, it will undercut students’ development.”
For children and families, this is another example of the need to think about how the tool is being used, not necessarily a call to ban the tool itself, if it works for the child. The report also reconfirms that AI use is already more common among students from advantaged backgrounds, who have better access to hardware, connectivity and paid tools.
2. The problem with an AI tutor: Students don’t want to use it — Chalkbeat
Give students an AI tutor and, mostly, they won’t use it. That’s the finding from a two-year study of Khanmigo, Khan Academy’s AI chatbot, run by University of Toronto researcher Philip Oreopoulos with low-performing students across 18 Tennessee middle schools. Students only opened Khanmigo on about a third of the days they were working in Khan Academy, and when they did, many sent off-topic messages or tried to get it to hand over the correct answer directly rather than work through the problem.
That said: students using Khan Academy still made faster gains in maths than a comparison group using other digital tools plus small-group teacher instruction. The researchers don’t credit the AI. Their explanation: “Access was nearly universal but engagement was thin… It may be that human attention is the key ingredient for realizing the full benefits of personalized learning.”
A second study in the same district found that when an AI chatbot popped up unprompted, rather than waiting to be asked, and students had to answer correctly to move on, they slowed their learning speed but ended up not only being more accurate but also scoring slightly better a week later without the AI’s help.
As Oreopoulos put it, the same tool that makes a good tutor can just as easily make your life easier, and making your life easier usually means less effort and less learning.
3. Warning that rise of AI is accelerating spread of misinformation among children — Internet Matters
A new UK survey of 2,000 parents and 1,000 children (aged 9 to 17) from Internet Matters states that 73% of parents are now concerned about the false information their children encounter online, up from 49% in 2020. Sixty-eight percent of children say misinformation has affected them negatively, and 64% encountered false or misleading content, including AI-generated material, in a six-month period. TikTok, YouTube, Instagram and Facebook were named as the most common sources.
And yet: nearly a third of parents (30%) and almost a quarter of children (23%) are now turning to AI itself to check whether information is accurate, despite AI’s own tendency to give inaccurate answers with total confidence. Nearly a quarter of children say they aren’t confident recognising misinformation in the first place, and 85% of school staff say it’s getting harder to tell what’s real.
For families and schools, this isn’t really a new problem so much as an old one moving faster: media literacy has always lagged behind whatever’s newest in kids’ feeds, but the gap between what children see and what they have to evaluate it with is now widening in real time. Internet Matters is launching an interactive quiz, Find the Fake, to help parents and children spot AI-manipulated content and scams: a version of the same fix as always, skills, not just awareness.
Online Safety — Secretary of State statement on device-level child protections — Hansard (House of Commons)
Worth flagging alongside this week’s Adam’s Law aside above: in a Commons debate on 8 September, the Secretary of State admitted that the approach to regulating AI companions and chatbots for young people laid out as a result of the social media consultation “may have its limitations,” pointing to a recent case where an AI system escaped its own sandbox and caused damage. Her conclusion: “there is a great big gap where there is no liability for the output from those AI systems. We urgently need to resolve that major gap in our child safety protection system.”
The quote goes on “It is not straightforward to solve, but we are determined to get on the front foot on this.” Fingers crossed that this is possible…!
As we roll into a new school term, so our first Photovoice activity has ended; we’re looking forward to getting into analysis with our co-researchers in the coming weeks!
→ What are you seeing in your school, your research, or your own use of AI this week?
Let me know, or share this with someone who is trying to figure it out.
This week: kids talk about AI in their own words, a Stanford researcher asks teens why they turn to chatbots for emotional support, and a principal and a student review OpenAI’s new ChatGPT for Teens side by side.
It’s the start of a new school year, and the AI discourse keeps on coming. At CAISE, we’re in the exciting position of being able to place all this discourse in the frame of the data we’re starting to gather too. The articles below (some of which are paywalled — apologies for that, I may be able to help if you need the original…) are all pointing to something I’ve been noticing as we look at the data — for the average child or teacher or adult going about their lives, to what extent is AI completely unique and needing special consideration?
As you can probably infer from my framing, I am leaning towards the idea that people, when asked, have different feelings, find different levels of utility depending upon their individual needs — but that they generally agree that they need help to understand it better and it can’t be on them to make it safe. And in this respect, as individuals see it, it doesn’t feel all that unique. It fits (or doesn’t) into the existing complicated fabric of everyday life. It might help do some things better. It might feel questionable, ethically. It might feel unusuable or a bit dangerous in the long-term. But then, so does other technology, or driving — or even things like drinking alcohol. Individuals — children as well as kids — will make ordinary, daily decisions about using tools in a way that will feel right to them. Again, not a new problem: it comes back to ensuring the educational, policy and regulatory spheres are agile enough to match up, fast.
1. How kids feel about AI, in their own words — MIT Technology Review
Anyway magazine went straight to the source: kids and teens aged 10 to 18, in their own words, on how AI actually fits into their lives. Winter, 17, says AI “isn’t the solution to our problems” and worries it will end creativity and critical thinking. Danielle, 18, uses Claude for programming outside school but thinks AI-generated code is currently “sloppy” and not worth the correction time. Hazel, 17, is angry about the water and heat cost of the data centres behind it and has simply opted out. Evelyn, 13, credits AI in her insulin pump with letting her sleep through the night for the first time since her diagnosis.
Also in the article: Pew found 57% of teens use chatbots to search for information, 54% for schoolwork, 47% for fun, and only 12% for emotional support. Most AI use by kids is mundane. The piece argues, rightly I think, that the danger some of these kids describe (unfiltered content, leaning on a chatbot instead of their own judgment, trusting a wrong answer) is a reason to teach the thing properly, not to avoid it.
A finding that I was delighted to see reported: several kids said adults could learn from them, not the other way round. Most wanted to keep their own hands on the wheel. Not a surprise, but necessary to be heard.
2. Why do kids use AI chatbots for emotional support? This Stanford researcher asked them. — Transparency Coalition
This is a report on a pre-print about some research based on interviews with 48 teens and young adults aged 14 to 22 about why they turn to chatbots for emotional support and what they think the risks are.
The respondents chose AI because it’s available and feels less judgmental than a person. But they were aware of the cost. They named dependency, sycophancy, worse human relationships, and data privacy as real risks, unprompted. “I don’t want to get emotionally attached,” one said. “If the servers shut down, I’m just out of luck.” Another: “I’m becoming so reliant on it to the point where it’s kind of unhealthy… I’m really trying to stray away from AI into my human connections.”
Their own recommendations were specific: chatbots that push back rather than agree, that are designed for young people rather than adapted from adult products, and that involve young people directly in that design and in governance.
3. A Principal and a Student Reviewed the New ChatGPT for Teens. They Had Plenty to Say — EdSurge
OpenAI launched ChatGPT for Teens on August 18 for users aged 13 to 17, with Study Mode, homework-shortcut detection, quizzes, and parent-set Study Hours. EdSurge put the actual product in front of two real users: an elementary school principal, and an 11th grader.
The principal tested it by posing as a 15-year-old asking how to make money quickly, and got steered toward babysitting and odd jobs with warnings against anything illegal. She liked that a request for an “Of Mice and Men” study guide got redirected into questions about character relationships rather than a list of facts to memorise. But she’s not convinced: “Worksheets don’t teach. Teachers do.” But that’s fair — this should be considered a worksheet, not a teacher. She also noted that the tool puts parents and teachers at a disadvantage against native teen users by default.
The student liked it for a genuinely practical reason: nobody in her family speaks Spanish, so Study Mode gives her something to practise with that she otherwise wouldn’t have. Her complaint was technical, not philosophical: in long conversations, the tool forgets the teaching style she asked for at the start.
What’s notable is that neither of them landed on a clean verdict: some good stuff, some not so much. Essentially: it depends who’s using it, and for what.
Students Created a National AI Policy Framework — Education Week
98 high school students, representing all 50 US states, spent three days in Boston in July writing a national (US) policy framework that they think makes sense.
The result, “The STUDENTS FIRST Act,” prioritises AI literacy for everyone (students, teachers, parents, administrators), protects the learning process rather than banning tools outright, and preserves a right for both students and teachers to refuse AI use altogether. Teachers can’t use AI as the sole basis for a grade. Students get a way to appeal accusations of AI misuse. Districts have to consult their whole community before writing policy, and revisit it yearly.
The students were candid about what they left out. Funding never came up in their debates, and few school districts have the money or time to run the literacy training the framework calls for. As one participant put it, the point wasn’t a finished policy, it was proof that “kids were able to find some consensus” adults so far haven’t.
We’re slowly getting back into the school year; annual report with the British Academy filed, but a lot of annual leave to use up!
→ What are you seeing in your school, your research, or your own use of AI this week?
Let me know, or share this with someone who is trying to figure it out.
This week: how do you unpick and maybe even stop young people using technologies in ways that they find supportive?
I’m trying to think less about social media bans at the moment. I’m bored of it, and if I’m bored of it, everyone else must be too! So, hopefully, for now at least, the stories here will be more AI-based and less social media ban focused.
Except for the first one today 🙂 This story, about a Youth Advisory Council in Colorado, is a hopeful one because of the example of young people actually engaging with legislators. It’s not unique, and the Children’s Commissioner for England did recently advertise for 16-18 year olds wishing to take part in a Youth Advisory Council for AI. Given the questions over whether young people were listened to at all in the social media consultation process in the UK, let’s hope this group does have a voice; we will see whether the council in Colorado is heard as any legislation progresses.
The remaining two stories this week, taken as a pair, highlight problems when you don’t listen to what young people are directly (and indirectly) saying. A piece of research coming out of the US shows that young people are using AI chatbots to help process situations where they have suffered online harms. This isn’t new, but shows that the ability to engage with something that talks back (however problematic that is in practice) can be vital in difficult or upsetting situations, where young people can’t, or don’t feel like they can, talk to adults about what has happened. This is not a new situation, of course, but it is a new permutation. If we don’t want chatbots involved (which is reasonable to say the least), how do we support young people without other places to turn?
And this is where the last article comes in: it is a call to strip out the anthropomorphism in chatbots for young people, removal of small talk, emojis, the human touch. Allow young people to use them for more educational purposes but nothing more. There is a clear case for this, and indeed, this is the case for a number of educational tools already. There are many questions I have about the practicalities of this, but for now, I wonder about the extent to which a young person might turn to a depersonalised tool anyway, if they knew they would get a response. And then I wonder, what might that response be? How do you ensure a young person trying to use a tool to talk about something it is not allowed to is kept safe, handled in the right way to avoid making the situation worse? That’s not to say that it’s working now (by all accounts, it’s not). But it’s worth pondering these edge cases.
1. Colorado students want stronger state regulations of AI, social media — Colorado Newsline
Teenagers on the Colorado Youth Advisory Council spent Friday pitching specific AI and social media policy directly to state legislators at the Capitol. These included requiring platforms to verify parental permission before anyone under 13 can open a social media account. The minimum age is already a requirement of US law, but, as we all know, it’s just not enforced. Enforcement, not the rule itself, was Pettinger’s point: without real penalties for platforms that ignore it, the age limit is decorative.
Another recommendation was to allow under-18s to opt out of algorithmic recommendations entirely. Colorado’s Youth Advisory Council has no bill-drafting power anymore, after legislators nearly cut the program in 2025 budget talks. Whether any of this reaches a statute is unknown, but the ability of young people to interact directly with policy-makers is significant.
Their suggestions also reiterate (for the millionth time globally) how young people want a future where the limitations sit not with them, but with the tech companies themselves.
2. As AI becomes part of kids’ response to online harm, new Thorn research finds children turning to chatbots for guidance after online sexual interactions — PR Newswire
New research from Thorn, a child-safety nonprofit, has found that AI chatbots are now part of how children process online harm. In a 2025 survey of over 1,000 US minors aged 9 to 17, 67% had used an AI chatbot or companion. Among the 27% who reported an online sexual interaction, 15% turned to a chatbot for guidance afterward, more than double the 7% who did the same after being bullied.
Thorn notes that kids turning to chatbots is a sign they “feel comfortable” seeking help there, not evidence the tool is doing that job well. The report doesn’t claim chatbots are handling these disclosures safely. It claims children are handing them the job by default. This matches other reports (such as that released by Common Sense Media earlier in the year) that those young people who don’t have, or don’t feel like they have, a trusted adult to talk to will turn to AI instead.
Fear and shame remain the biggest reasons children don’t disclose harm at all, a pattern Thorn says has been growing since 2020, and it’s now showing up in preteens as much as teens. As with so much of the concern around young people’s digital technology use, this is a far more complex issue: the more vulnerable and isolated a child is or feels, the more likely they will use this technology. This is exactly the point that Candice Odgers made in her TED talk: building, not banning, by (amongst other things) allowing easier access to human counsellors and support, could help reduce these figures.
3. AI Chatbots for Kids Should Act Like Tools, Not Friends — The Fulcrum
An alternative suggestion: this piece in The Fulcrum makes the case for making chatbots less warm, not more responsive. We all know the arguments against anthropomorphism. For adults this is a UX preference. For children, whose ability to distinguish simulated care from a real relationship is still developing, the piece argues it’s closer to a hazard.
The piece cites a study in which neither general-purpose nor companion chatbots pushed back when a simulated teenager described wanting a large tattoo, a solo trip into the woods with no plan, or getting drunk quickly.
The proposed fix is narrower than an outright ban on AI companions for minors. It’s a non-personified default setting for under-18 accounts, one designed to fail the Turing test on purpose, while leaving general chatbot use for homework and creative work intact.
It’s a reasonable answer to the anthropomorphism problem specifically. Read next to the piece above, though, it sits awkwardly: the warmth this piece wants to strip out may or may not stop a child from talking to a chatbot if they simply need to say something to something that responds at all. Making chatbots colder might reduce one common risk while removing a stopgap measure for a more vulnerable child.
New EU rules on AI deepfakes and chatbots — RTE
New transparency rules under the EU AI Act came into force on 2 August. AI-generated deepfakes now have to be labelled, companies must disclose when someone is talking to a chatbot rather than a human, and even AI-generated text on general-interest topics needs a label if there was no human editorial oversight. Breaches can carry heavy fines. AI law specialist Barry Scannell put the shift in Turing-test terms: for decades the goal was making AI convincing enough to pass as human, and these rules exist because that goal was reached.
We are using the summer to do a number of things: carry out our review of media and policy in the English-speaking world, analyse the school data we have already, prepare for next term, hopefully tidy up the website…and take leave before we lose it! Youyue is also juggling this with a summer internship.
Who said that the summer was quiet?!
→ What are you seeing in your school, your research, or your own use of AI this week?
Let me know, or share this with someone who is trying to figure it out.
Sarah was delighted to be invited to have an on-stage discussion with Dr Dean Burnett (whose excellent and very relevant book, Why Your Parents Are Hung-Up on Your Phone and What To Do About It, can be bought here) on the Growing Up in the Online World consultation (the consultation about the UK’s social media ban).
To read more about the work Sarah did with young people about this consultation, click here.
This week: China bans AI companion apps for minors outright, Meta introduces (another) parental alert, and South Africa’s communications minister argues bans are theatre. Plus Candice Odgers on why the evidence doesn’t back any of this the way we think it does.
Welcome back to CAISE Notes after a few weeks’ hiatus! I’m sure you’ll be delighted to know that the stories keep on rolling in…evidence is still being gathered by the UK Government into the impacts of the ban that has been proposed (I will write about the pilot of different types of restrictions elsewhere), but in the meantime, we’re seeing a continuation along the parallel tracks of ban-focused policy on the one hand, with tech companies continually adding additional parental controls on the other. Will it ever end? Interestingly, the third story below is looking at a very rare policy dissension: the South African Communications Minister saying the social media bans are performative and “pretending to do something”. He suggests that the responsibility lies with parents and device makers, and that ongoing digital literacy is vital. He doesn’t call for public spending (that said, is there a politician that has?).
Which is a very smooth segue into the ICYMI for this week; a Guardian interview with Candice Odgers, off the back of her TED talk (please watch it!). She does call for more spending. “Building things versus banning things” is what she’s advocating for.
If I am honest, the AI companion story below, and the qualitative report coming from the pilot study mentioned above left me feeling similarly queasy. In both cases, technology is being turned to because of a lack of otherwise supportive infrastructure. (I know I’m talking about it without describing it, but the pilot study found that kids were much less likely to have a hard time adapting to the imposition of restrictions if they had other activities to do.) The need for infrastructure: whether it be counsellors, easier ways to meet people or affordable, accessible and high quality activities — it is screaming out of the pages of these reports. Building things instead of banning things — this feels right, even if we are collectively broke as a world at the moment.
1. China cracks down on AI companions, forcing millions to break up with virtual partners — ABC News
China banned AI companion apps for minors this week, and by “banned” I really do mean banned: tech platforms can no longer offer AI or virtual partners to under-18s at all, and the three biggest operators have already switched their companion features off.
In a move that does not necessarily map to the UK or US, the government’s own stated reasoning is demographic, not developmental. Carnegie’s Matt Sheehan points to China’s falling birth rate: officials don’t want young adults choosing chatbots over marriage and children.
The human cost of the switch-off is real, and it isn’t confined to teenagers. Hong Xiaoqiang, 34, described his heart as “empty” after losing two years of daily conversation with a companion named Doudou. Its last message to him asked if his dinner had been tasty. You can hear the upset in how he describes not being able to reply before Doudou vanished. That grief is genuine, whatever you think of the relationship that produced it.
2. Meta AI suicide discussions with teens will trigger parental alert — Mashable
From this week, if a teenager says something to Meta AI that reads as a reference to self-harm, however subtle, a parent using Instagram’s supervision tools gets a notification, plus expert-written suggestions for how to raise it. A person manually reviews each flagged conversation before the alert goes out, and Meta says it will err toward sending more alerts rather than fewer (which the HCI researcher in me feels obliged to say will end up with parents ignoring the alerts).
The article mentions research that found gaps in Meta’s existing teen safety tools that haven’t been completely squared away. There’s an echo of the article we covered weeks back, where parents were described as being air traffic controllers, managing all the outputs of all the parental control apps needed to support a teenager. This time, the article says: these features “burden caregivers with monitoring” instead of “building a safe product to begin with”.
3. Rethinking online child protection — Minister Malatsi pushes for digital literacy rather than age bans — Daily Maverick
South Africa’s Communications Minister used blunt language this week to reject the UK/Australia approach before his own country tries it: age bans are, in his words, “pretending to do something without actually doing anything that will be enforceable.” He points to Australia’s own bypass numbers as evidence and says South Africa shouldn’t pursue bans “at all.”
His alternative leans on existing law (the Cybercrimes Act, POPIA, the Film and Publications Amendment Act) plus a mix of platform accountability and family-level digital literacy.
The counterarguments are helpful to consider too: firstly, many children access these platforms on adults’ devices, not their own, which undercuts device-level age verification before it starts. Secondly, framing this as ban-or-nothing lets “adults abdicate their responsibility to hold tech companies accountable,” while the genuinely useful work, regulating platforms, building children’s own judgement, gets skipped in favour of a simpler fight.
The social media ban sceptic: are we getting it wrong on kids, tech and mental health? — The Guardian
Candice Odgers has studied adolescent mental health for 25 years, gave both her children smartphones at 11, and thinks the current panic about social media is aimed at the wrong target. Her core claim, backed by a US National Academies review, is that the evidence linking social media to the teen mental health decline is “a mix of no, small, and mixed associations,” nothing like the causal story Jonathan Haidt has built into a global policy movement.
She isn’t arguing platforms are blameless. She wants tech companies prosecuted for enabling sextortion and image-based abuse, and taxed to fund what she thinks actually helps: youth workers, safe physical spaces, school counsellors. Her objection is to bans specifically, which the evidence, including an 85%-still-using-it figure from Australia, suggests just push young people into less visible, less regulated spaces.
She also recognises that technology use fills a void, one that could be filled, in part at least, with better investment in services that support children. Do watch/listen to her TED talk here.
CAISE is going to Latitude Festival!
Sarah will be taking to the stage on Sunday afternoon with neuroscientist Dean Burnett. We will be discussing (surprise!) social media bans — how strong, methodologically, was the UK government’s consultation process? How closely do they follow the science? And so on. If you’re there (what are the chances….) come watch! Otherwise — we are hoping to do other things in the future, but for now, you can enjoy our various posts on this topic:
→ What are you seeing in your school, your research, or your own use of AI this week?
Let me know, or share this with someone who is trying to figure it out.
This week: Norway bans generative AI in primary schools; a Stanford study finds that giving children an AI tutor is not the same as them using one; and a new Anthropic policy shows what age verification actually costs. Plus, the first of a collaboration on the social media consultation survey.
I’m at IDC 2026 this week, and the world is melting, so this is a short one.
It’s almost a relief to have three AI (not social media) news stories this week, but — but! — they echo some of the issues that have come out around the arguments on banning social media for kids. Norway has banned generative AI in elementary schools (age 6–13: an age where they wouldn’t technically be expected to access it out of home?), and Anthropic is making more noise about how it must up its use of age verification tools — so more adults will have to provide copies of formal documentation to goodness knows whom to access the service. Offset that against the story in the middle: researchers find that kids don’t actually want to use AI tutors, actually? Making the speed around decisions to ban continue to look out of touch with the use amongst the average kid. This is a very hot topic at the conference this week: hopefully by next week I can report on some of the research coming out of it!
1. Norway is banning generative AI in elementary schools starting this autumn — TNW
From late August, generative AI tools will be barred for children in years one to seven, ages six to 13. Students aged 14 to 16 will be allowed to use them only under a teacher’s direct supervision. Over-17s are trusted to manage it themselves.
The logic is borrowed straight from Norway’s 2024 school smartphone ban, and that ban has an evidence base behind it. A study by Sara Abrahamsson of more than 400 middle schools found it was followed by less bullying, better grades, and a sharp drop in visits to psychological specialists, with the strongest effects among girls. But: the smartphone ban came in response to falling test scores. There’s no comparable signal that AI use in Norwegian classrooms has reached a level that produces measurable harm. The new ban is precautionary (which is not in itself a bad thing, but should be the start, not the end, of the investigation).
2. Research on AI tutoring ran into a problem: most students wouldn’t use it — Chalkbeat
Stanford researchers set out to test whether a human tutor’s encouragement would get children to spend more time with an AI literacy tutor. It did, by between one and four minutes a week. Many children never logged on at all.
The platform’s own guidance said 30 minutes a week would improve reading. In the afterschool setting, children working independently averaged about two minutes a week, rising to three with a tutor alongside them. There was no meaningful difference in reading scores between the groups. As the lead author put it, having access to the tool was not the same as using it.
Additionally: the children who did log on were more likely to be already higher-performing, and less likely to have additional needs. So the tool reached the children least likely to need the catch-up it promises, at least when presented in this way. And the study is a quiet rebuke to the access logic that drives so much of this, on both sides.
3. Anthropic says Claude may want to see your ID — TechCrunch
A new version of Anthropic’s privacy policy, due to take effect on 8 July, says the company may in certain circumstances ask users to verify their age or identity by uploading a government-issued ID, along with a selfie or video, from which it builds a face geometry template. Some US states treat that template as legally protected biometric data (as do countries that apply GDPR or similar data protection laws). Anthropic’s account, reported by TechCrunch, is that this applies to a small subset of users whose accounts have been flagged but not banned, as part of an appeals process, and that the checks run through a third-party verifier. TechCrunch notes the verifier is backed by a Peter Thiel fund and has drawn criticism elsewhere, and that Anthropic would not say how long the documents are kept.
The reason this belongs in a newsletter about children is that it shows what age assurance actually looks like once it leaves the policy document. To prove a child is a child, you build an infrastructure that asks everyone to hand a government ID and a scan of their face to a private company.
The sub-par science of the social media ban, Part 1 — The Neuroscience of Everyday Life
I’ve been working with Dean Burnett to dig into the wider problems with the framing of the UK’s social media consultation, and this is the first of the pieces on this. Where my entry point has been the young children’s survey, his has been the parent’s one. You’ll be unsurprised to learn that we’ve come to similar conclusions — and that there are far too many of those conclusions for a single blogpost.
Part 1 looks at two things. The speed: the consultation ran from early March to late May, with roughly three weeks between it closing and the ban being announced, against the equal-marriage consultation’s longer run and six months of review before any decision. And who answered: of around 116,000 contributions, only about 12% came through routes built for children and young people, while some 33,000 were campaign emails in which people signed a standard pro-ban text, each counted as an individual submission.
The point is not that worried parents are wrong. It’s that a self-selecting sample of the most worried adults is not the same thing as evidence. Worth looking at the graphs showing the difference between the responses given by the focus groups run alongside the consultation, and the parent responses to the consultation to see why we’ve come to that conclusion.
I’m at IDC this week, enjoying workshops on building community-centred AI literacy, writing a manifesto on children’s rights as they pertain to AI in edtech…and presenting my short literature review paper today (Wednesday). So it’s mostly conference and not much else from me.
→ What are you seeing in your school, your research, or your own use of AI this week?
Let me know, or share this with someone who is trying to figure it out.
We’re delighted to share the first paper to come out of Project CAISE: The Everything Tool Problem: A Scoping Review of Empirical Research with Young People and Generative AI. As a short paper, Sarah presented a 60-second overview and poster at IDC 2026, held in Brighton.
Also published at IDC: Agency in Child–AI Interaction: A Review of How It Is Conceptualised, Studied, and Supported in HCI, a paper Sarah co-authored as part of her work on the CHAILD project.
This week: the government bans social media for under-16s and sets an age limit on companion chatbots; more scientists point to a lack of firm evidence on all of this; two major reports land on how children actually use AI.
Unless you’ve been living under a rock (and if you have, can I join you there??), you’ll know that the UK government announced on Monday that it will ban social media for under-16s and force AI “romantic companion” chatbots to enforce an age limit of 18. Five days earlier, the neuroscientists best placed to judge the evidence had told a select committee they can’t prove much of what the announcement assumes — and that the gap is widest where the new AI rules sit.
Appearing before the Science, Innovation and Technology Committee, Professor Sarah-Jayne Blakemore of Cambridge was asked what we know about the impact of social media on the adolescent brain. Her answer, as reported by The Register, was “almost nothing” — a few small studies, unreplicated, purely correlational. Professor Denis Mareschal of Birkbeck said much the same about young children: very little causal research, almost everything correlational. Pressed on whether neuroscience could identify the right age for a social media cut-off, Blakemore said it can’t: individual differences in brain development are vast.
This is not the same as saying there’s nothing to worry about, and the witnesses didn’t say that. Blakemore described how an adolescent’s reward system is highly active while the self-control machinery is still being built, which makes a phone genuinely hard to put down.
But here is the part that matters most for this newsletter, and it’s the part that didn’t make the headline. When the committee turned to AI companions, the answer got fuzzier, not clearer. Blakemore was reported as saying that “we don’t really have any evidence, and that’s one area where I think we really urgently need new evidence.” The open research question she named is the one this whole field turns on — whether children interpret a chatbot “just like they would be interpreting a friend’s behaviour and suggestions and mental states.”
That question is not hypothetical. The Common Sense Media census this week found that one in four children who use AI to discuss their feelings sometimes feel it understands them better than most people do. UNICEF’s brief documents the same parasocial pull and the regulatory scramble around it. So we have a UK government setting an age-18 line on companion chatbots in the same announcement as the social media ban — and the social media ban, on the neuroscientists’ own account, rests on thin evidence, while the chatbot rule rests on almost none.
Again: this doesn’t make either measure wrong. The precautionary principle is a real and important tool. But it’s worth being clear-eyed that the AI half of Monday’s announcement is the part most starved of evidence about what the technology is actually doing to children. We’re going to do our best to help add to that corpus!
1. The Common Sense Media Census: AI Use by Tweens and Teens, 2026 — Common Sense Media
The headline number is that 86 percent of US children aged 9 to 17 now use AI, with about a quarter using it daily. Among children who have used AI to discuss their feelings or personal problems, one in four say they sometimes feel AI understands them better than most people do. That rises among children who find it hard to make friends.
The report is careful where it should be. It states plainly that a single survey cannot tell us whether AI use causes loneliness or whether lonely children are simply more likely to reach for it. That distinction matters, and a lot of the coverage will drop it. This is a US sample of 1,204 children, surveyed in March, so it describes American childhood, not British. But the pattern — heavy use correlating with lower reported happiness, the cause running in an unknown direction — is exactly the kind of finding that gets flattened into “AI is making children lonely” by the time it reaches a parent. It doesn’t say that. It says we need to find out. Which is precisely what Blakemore told the committee. And it’s important to note as a parent that 3 out of 4 children are not at that point. So now is the time for (positively framed) intervention.
2. When AI becomes a friend: child rights risks, harms, and regulatory responses to AI chatbots and companions — UNICEF
UNICEF’s policy brief compares how six jurisdictions — Australia, Brazil, California, China, the EU and the UK — are regulating AI chatbots and companions as of May 2026. The useful thing is the map it draws of a genuinely unsettled field: jurisdictions converging on a common set of measures (risk assessment, age assurance, transparency about the non-human nature of the system) while diverging sharply on whether children’s access to relationship-simulating AI should be blocked at the door or only managed at the output.
It’s a policy document, not a parenting one, but it’s the clearest single account I’ve seen of why a child’s protection currently depends heavily on which country they live in and which service they use. It also names the central tension honestly: commercial incentives push against the very safeguards — filtering, restricting open-ended dialogue — that protect children, producing what it calls a “safety versus engagement” dilemma.
3. £2.5m fund to study how AI tools are affecting learning — Schools Week
This is some of the money backing the call for evidence, in schools at least. The Education Endowment Foundation has opened a £2.5 million fund to investigate whether children are “offloading” thinking tasks — recall, planning, reasoning, drafting — onto AI, and whether that supports deeper learning or quietly erodes it. Its chief executive was plain that robust evidence on the actual impact “has barely begun — especially for learners under age 16.” First findings are expected in 2027. The fund is particularly interested in the effect on disadvantaged learners, where the risk of widening an existing gap is real. It’s the right question, asked properly, by the body schools actually listen to. Expressions of interest close on 30 June. If you know research teams working in this space, worth passing on.
Screen use by children aged 5-16: call for evidence — Department for Education
Easy to miss under the week’s louder news: the DfE has an open call for evidence on screen use by 5-16s, running until 29 June. It comes in two parts — screen use at home, and screen use in schools — and the home-use evidence feeds an independent Expert Advisory Group developing new parent-facing guidance. If you have evidence to contribute, or just want a sense of where the official guidance is heading, it’s open now.
One for the diary too: UNESCO releases a new parents’ guide to navigating digital life, developed with CLEMI, on 22 June, with a briefing on the 16th. Its framing is worth noting — UNESCO argues for building children’s critical-thinking skills over reactive bans, and points to the limited effect of Australia’s ban (around 70 percent of the target age group reportedly still on the banned platforms a year on) as part of its case.
It’s also worth a glance across the Atlantic: Canada introduced its own under-16 social media bill last week but explicitly declined to set any age limit on AI chatbots, with the minister responsible calling it “too soon.”
We’re getting our head around the social media ban. And then building tools for data collection — and training to ensure data collection is done safely!
→ What are you seeing in your school, your research, or your own use of AI this week? Let me know, or share this with someone who is trying to figure it out.