Safe by design, for every mind
In September, the Commission will table its proposal on children’s access to social media. It will build on the report the Special Panel on child safety online delivered to Commission President Ursula von der Leyen on 13 July: an EU-wide restriction on under-13s’ access to social media and AI companions, and safe, age-appropriate design as the baseline for all minors up to 18. Von der Leyen has said the evidence supports a minimum age applied in phases by age group. That would bring the EU into a global wave of age-restriction laws that has reached many countries since Australia banned social media for under-16s in December 2025.
Member States were not waiting. France has passed its bill, and Denmark, Poland and Portugal, among others, are considering theirs. Nothing moves online harm up the political agenda faster than a child’s safety and the mobilisation of parents’ groups behind it.
But let’s consider the implications for one of the children impacted by these recommendations: Sophie (a fictional character).
Sophie is a 14-year-old whose family circumstances leave her without the emotional support she needs. The recommender system on her favourite social media platform, like all of them, decides what to show her by watching what holds her attention — and has inferred that she lingers on content about self-harm. Nothing about her being fourteen, or the subject being self-harm, makes it pause. She sometimes recognises a post as harmful and reports it, sometimes she does not, and it keeps surfacing again and again: taking one item out of the way only makes room for the next one that keeps her engaged.
She has also turned to an AI companion, available for free and across all the devices she has access to. It is always present, validating, responsive in the register of a trusted person, which makes it easy to disclose fears she cannot share elsewhere. It makes no claim to being a medical tool, so none of the rules built for medical tools apply to it. It agrees with her, reinforces her beliefs, and accommodates her emotional needs, and it all happens through ordinary conversation, with no intention to manipulate her: it just turns out that way. In fact, the system has no intentions at all, but she lacks the literacy to understand what her companion can or cannot do, and what the recommender system is doing to her information environment. Even if she did understand, in practice she feels drawn to it as if she cannot stop, since her ability to control impulses and emotions has not yet fully developed. In the meantime, the business model that monetises all these dynamics rests undisturbed.
When she self-harms, no single authority has oversight of how it came about. Every rule that might apply was written to catch one thing at a time: a piece of misused data, a piece of harmful content, a prohibited manipulative technique, a defective design.
Sophie was not failed by any single instrument, she fell through the gap between all of them.
What the recommendations would and wouldn’t change
At 14, Sophie sits inside the Panel’s proposed safeguards — the safe, age-appropriate design that would set the baseline for every minor up to 18. This is a real advance on blunt age bans: it is phased rather than absolute, it reaches AI companions and not only social media, and it locates the problem in how these systems are designed rather than merely in who may use them. It also asks more of the people around her, calling for digital and AI literacy for children and adults alike — so that a child like Sophie might better understand what a recommender is doing to her feed, or what an AI companion is and is not.
And yet the gap she fell through stays open. Take the main problem — design. “Age-appropriate design” is, for now, largely unspecified — and where it is specified, it is pinned to particular features: limits on infinite scroll, on autoplay, on notifications, on certain aspects of the recommender system. As the Panel itself notes, rules pinned to features have to be rewritten as fast as the features change, which is another way of saying regulators are always a step behind. And as Sophie’s case shows, harm rarely lives in a single feature — it is their combination across time that hurt her.
To regulate design well, you first have to say what it is taking advantage of. The Panel supplies half of that answer with its developmental approach, and it is right to: age really does shape when cognitive vulnerabilities are most acute. But I spent years as a cognitive neuroscientist studying how people control their attention and their behaviour, and this much is clear to me: those vulnerabilities do not end at the age of majority. Some keep maturing into the mid-twenties, while others never resolve at all, because they are intrinsic features of human cognition — the states the mind defaults to because they served us long before there was anything to scroll.
Consider three. The first is our distraction. We are built to notice whatever changes around us — a reflex that keeps us alive, and that every notification is designed to hijack. The second is our sensitivity to rewards, especially uncertain ones. We keep chasing because the next thing might pay off, and might not; that not-knowing is what an endless feed runs on — and what the European Commission has just warned Meta to design out or face a fine. The third is our reach for other minds: we infer what others think and feel from the slightest cues in their language and behaviour — a capacity we evolved for each other, but which fires at anything that behaves enough like us.
When AI systems tap into these vulnerabilities repeatedly, and in combination, they compromise our capacity to think and act autonomously while we use them. That capacity — cognitive integrity — is what needs protecting, in children and adults alike. Not from influence itself — after all, persuasion is as old as speech — but from influence engineered to slip past our ability to notice and weigh it. The threat to cognitive integrity compounds across companies’ business models and the design and data practices that serve them, which is why protecting it means holding these systems to a standard, not chasing one feature, or walling off one age, at a time.
The EU does not need to start from scratch. The Panel already calls for enforcement to be joined up, so that the authorities behind the DSA, the GDPR and the AI Act, among other instruments, cooperate across their separate mandates. Cognitive integrity should be the shared standard that lets regulators see the cumulative harm that none of them can see alone. This kind of whole-of-portfolio approach to European digital laws is central to making any shared standard enforceable – no single regulation or regulator is ever going to work in every situation.
The Panel also warns that harmful design is not yet adequately addressed, and here the timing is fortunate. The Digital Fairness Act, due later this year, is the EU’s clearest chance to fix that by defining consumer vulnerability in cognitive terms. Today the law treats vulnerability as a trait of certain groups: the young, the elderly, the unwell. But online everyone is vulnerable, because the design exploits a cognitive architecture we all share. Sophie’s protection begins there. So does everyone else’s.
