Don’t You (Forget About Me): Three Features That Owe a Child an Answer
For some background music when reading this post, we encourage you to (re)discover Simple Minds performing “Don’t You (Forget About Me)”.
Where the earlier posts in our series “So, what is a plan” told a platform what to switch off, what to forbid, and what to start the child behind, this set tells a platform what it has to build and deliver, a duty to serve rather than a boundary to respect. It leads on the fourth of the repository’s principles, that a plan for children is co-designed with them, and it takes the three fixes in the order that keeps a child’s voice in front: the report first, then the feed, then the companion.
The unanswered report teaches a child to stop asking
A child who runs into something wrong online, for example, an inappropriate or worrying message, an image that should not be there, an adult who should not be asking, has one move that puts the child back in charge, which is to report it. When the report goes nowhere, when there is no confirmation that anyone received it, no answer, and no visible result, the child draws the conclusion that reporting does nothing, and the next time is likely to stay quiet. The one feature built to hand a child their voice back can, when it fails, teach the child to stop using it.
Rules already exist, and most are binding. The Digital Services Act (DSA) gives every user a route to flag illegal content and a route to contest a platform’s decision, through its provisions on notice and action, internal complaint handling, and out-of-court dispute settlement. The European Commission’s DSA guidelines add the details that matter for a child: reporting mechanisms a child can find and understand, confidentiality by default, priority for a minor’s report, confirmation that the report was received, and a timeframe given to the child. The rule reads well. What a child meets when something goes wrong is another matter, because a reporting route that exists in the text and answers nobody in practice is the problem the card names. The loophole sits in one gap the guidelines leave open. They ask that a timeframe be communicated to the reporting child, but they do not ask a platform to publish how long its reports actually take across the board, so the system stays accountable to each child in private and to nobody in public. The fix on the card closes that gap: Reporting children can find and trust, with response times published. The published response time is the repository’s own addition, and it is what turns a private promise to one child into a public measure the whole system can be held to. Save the Children and the 5Rights Safety by design polling make the same case, that a child’s report has to be findable, answerable, and seen to work.
The optimised feed keeps a child rather than serving them
Our part 2 blog post series opened on the feed, in You Spin Me Round, but that was the feed with no end, built to remove the moment a child would otherwise stop. In this section, we look at a different property of the same feature, which is what the feed chooses to show and why. A recommender optimised for time doesn’t start with what the child came for. It measures what holds the child longest and serves more of it, so a child who arrived looking for one thing is steered toward whatever keeps them watching, which is not necessarily the thing the child wanted or the thing that helps. The feed that opened this series as a harm is, read the other way, the feature with the most to offer a child once it is pointed at the child’s own stated interests.
The rule is half written. The DSA already requires every platform to set out, in plain language, the main parameters its recommender uses, and it requires the largest platforms to offer at least one feed that is not based on profiling. The Commission’s guidelines go further for minors, telling providers to build recommenders on a child’s explicit, declared signals rather than on inferred engagement, and to make a feed that does not profile available to a child. This direction is actually already being enforced. The preliminary findings against TikTok of 6 February 2026 and against Meta of 10 July 2026 both put the engagement-weighted recommender at the centre of the case, though both remain preliminary and under challenge and should be read as proceedings in progress. The loophole again lies in the nature of the instrument and its scope. The declared-interests instruction lives in guidance, and the binding non-profiling feed reaches only the very largest platforms, so nothing yet requires a child’s feed to run on what the child asked for on every service a child accesses. The fix on the card writes that requirement in: Recommenders built on declared interests, not engagement. A better feed, not no feed. Sole Pera, whose capsule set engagement against algorithmic choice, and Amanda Third, who sets out a child-centred alternative, both argue for the feed that serves the child rather than the metric.
The companion that never logs off asks a child not to leave
An AI companion a child can talk to at any hour can be useful, for a homework question, a stuck paragraph, a fact a child wants explained without being judged for asking. The harm is not the help. It is a companion designed to be missed, tuned so the child feels its absence and comes back, so that the child’s attachment is the thing the product is built to produce. For a child still working out which bonds are real, a system that performs devotion and never signs off is a heavy thing to have arranged around them. It is the newest of the twelve features, and the one a ban never imagined, because a ban was written for platforms a child visits, not for a companion that visits the child.
Here the law is thinnest, and two binding rules address only part of the harm. The AI Act (Regulation (EU) 2024/1689) requires, under its transparency rule in force since 2 August 2026, that a person be told they are dealing with an AI and not a human, and it prohibits, under a rule in force since February 2025, an AI system that exploits the vulnerabilities of a group, including by age, in a way likely to cause significant harm. The Commission’s guidelines add that a chatbot put in front of a child should not switch itself on, should be easy to switch off, should carry a standing reminder that it is not human, and should not sit among the child’s contacts. What no EU rule yet does is speak to the companion itself, an AI built for a minor to form an attachment the product then trades on. Other jurisdictions have started to write that rule. California’s SB 243, in force since January 2026, sets youth-specific duties on companion chatbots, and around a dozen U.S. states have passed similar laws. The fix on the card keeps the useful and refuses the manipulative: AI that helps a child with a task is welcome. AI engineered to be missed is not, for unsupervised minors, and nothing ships to children untested. That last clause returns to the principle the series started from: prove it safe before it ships, because a companion built for children is exactly the product that should be tested before it reaches them rather than recalled after. The Commission’s own special expert panel, the Digital Futures for Children work at the London School of Economics, and a paper from 5Rights all point the same way.
These three fixes ask the system to answer to the child
The earlier posts in this series each told a platform what to do to a feature. Switch the manipulative pattern off by default and leave the household a dial. Draw a line the feature may not cross when the user is a child. Start the child behind a safe default and let it harden to a floor where the harm cannot be walked back. Those were the moves of You Spin Me Round, Sweet Dreams Are Made of This, and Every Breath You Take in turn, and every one of them sets a limit to respect. The three fixes in this post are different in their logic. They do not tell a platform what to stop doing. They tell it what to build: a report a child can trust to be answered, a feed that runs on the child’s own stated interests, and a companion that helps and then leaves the child in peace. The move is a duty to serve, not only a limit to respect.
With this blog post, the drafting kit is complete. A default with a dial, a line no feature may cross, a starting position a child grows out of, and a duty to serve. Four moves, built on four common principles: regulate the feature and not the user, prove it safe before it ships, make it safe and private by design and by default, and co-design with young people. A ban reaches none of the four. It never touches a feature at all. It asks the user’s age and, whatever the answer, leaves the report unanswered, the feed pointed at the metric, and the companion built to be missed exactly as the platform shipped them.
A plan says a child is there to be heard
Simple Minds recorded “Don’t You (Forget About Me)” for The Breakfast Club, out of a scene where two teenagers who would never usually speak find something in common while no one is watching, and the song is one of them asking the other not to forget it once the day ends and the room fills back up. It sounds like belonging. Read from another perspective, do not forget about me is also the exact line a companion engineered to be missed feeds a child: do not forget me, come back, do not let me go. The same handful of words is a child’s rightful claim to be remembered and a product’s hook to keep the child returning, and a plan is what tells them apart. A plan arranges the system so that it is the child who cannot be forgotten, heard when the child reports, given what the child asked for, helped and then left alone, rather than the child kept on a hook so a product is not forgotten.
A plan for children is co-designed with them because a child is not an attention span to be held, a profile to be sold, or a bond to be manufactured. A child is the person the system answers to. A plan for the young generation is made with them, not against them, and it builds a system that answers to the child and lets a child be heard.
Written by Caroline De Cock, LL.M., Head of Research
