What meaningful safety really looks like in youth-facing platforms

In the race to prove compliance with youth safety laws like COPPA, GDPR-K, or the Online Safety Act, it’s easy for digital platforms to treat trust and safety like a checklist. But what if the real risk isn’t non-compliance? What if the real risk is designing safety systems around an imaginary version of childhood that doesn’t reflect how kids actually behave online?

Agnes Evrard, former Director of Trust & Safety at Brainly, has spent her career questioning assumptions baked into platform design. “A lot of companies think about policy as just a version of compliance,” she told Ailís Daly, Head of Trust & Safety for EMEA at WebPurify, an IntouchCX company. “But if you’re working with children, you have to understand child development. You have to understand the way that children communicate and learn.”

Evrard’s work goes beyond regulation. It probes a deeper question: what does it actually take to make a digital space that’s safe for kids? She argues the answer isn’t found in a ‘terms of service’ update, but rather it lives in moderation logic, UX language, community reporting structures, and perhaps most critically, in the willingness to observe how young people behave in real time.

Meaningful safety doesn’t begin in the legal department, Agnes argues. It takes shape in design sprints, community testing, and the quiet decisions teams make about what to prioritise and who they’re really building for.

What meaningful safety actually looks like

To Agnes, the phrase “child-safe” has become dangerously diluted. Too often, it’s conflated with parental controls, rigid filters, or worse, blanket bans that strip away nuance. But these approaches fail to account for what kids are actually doing online.

She points to AI prompt misuse as one example. “You might think your AI prompt is going to get people to be more productive or learn better,” she says. “But children are not going to use it like that. They’re going to be silly. They’re going to try to get it to say a swear word.”

That behaviour isn’t malicious. It’s exploratory. And it reveals a key insight: safety is more about prediction than it is about restriction. If you can’t anticipate how kids will stretch, bend, or subvert your tools, you’re not really building for them.

Brainly’s guidelines and systems

At Brainly, Evrard helped redesign community guidelines and enforcement systems to better reflect how young users interact. But what made the work effective wasn’t just the language. It was the framing.

Instead of punitive commands, the team opted for contextual, supportive cues. “It was really important to make it welcoming, especially for children who are in an environment where they may be struggling already,” she said. This meant surfacing shorter, more readable policies and pairing them with tone-appropriate copy that guided behaviour rather than reprimanded it.

It also meant working closely with educators and behavioural experts, whose insights helped strip out the assumptions that often creep in when adult teams try to write for kids.

She also flagged a common industry pitfall: reusing or adapting policies from other platforms without revisiting core risk models. “I remember a time when companies were hiring from almost cheating talent – being like, ‘Oh, you have experience there, come write my policies,’” she recalls. “But those can’t just be recycled. If policies are tailored to the need of the platform or the product, then it becomes your first line of defense.”

That means understanding the full scope of your platform’s ecosystem. “What are you trying to sell? Who are your users, who are your vendors?” she asked. “The risk model for a marketplace is not the same as the one for an educational tool. You have to map the full ecosystem and that includes your moderation architecture.”

Moderation that actually understands kids

Most content moderation systems weren’t built for children. They were built to manage adult threats: harassment, hate speech, spam. So when those same classifiers are applied to a platform full of 13-year-olds, false positives skyrocket and the real issues often slip through the cracks.

“You can’t just repurpose adult classifiers,” Agnes says. “You’ll either miss the real risks or you’ll flag everything and frustrate your users.”

She highlights how children communicate in layered, often coded ways: emoji chains, invented slang, playful aggression. If your filters aren’t calibrated to that context, moderation becomes a blunt instrument. At best, it alienates users. At worst, it creates safety theater.

Agnes also warns against borrowing moderation practices from adult social platforms. Their risk assumptions and communication patterns differ vastly from those in education-focused or youth-specific spaces.

So what is the solution? Agnes advocates for building moderation logic specifically for youth spaces. That means new taxonomies. Different severity thresholds. And human reviewers who actually understand the population they’re moderating.

But Agnes is clear that this doesn’t mean AI has no role. In fact, when policies are written in clear, readable language — where definitions are precise and context is well described — AI systems thrive and can be surprisingly good at catching nuanced behaviours. But she cautions that no amount of machine learning can replace human instinct in ambiguous situations.

“AI isn’t very good with any kind of sensitive escalation or unclear workflows where instinct and judgment matter,” she explains. “That’s where you need trained humans who can understand the context, ask questions, and make thoughtful decisions.”

Assumptions are the real risk

What separates meaningful safety work from checkbox compliance isn’t budget or headcount. It’s epistemology. Evrard makes this clear: if your safety policies are based on what adults think children are doing, you’ve already lost.

At Brainly, her team treated students, moderators, and even parents as co-designers. “We would speak to moderators who were students themselves, and their feedback was some of the most valuable we got,” she says. “They helped us see which rules made sense and which didn’t.”

This isn’t a feel-good anecdote. It’s a governance principle. Platforms can’t protect what they don’t understand and they can’t understand unless they ask the right people.

The future is designed, not legislated

Regulation will always be the baseline. But if that’s the only benchmark a platform is aiming for, it’s failing its youngest users, Agnes argues.

She sees meaningful safety as a product function, not a legal one. It demands cross-functional design. Behavioural insight. Curiosity. And a willingness to question the industry’s most comfortable defaults.

“You can make your platform safe and welcoming,” she says. “But only if you build it with a real understanding of the kids who are using it.”