On May 29, in Paris, the G7 digital ministers signed a joint statement on the protection of minors online—a first. Three weeks later, in Évian, the question is no longer whether children should be protected from artificial intelligence, but how to prove it.
In 1959, Swedish engineer Nils Bohlin developed the three-point seat belt. Volvo, which held the patent, chose to make it available to the entire automotive industry free of charge: a manufacturer voluntarily gave up a monopoly so that safety could become a shared standard rather than a competitive advantage. Sixty-seven years later, on the Viva Tech stage, this anecdote resurfaced in the words of Stéphie Herlin, a member of the founding team of KORA Benchmark—the first independent, open-source framework measuring the impact of artificial intelligence models on children. The parallel is significant: it highlights the difference between a statement of intent and technical proof.
What Twenty-Six Risks Reveal
KORA subjected 39 models and 11 applications to more than 100,000 simulated conversations, which were evaluated based on 26 risk criteria defined in collaboration with about 30 child protection experts. The result: an average score of 39%. The gap between the models is wide—ranging from 6% to 70%, depending on the systems tested. An even more unexpected finding, according to Stéphie Herlin: the safety of some models declined over time, even as their capabilities improved. The explanation can be summed up in one sentence: what isn’t measured doesn’t improve.
This is the same principle championed by Camille François, founding president of ROOST (Robust Open Online Safety Tools), an organization that develops open-source moderation and detection tools. Her analogy, also drawn from the automotive industry: selling a car by saying, “I think there’s a seatbelt—maybe it works,” would be deemed unacceptable by any regulator. When applied to digital platforms aimed at children, this standard of proof remains, for now, the exception rather than the rule.
Details. The Digital G7, under the French presidency, convened its ministers on May 29, 2026, in Paris ahead of the Evian summit. The adopted text—the first set of shared principles for a safer digital space for minors—was endorsed by G7 partner countries, including Kenya, Brazil, Egypt, India, and South Korea.
Diplomacy as a Way to Prove Something
For Anne Le Hénanff, Minister Delegate for Artificial Intelligence and Digital Affairs, the issue goes beyond simply restricting access: it is about establishing, through legislation, enforceable safeguards—non-intrusive age verification, an absolute ban on the generation of child sexual abuse material by AI, and the design of products with protection in mind from the outset. Clara Chappaz, who has served as Ambassador for Digital Technology and Artificial Intelligence since December 2025 after previously holding that same ministerial portfolio, promotes this set of principles in diplomatic forums, from Japan to South Korea to Kenya, where the issue of online child protection found an unexpected place on the agenda of an economic forum dedicated to African investment.
In France, a study by the Observatory on Digital Parenting published in February reveals that 74% of parents support banning social media for children under the age of 15—a figure that ministers now cite as a mandate rather than a political gamble. The same survey reports a more nuanced finding, gathered directly from teenagers themselves: many of them are waiting for just one thing before leaving a platform—for their friends to leave at the same time. In short, the algorithm sometimes holds a stronger grip than the will of the user or that of their parents.
What Remains to Be Measured
None of the speakers claimed to have a definitive answer regarding the effects of conversational artificial intelligence on children’s development. As Stéphie Herlin points out, scientific research takes time—time that political urgency does not allow. What can be demanded right now, however, is access to data: without it, tens of millions of minors continue to interact every day with systems whose impact remains, by design, unverifiable.
The seat belt eventually became a legally mandated standard, after having been an optional choice. Will artificial intelligence designed for children follow the same trajectory—from voluntary adoption to a mandatory standard—or will the measure remain just one argument among many in a race that is, incidentally, said to have no boundaries?
Cette publication est également disponible en :
