The document, which was twenty-eight pages long and written in the meticulous, almost bureaucratic style of an international organization that has been conducting this kind of work since 1948, arrived on a Thursday. OMEP, the French acronym for the World Organization for Early Childhood Education, is not the kind of organization that typically unnerves people. It hosts conferences in cities like Nairobi and Athens.
Directors of kindergartens, researchers studying child development, and occasionally retired ministry officials make up its membership. Nevertheless, two product managers at a significant San Francisco AI lab were discreetly instructed to halt their roadmap for an under-six chatbot companion a few days after the position paper was released. Confirmed by someone who saw the internal Slack thread, that detail is the kind of thing that indicates something has changed but doesn’t make headlines.
On the surface, the paper’s actual content is not radical. It makes the case that AI products made for or sold to kids under the age of six ought to fall under a different regulatory category than the more general edtech category. It demands restrictions on data collection, limitations on emotional simulation, and a categorical rejection of any product that presents a chatbot as a friend or companion to a child who is still figuring out what a friend is. But after reading it twice, the tone becomes more acute. The writers seem to have been holding back for a very long time.
The timing is important. A few well-funded startups have been racing for the past eighteen months to create what one founder described as “the Duolingo for toddlers” in a since-deleted tweet. When a parent is at work, some of these products mimic their voice through voice cloning. Others pledge to use a tablet camera to monitor a child’s emotional state and modify the lesson in real time. Investors appear to think that this is a developing category. The OMEP paper basically makes the case that the category shouldn’t exist at all in its current form, citing recent scoping reviews and the Brookings policy guardrails work from late 2025.
It’s difficult to ignore Silicon Valley’s lack of readiness for this kind of criticism. The industry has practiced responses for parents who are worried, regulators who are skeptical, and academic ethicists. However, early childhood educators occupy a peculiar middle ground: they are highly credentialed in a field that most tech executives are ignorant of, trusted, and nearly impossible to mock as Luddites. Before the paper was released last month, I spoke with a venture partner who informed me that the children’s AI category was “the safest bet in consumer.” Since then, he hasn’t said much.
Overanalyzing the situation can be dangerous for a number of reasons. No matter how persuasive they are, position papers do not enact legislation. In product circles, where personalization is viewed as a self-evident good, the OMEP document heavily relies on Vygotskian frameworks and the social aspect of early learning. Additionally, the paper is overly cautious in some places, avoiding the more difficult issue of what to do about the AI-generated content that already permeates the platforms used by young children, a problem that Bloomberg painstakingly detailed in April.
Nevertheless, something is taking place. The paper has been referenced in preliminary briefings by three European education ministries. A translation was requested by the office of a US senator. After being quietly circulating, the UNESCO Blue DOT issue on AI in early childhood is now being downloaded five times more frequently than usual. It’s still unclear if any of this will become policy. It’s more obvious that the discussion has shifted and that there are now other people in the room besides the manufacturers of these goods.
