Microsoft accepts 10 binding rules on AI in classrooms to reopen a market that bans were closing

By Zendoric · September 11, 2026.

Microsoft has committed, through agreements with two of the largest teachers' unions in the United States, to abide by ten legally enforceable principles on the use of artificial intelligence in schools. As reported by GuruFocus, the company will not use student or teacher data to train its models, will limit data collection and must explain how its systems work to families in plain language. Also banned are «AI companions» —chatbots designed to simulate emotional or companionship bonds— and any decision with serious consequences for a student must undergo human review. Microsoft shares rose around 0.7%, to $495.93, on Friday morning.

The timing is no accident. New York and Los Angeles had temporarily restricted the use of student-facing AI tools in their districts, and that kind of local veto is the worst possible news for a provider whose business depends on selling software under contract to public administrations. From November, the new commitments can be folded into existing or new contracts without reopening the full agreement, which drastically lowers the cost of saying «yes» for a cautious school district. GuruFocus reports no figures on adoption, contract value or revenue impact: for now this is a bet on trust, not a commercial announcement.

What is interesting is not that Microsoft promises to behave responsibly with minors' data —US educational privacy law already largely requires that— but whom it is promising and how. It is not a federal agency imposing these rules: it is teachers' unions, contractually, in the absence of a specific federal law on AI in the classroom. It is the same pattern we have been observing on other fronts: with no framework law, AI governance in the US is advancing district by district, lawsuit by lawsuit, union by union. It is regulation by specific application, not by model, and child protection is confirmed as one of its most active axes, and the one with the broadest political consensus.

For the rest of the sector, this sets a benchmark. Google, OpenAI and any other provider hoping to sell AI to US public schools now has a reference catalogue —no training on minors' data, plain-language explanations of systems, a ban on AI companions, mandatory human oversight— against which districts, and their unions, will be able to measure any future offer. Microsoft has not done this out of generosity: it has done it because the school market was closing on it, and trust, turned into a contractual clause, is cheaper than losing entire districts.

This connects with something we have long argued at Zendoric: distrust of AI in the classroom is not unfounded alarmism. AI companions for minors are a documented risk —there have been lawsuits and real cases of psychological harm linked to chatbots that simulate affection— and Microsoft's acceptance of an explicit contractual ban is a tacit acknowledgement that the problem was real, not hypothetical. In the short term, friction and local bans are a legitimate brake, not a regulatory whim. But in the medium term, if governance is done properly —with real auditing, not just for show—, AI-assisted personalised tutoring is exactly the kind of tool that can free up teachers' time for what no machine can replace: the human relationship with the student. The road to that educational abundance runs, paradoxically, through accepting more controls and less room for manoeuvre over data today, not fewer.