Learning AI from its failures: American school districts rely on proof of error

Thirty-seven states have now issued guidelines. In Utah, an eight-person team has trained 7,000 teachers, or one-third of the state's public teaching staff.

The U.S. re-entry confirms a pedagogical shift: instead of banning or bragging chatbots, a growing number of districts choose to make them fail in front of students. Hallucinations, biases, factual errors: the most cited demonstration is to ask a model for a world map and let students see the distorted country names it produces.

Thirty-seven states have now issued official guidelines on the use of AI in the classroom. Utah is a laboratory: a team of eight in the Canyons School District conducted more than 150 sessions and trained approximately 7,000 teachers, nearly one third of the public instructors in the state, with a $500,000 grant. Maine, West Virginia and Georgia have since taken over the model.

Whole districts on board

The Charleston County School District in South Carolina applies the approach to its 50,000 students; Chicago Public Schools and Houston relied on the recommendations of AI for Education, led by Amanda Bickerstaff. The issue at Brookings Institution's Rebecca Winthrop is less the tool than the skill at hand: learning to doubt a plausible answer remains, at this point, the most difficult part to teach.