Schools stop banning AI, start teaching kids where it gets things wrong
A growing number of U.S. schools now use flawed chatbot answers as teaching tools, while a new classroom trial shows AI tutors help but also slow kids down.
What happened: After trying outright bans, a growing number of U.S. public schools are inviting AI into class on purpose — partly to show students where it breaks. At a Charleston County, South Carolina, teacher training this summer, an instructor asked an AI tool to draw a world map; it mangled country names into gibberish (Mali became "Mail," Egypt became "Sopth"), drawing laughs meant to teach kids not to over-trust chatbots. Thirty-seven states now publish official AI guidance for schools. Separately, a trial of 6,997 Tennessee middle-schoolers using a guardrailed AI math tutor found real but narrow tradeoffs: better recovery after mistakes, slower work, and only marginal gains a week later.
Why it matters: Surveys cited by researchers show most teens and teachers already use AI for schoolwork, but taught themselves, with little formal training in how it actually works or when to doubt it. Educators explicitly want to avoid repeating social media's mistake — letting kids explore powerful, addictive technology unsupervised until the harm is already visible in mental health and attention spans. The real question isn't whether students use AI, but whether they learn to question its confident-sounding wrong answers, and whether AI tools built into schoolwork actually deepen understanding or just speed up guessing.
What 'AI literacy' actually teaches: Programs like Charleston's train teachers and students to spot hallucinated facts, verify sources, and avoid feeding personal data to chatbots that store and reuse conversations. Utah's statewide model has trained over 7,000 teachers, negotiated discounted tool pricing for districts, and requires school AI policies by July 2027. Separately, the Tennessee NUMI trial shows what careful design can and can't do: a tutor built to withhold answers and force step-by-step reasoning raised next-attempt accuracy by 8.5 points after a mistake, but slowed practice and produced only a borderline, not statistically strong, gain on a delayed test.
The rollout: Coverage is uneven. Utah, Maine, West Virginia and Georgia have created full-time state AI-education roles; South Carolina has none, so its second-largest district built its own policy with an outside nonprofit's help. Training costs money — one Utah district's statewide outreach ran on a $500,000 grant from a local healthcare provider. The Tennessee researchers call their finding a first test, not proof: students used the tutor for one 50-minute session, and the team says repeated use over weeks is the next thing to check before schools can trust embedded AI tutors to change real learning, not just classroom behavior.