Most states are still writing guidance on artificial intelligence in schools. Arizona approved a school built on it.
Unbound Academy, an online only charter serving grades four through eight, was cleared by the Arizona State Board for Charter Schools on a 4 to 3 vote. Its model compresses core academics into roughly two hours a day, from 8:30 to 11:00 in a school day that runs 8 to 2, delivered through AI powered educational software.
The rest of the day goes to workshops in things like financial literacy, public speaking and entrepreneurship.
It does not employ classroom teachers in the usual sense. It employs guides, who coach students through the software, step in when someone is stuck, and run the afternoon sessions. The approach comes from Alpha School, a private school in Austin, Texas, that has used it since 2016. Unbound is open and enrolling now.
Whatever one makes of it, it is the most direct test in the country of a question the research has just started to answer properly. And the research says the details decide everything.
The trial that split the difference
Researchers gave roughly a thousand high school students access to GPT-4 during mathematics practice. Practice scores rose 48 percent against students working without it.
Then the tool was withdrawn for an exam. Those students scored 17 percent worse than classmates who had never used it.
The study, published in the Proceedings of the National Academy of Sciences, found the students had used the model as a crutch. What makes the result genuinely difficult is that nothing looked wrong while it was going on. Practice scores, homework quality and engagement all improved. The deficit appeared only once the tool was absent.
The configuration that changed the outcome
A third group used the identical GPT-4 model, instructed to supply incremental hints and never the full answer.
They improved the most during practice, by 127 percent, and showed no significant harm on the exam afterward. Same model, same students, same material. The only difference was whether the software would hand over an answer.
That is the finding to hold against any AI driven school model, including one operating in the East Valley. The question is not whether software can teach. It is whether the software is built to make the student do the thinking, and whether an adult is checking that it does.
What the supporting evidence suggests about design
The strongest positive result available happens to point the same direction.
A World Bank randomized trial in Benin City, Nigeria, put first year senior secondary students through six weeks of after school English sessions using Microsoft Copilot, supervised by teachers. The gain was 0.31 standard deviations across the full assessment, which covered English, AI knowledge and digital skills, and 0.23 on English alone, the study's main outcome.
The researchers report the program outperformed 80 percent of the interventions in a comparison database of randomized trials in developing countries.
The two features that stand out are that the sessions were structured and that a teacher supervised them. That is much closer to the hint giving arm of the Turkish trial than to a child working alone.
