California decided some time ago that the answer to artificial intelligence in schools was to teach it rather than fight it. The research that has arrived since suggests that instinct was probably right, and that the details will decide whether it works.
Start with the sharpest finding available. Researchers gave roughly a thousand high school students access to GPT-4 during mathematics practice. Practice scores rose 48 percent. Then the tool was removed for an exam, and those same students scored 17 percent worse than classmates who had never used it at all.
The study, published in the Proceedings of the National Academy of Sciences, describes the students using the model as a crutch.
Why that result is worse than it first sounds
The failure was invisible while it was happening.
Practice work improved. Homework looked better. Engagement looked better. Every number a school actually collects pointed upward, and the loss only surfaced on a test where the tool was absent. A district measuring assignment completion through an AI saturated year would record success while measuring the software instead of the student.
The same model, configured differently, did no harm
A third group in the study used the identical GPT-4 model with one change. It had been instructed to offer incremental hints and never hand over the answer.
Those students improved most during practice, by 127 percent, and showed no significant deficit afterward. The technology was not the variable. The instruction given to it was.
What California has actually required
Assembly Bill 2876, signed by Governor Newsom on September 29, 2024, folds AI literacy into the state's core curriculum work. It directs the Instructional Quality Commission to consider incorporating AI literacy into the mathematics, science and history social science frameworks as those are revised, and into the criteria used to evaluate instructional materials.
The framing in the bill covers both halves of the problem: learning about AI, meaning how it works and how it affects society, and learning with it, meaning using it effectively and ethically. The California Department of Education has since published statewide AI guidance for districts, county offices and charter schools, developed with an artificial intelligence working group.
That is a more considered starting point than most states have. It also leaves the hardest question open, because a framework can require that students learn about these tools without specifying whether the tool in front of a child hints or answers.
The Stanford finding that matters most in the Tri-Valley and the South Bay
There is a second research result with unusually direct local relevance, and it concerns enforcement rather than learning.
Stanford researchers, led by Weixin Liang with James Zou and colleagues, tested widely used GPT detectors against essays written by human non-native English speakers. The detectors consistently misclassified that writing as AI generated. More than half of the non-native TOEFL essays were flagged. On essays by native speaking US eighth graders, the same detectors were nearly perfect.
The mechanism is unforgiving: the tools read the plainer sentence construction of a second language writer as machine output. The researchers showed the bias could be removed simply by prompting for more varied phrasing, which means the detectors were penalizing limited linguistic range rather than detecting cheating.
In districts where a large share of families speak a language other than English at home, that is not an abstract fairness concern. It is a prediction about which students get accused. Universities have acted on exactly that: Vanderbilt disabled Turnitin's AI detector in August 2023, citing its reliability, false positives and the disparate impact on international students.
