Trump’s Executive Order 14277 is the most aggressive federal push into classroom technology in American history. The evidence it’s built on is thinner than anyone is admitting.
In June 2025, researchers at Wharton published a randomized controlled trial in the Proceedings of the National Academy of Sciences. They gave roughly 1,000 high school students unrestricted access to GPT for math tutoring. While using the AI, students improved 48% on practice problems. When they took unassisted exams afterward, they scored 17% worse than the students who had studied without AI.
The mechanism was not complicated: students who had access to GPT mostly used it to get answers, not to think through problems. The AI provided correct answers only 51% of the time and made logical errors 42% of the time. Students who were nominally doing better during practice had actually learned less. The tool had done the cognitive work for them.
This study was published three months after President Trump signed Executive Order 14277: “Advancing Artificial Intelligence Education for American Youth.” The two things coexist in the same policy environment, mostly without touching each other.
What the Order Actually Does
EO 14277, signed April 23, 2025, is ambitious on paper. It created a White House Task Force on AI Education, chaired by OSTP Director Michael Kratsios and including the Secretaries of Labor, Energy, and Education, the NSF Director, AI and Crypto Advisor David Sacks, and (added after signing) Melania Trump. The order directed the Department of Education to issue guidance on using existing grants for AI tutoring resources, prioritize AI in teacher-training grants, and build K-12 AI literacy resources within 180 days. The NSF was directed to reprioritize AI education research. A Presidential AI Challenge for students was to be fully operational within 12 months.
The DOE issued its implementing Dear Colleague Letter in July 2025. The Final Supplemental Priority Rule, the piece that actually redirects federal grant competition toward AI education, took effect in May 2026.
The mandate is real. What is not real is any new money to carry it out.
EO 14277 appropriated zero new dollars. Every requirement runs through the reprioritization of existing grants: ESEA, HEA Title II, WIOA, and NSF discretionary funding. Schools were told to integrate AI; the federal government would help them redirect funds they were already getting. The assumption was that the private sector would do the rest.
The private sector obliged. Google pledged $150 million in AI education grants and made Gemini free for all U.S. high schools. Microsoft deployed free Copilot for every K-12 student and teacher. Amazon, OpenAI, NVIDIA, Meta, and xAI joined a coalition of more than 200 organizations that collectively pledged over $1 billion in AI education support.
Whether you call this corporate philanthropy or market development depends on your priors. Either way, it is the mechanism through which federal AI education policy is actually being implemented: a mandate with no funding, filled in by the companies selling the products.
The Irony in the Infrastructure
Here is what makes EO 14277 structurally unusual: it tasked the Department of Education and the National Science Foundation with leading AI education implementation while simultaneously dismantling the infrastructure those agencies use to do policy work.
The Trump administration cut the DOE by roughly 50%, from approximately 4,054 to 2,000 full-time employees. The Office of Educational Technology u2014 the unit that would ordinarily be responsible for evaluating and implementing something like an AI education mandate u2014 was eliminated. The NSF saw more than 400 STEM education grants canceled.
Former DOE Senior Policy Adviser Bernadette Adams put it plainly: the EO “lacks provisions on student data privacy and bias,” and the staff cuts mean “you need people who know how to do that,” and they have been let go.
The result is an executive order that told federal agencies to do more with considerably less, while filling the implementation gap with the companies being regulated.
What the Research Actually Shows
The Bastani et al. study is the sharpest data point, but it is not alone.
A 2025 randomized trial in the Philippines found that students in an AI-assisted learning group scored 57.5% on a retention test 45 days after instruction. The non-AI control group scored 68.5%.
OECD’s PISA 2022 assessment, one of the world’s largest and most rigorous international education studies, found that math scores fell 15 points globally compared to 2018, the largest decline in PISA history. Sixty-five percent of students reported being distracted by devices during math class.
The RAND Corporation’s 2025 assessment of AI in education concluded that “whether AI will improve learning u2014 the answers are not yet clear and likely will not be for some time.”
UNESCO, in its 2025 guidance on AI in schools, was more direct: “Generative AI arrived in schools without evaluation, assessment of risks and benefits, or training for educators.”
None of this is an argument that AI cannot be useful in education. Some tools, deployed carefully with trained educators, probably do help some students in some contexts. The research is genuinely mixed, not uniformly negative. But “mixed” is a very different claim than the one embedded in EO 14277, which treats AI integration as a national competitiveness imperative and moves to implementation before the evidence base is established.
The Union That Fought the Order, and Then Joined It
The American Federation of Teachers, one of the largest teacher unions in the country, was publicly critical of EO 14277. AFT President Randi Weingarten warned it was opening schools to “unaccountable tech companies” pushing “unproven software” into classrooms.
She is not wrong about the evidence.
She is also, simultaneously, a partner of those companies. In July 2025, the AFT announced a joint initiative with OpenAI, Microsoft, and Anthropic to build the National Academy for AI Instruction, a physical facility in New York City intended to train 400,000 teachers in AI tools. OpenAI and Microsoft, specifically, are among the companies whose products Weingarten was warning against.
This is not hypocrisy so much as the predictable outcome of a policy environment that creates mandates without resources. If the federal government is directing schools to adopt AI and the AFT wants a voice in how that happens, partnering with the companies delivering the AI is how you get a seat at the table. The alternative is watching it happen without you.
What It Looks Like on the Ground
The Salamanca City Central School District in western New York is one data point in what EO 14277 looks like in practice. Salamanca, with 1,300 students (79% economically disadvantaged) on the Seneca Nation reservation, spent $57,590 on a hyperrealistic humanoid robot from Realbotix, a company that previously operated as a cryptocurrency firm and whose parent company still manufactures RealDoll sex dolls.
Sally, as the robot is called, will deploy as a classroom teaching assistant this fall.
Nobody at the federal level directed Salamanca to buy Sally specifically. But the policy environment (aggressive federal push into AI adoption, tech industry eager to find school customers, grant structures being redirected toward AI integration) created conditions in which a school board decided a $57,000 humanoid from a sex doll company was reasonable classroom infrastructure.
EO 14277 does not mention robots. It does not mention any specific technology. It creates a mandate and a structure, and the market fills in the rest.
The Bipartisan Problem
It would be convenient if concerns about EO 14277 broke cleanly along party lines. They do not.
The RAISE Act u2014 legislation proposing guardrails on AI in education u2014 is sponsored by Democratic Senator Lisa Blunt Rochester of Delaware and Republican Senators Jon Husted of Ohio and Bill Cassidy of Louisiana. The NSF AI Education Act similarly carries bipartisan cosponsors.
When a December 2025 Trump executive order attempted to preempt state AI regulations, sixteen Texas state senators opposed it: eight Republicans and seven Democrats. The lead quote came from Republican Senator Angela Paxton: “I want Texas to be a leader in AI, but I don’t want Texas to be a leader in AI generated child porn.”
The concern is not that AI has no place in education. The concern is that a federal mandate moving faster than the evidence, with no new funding, no privacy guardrails, and reduced regulatory capacity, creates conditions for exactly the kind of outcomes the research has started to document: students who appear to be using tools effectively but are actually offloading the cognitive work that makes learning stick.
The Gap Between Policy and Evidence
The Bastani study published its findings in a peer-reviewed journal while EO 14277 was already nine weeks old. The federal government was not in a position to wait for better evidence before establishing AI education as a policy priority. The competitive logic of AI development does not allow for that kind of patience.
But there is a difference between a policy environment that acknowledges the uncertainty and structures implementation carefully, and one that moves to mandate first and research later.
The 58% of teachers who had received zero professional development on generative AI before EO 14277 was signed are now being told that AI integration is a national competitiveness imperative. The schools serving low-income and rural students u2014 often the first to receive experimental technology deployments u2014 are making purchasing decisions in a policy vacuum.
The evidence will catch up eventually. In the meantime, the robots are showing up.
Sources: Bastani et al. (PNAS, 2025, DOI: 10.1073/pnas.2422633122); White House EO 14277; DOE Dear Colleague Letter (July 2025); PISA 2022 (OECD); Li et al. (JAMA Network Open, 2025); RAND (2025); UNESCO (2025); Chalkbeat; EdWeek Research Center (Oct 2024); AFT/OpenAI/Microsoft partnership announcement (July 2025).
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com
