AI in US Schools: New Research Sharpens the Debate Over Learning and Student Privacy

Artificial intelligence in US schools is facing a more specific test: whether it helps students build skills they can still use when the chatbot is gone.

An October 2026 expert-consensus report hosted by the University of Minnesota examines how generative AI could support or disrupt learning. It arrives after New York City announced a one-year moratorium for younger students and Microsoft joined teachers' unions in announcing school-specific privacy protections. These developments address different questions: learning, permission to use a tool, and control over the data it collects. University of Minnesota report, NYC announcement, Microsoft announcement.

What has changed in the US school AI debate?

Three developments make the current debate concrete.

The developments point toward a more demanding form of adoption. Schools must specify who uses the system, for what task, under which protections, and how success will be measured.

There is a substantial difference between a teacher using AI to draft a worksheet and a child using it to produce the reasoning the worksheet was designed to teach. A policy that treats both as the same activity misses the educational issue.

Why New York restricted generative AI for younger students

New York City's September 2 announcement pauses student-facing generative AI for children in 2-K through eighth grade during 2026–2027, affecting nearly 600,000 students. Companion chatbots are prohibited across all grades.

Approved teacher uses continue. Exceptions cover assistive technology, multilingual learners and certain career-readiness programs. High school students receive AI literacy modules, and selected classrooms can run supervised pilots. A stakeholder coalition will evaluate the policy. New York City's official policy announcement.

The policy distinguishes AI literacy from unrestricted chatbot access.

What the October research says about AI and student learning

The University of Minnesota's October report involves 30 learning scientists and educational researchers. They examined 65 learning processes, reaching consensus on 42.

They identify practice as a promising stage for appropriately prepared students to use AI. Delegating thinking during initial knowledge acquisition or later application can undermine skill development, they warn.

These are structured expert judgments, not trial results for a particular product. The report also highlights weaknesses in the quality of existing educational AI research. Expert consensus on generative AI and learning.

For procurement, the report raises a practical question: which learning process does the product actually support?

Can AI improve homework performance while weakening learning?

A randomized field experiment published in PNAS in June 2025 illustrates why those questions matter.

Researchers studied nearly 1,000 high school mathematics students in Turkey using two GPT-4-based systems. One resembled a general chat interface. The other incorporated teacher-designed safeguards, including hints intended to avoid giving away the solution.

Both improved performance during assisted practice. But when students subsequently completed an exam without AI, the general-interface group scored 17% lower than the control group. The guarded tutor largely removed that negative effect, without demonstrating an improvement over the control group on the unassisted exam. PNAS study.

This was a study in Turkey, not a measurement of all US schools. It used specific tools, mathematics tasks and a defined testing period. It does not establish that every AI tutor harms learning or that newer systems behave identically.

It does demonstrate why assisted task performance and independent learning should be measured separately. An impressive answer produced with a chatbot is not sufficient evidence that the student has acquired the underlying skill.

Students are using AI more, even while questioning its effects

RAND's 2026 report on the American Youth Panel presents another tension.

Among surveyed students in middle school grades and above, reported use of AI for homework help rose from 48% in May 2025 to 62% in December 2025. In December, 67% endorsed the view that greater AI use for schoolwork would harm critical-thinking skills.

The report draws on 1,214 respondents aged 12 to 29. Older students more often reported teacher-dependent rules. RAND's American Youth Panel findings.

These are self-reported behaviours and beliefs. The 67% figure is not a measured decline in critical thinking.

RAND recommends explicit guidance on when and for which purposes students may use AI.

A practical response is assignment-specific guidance distinguishing permitted feedback from submitting generated answers.

Why teachers' time savings do not settle the student-use question

The case for teacher-facing AI has a different evidence base.

A Gallup and Walton Family Foundation study published in June 2025 surveyed 2,232 US public K–12 teachers. Six in ten reported using AI for work during the 2024–2025 school year. Teachers using it at least weekly estimated average savings of 5.9 hours per week.

Reported uses included preparing lessons, creating activities and adapting materials to student needs. The savings were teachers' estimates, not independently timed observations. Gallup's teacher survey.

Reducing preparation time can create room for feedback and individual attention. But it measures a different outcome from whether students learn more. A school could reasonably permit staff assistance while requiring stronger evidence before allowing younger students to rely on a chatbot.

What Microsoft's school AI privacy standard covers

On September 9, the American Federation of Teachers, United Federation of Teachers and Microsoft announced a National AI Safety & Privacy Standard.

The announcement describes protections around data use, human oversight and transparency that US school districts can incorporate into Microsoft customer agreements. Its practical significance is contractual: a district can seek obligations against which a provider's conduct can be assessed. It is not a federal law applying to every vendor. Joint announcement published by Microsoft.

The 31-page agreement covers authenticated generative AI services primarily designed and marketed for education. A school licence does not automatically extend coverage to general-purpose productivity, search or cloud tools.

It prohibits covered data from being used for general model training, with a narrow safety and security exception. Protected data includes identifiable prompts and outputs. The standard also requires meaningful human review and restricts features designed to create emotional dependency. Full school AI privacy agreement.

A district should therefore verify the product and agreement covered by its licence.

Existing privacy rules still apply to educational AI

School AI does not operate outside existing privacy obligations.

The Federal Trade Commission's COPPA guidance explains that schools can, in certain circumstances, authorize collection of children's personal information on parents' behalf. That authority is limited to the educational context: the information must be collected for the school's use and benefit, with no other commercial purpose. Providers must give notice of their practices, and schools must be able to review information and request deletion.

The FTC also tells schools and operators to consider their obligations under the Family Educational Rights and Privacy Act, or FERPA. FTC guidance on COPPA and schools.

For an AI purchase, the resulting questions are practical: what information enters the system, what the provider does with it, how long it remains, and who can authorize a change. A product's "education" label does not answer them.

Federal support for AI does not prove a classroom product works

Federal policy has also encouraged adoption.

In July 2025, the US Department of Education issued guidance on using federal grant funds for responsible AI integration. It said qualifying uses were allowable under existing education programs when they complied with applicable legal requirements. Department of Education guidance announcement.

Permission to fund an activity is different from evidence that a particular deployment improves learning. Schools still need to evaluate the product, its intended use and the students affected.

The National Education Association's 2024 framework supplies a complementary perspective: human relationships, evidence, data protection, equitable access and AI literacy should guide implementation. It explicitly considers accessibility and the needs of students with disabilities. NEA's AI in education report.

What schools and parents should look for next

The most useful evidence will show what changes for students after an AI pilot, including when assistance is removed.

AI Tools Police's assessment is that districts should publish a small set of comparable answers:

• Independent learning: Can students explain and apply the material without the tool?
• Purpose: Does the system provide feedback and practice, or complete the assessed work?
• Scope: Which ages, subjects and tasks were actually evaluated?
• Teacher workload: Do preparation savings remain after checking outputs and managing mistakes?
• Data control: Which contractual terms cover the product and its current features?
• Accessibility: Are beneficial assistive uses protected when broader restrictions are imposed?

These are evaluation questions, not findings that every district already follows them.

A privacy agreement can constrain data use without proving educational benefit. A learning study can show promise without establishing that a particular vendor's contract protects children. Responsible deployment requires both kinds of evidence.

The next stage of the US school AI debate will be decided in those details: whether schools retain control of the tools they buy, and whether students retain the ability to think when the tools are unavailable.

Common questions about AI in US schools

Has the United States banned AI in schools?

No nationwide ban is established by these developments. New York City's policy is local and includes exceptions.

Does the Microsoft privacy standard apply to every classroom app?

No. Coverage depends on the educational product and applicable agreement.

Does faster homework completion mean AI improves learning?

No. Evaluations must distinguish assisted performance from skills demonstrated without AI.

Does the 67% RAND finding prove AI damages critical thinking?

No. It measures respondents' beliefs, not cognitive decline.