An AI Policy Isn't a Playbook: Educators Need Agency, Not Just New Rules

Synthesis: Schools have spent three years writing rules about generative AI, and this report argues that the rules are the least of it. Drawing on a nationally representative survey of 1,018 US public school teachers and principals plus interviews with 12 educators and 31 young people, the Center for Digital Thriving finds cheating dominating educators' accounts of AI dilemmas while fewer than a quarter of them work under a school policy that governs the tool at all. Its response is not another compliance framework: an Agency-Centered Framework for AI Literacy crossing three literacies with three modes of Learner Agency, three mindshifts, and five classroom "plays" that move educators from referees of AI use to coaches. The authors are explicit that this is a synthesis and resource, not an evaluated intervention.

Key Findings

1. Only 18 percent of surveyed teachers and 24 percent of principals reported that their school has specific policies or guidelines governing generative AI use — weighted, 17.5 percent of teachers and 24.1 percent of principals, counting both "Yes" and "Yes, but they need significant improvement."
2. Cheating dominated the dilemmas educators described: 74 percent of teacher descriptions (171 total) and 69 percent of principal descriptions (85 total) were coded as cheating-related, spanning essays, lab reports, state exams, art projects, and extracurriculars — though this is a coding of what respondents chose to describe, not a prevalence estimate.
3. Detection is where policy meets its limit. The report cites evidence that marketed detection tools are unreliable and describes "AI-DAR," teachers' informal radar for suspecting AI use, alongside students who report dulling their own answers to avoid false accusations.
4. The authors frame the deeper cost as relational rather than academic: in one student's words, "I don't usually go to my teachers. I kind of just go to Chat instead of them." Cheating, they argue, costs the teacher-student relationship as well as the learning.
5. The Agency-Centered Framework for AI Literacy separates three literacies — technical, market, and self — and crosses them with Bandura's three modes of agency (individual, collective, proxy) so a school can diagnose which combination is missing rather than treating literacy as one thing.
6. Three mindshifts are named as the conditions each play needs: from assuming to asking, from tech-centered to teen-centered, and from us-versus-them to us-and-them.
7. Five plays are offered as free, repeatable resources co-designed with educators and young people: Find Your Feet, Scout the Field, Align on the Line, Watch the Replay, and Name the Game — each with its dilemma, its moves, and guidance on when to call it.

What the survey found, and what it could not reach

The survey drew two nationally representative samples from RAND's American Educator Panels: invitations to 1,400 teachers (1,390 eligible) and 998 principals (991 eligible), with 719 teachers and 299 principals retained for a total of 1,018. Fielding ran April 21 to May 27, 2025, online and in English; completion rates were 51.7 percent for teachers against 30.2 percent for principals, and both samples are weighted to their own national populations.

Two design features bound what the numbers can say. The teacher and principal samples are not matched pairs from the same schools, so comparing them describes a national-level pattern rather than any one school. And because the two populations were weighted separately, the results cannot be pooled into claims about "all educators" — which is why every percentage appears by role.

The policy finding is the one that reframes the problem. With roughly four in five teachers reporting no school policy governing generative AI, most educators are exercising individual Learner Agency in the absence of what the report calls proxy agency: the institutional rules and Administrators support that would act on their behalf. That gap is the report's reason for treating policy as necessary but insufficient.

From policy to the framework

The framework starts by refusing to treat AI literacy as a single competency. Technical literacy covers how the tools operate and where they fail; market literacy covers the commercial forces shaping how they are built and distributed; self literacy covers values, Ethics, goals, and knowing where one's own expertise ends. Each is crossed with a mode of agency: individual (you act), collective (we act together), and proxy (others act on your behalf through rules and norms).

The point of the crossing is diagnostic. A school can hold a strong technical curriculum yet leave students without collective practice in negotiating AI norms, or write a policy that supplies proxy agency while leaving teachers no room for judgment. The report's own diagnosis is that the missing cells cluster around collective and proxy agency in a moment changing faster than any rule can track.

Three mindshifts are offered as the conditions the plays need: ask rather than assume what students experience, put the student rather than the technology at the center, and treat AI as an us-and-them problem rather than a tug-of-war. The authors describe the skills the plays require as relational rather than technical.

The five plays

Each play is presented as a dilemma, one or two named moves, what it looks like in practice, and when an educator might "call it." Find Your Feet answers students making AI choices that feel compromised, using a Tech + Values move and a Values Bridge move to generate questions that guide choices; it suits the start of a year or unit. Scout the Field answers students hiding their AI use, with a Making AI Pressures Visible move built on Ethan Mollick's "secret cyborg." Align on the Line answers the gray area where rules are unclear, having students define their own lines and then align as a class before an assignment. Watch the Replay responds to moments that pass too quickly to reflect on, using a Map a Moment move anchored in Dan Siegel's work on reflection. Name the Game builds shared vocabulary through an AI to Z glossary and workshop.

The plays were co-designed with educators and students in workshops through 2025. The report's stance is that the whistle will still be needed, mostly to get students' attention because they are talking.

What this means for practice

• Instructors. Treat suspicion as the weakest tool available. The report's teachers had unreliable detectors, no policy to fall back on, and students who deny use — so the practical move is to open the conversation rather than build a case: surface pressures and values before an assignment, not after an accusation.
• Administrators. Check which cells of the framework your school actually fills. A written policy supplies proxy agency but leaves individual and collective agency untouched, and the survey suggests most teachers do not even have the policy. Pair any rule with time and permission for the classroom conversations that make it workable.
• Curriculum designers. Note that the report separates market literacy from technical literacy: understanding what a tool can do is not the same as understanding why it was built and distributed that way, and the framework treats the second as teachable in its own right.
• Faculty developers. The plays are deliberately lightweight and repeatable rather than a curriculum, and they work with adult groups — faculty or parents — as well as with students, which makes them usable in a staff meeting rather than a training program.
• Researchers. The report's framing is itself a hypothesis: if the constraint is agency rather than knowledge, interventions building collective and proxy agency should outperform tool-focused training. Nothing here tests that.

Limitations

• This is a funder-supported report from an education center with an explicit stance ("critical optimism," agency as the organizing value), not a peer-reviewed study. It synthesizes its own data and the wider literature to argue a position, so its recommendations carry the authority of a well-evidenced brief rather than of independent review.
• The survey was fielded in April and May 2025, roughly seventeen months before publication, and the report flags the date itself when reporting the policy figures. Generative AI capability and school policy both moved in that period, so the 18 and 24 percent policy figures describe that moment rather than the current school year.
• The two samples are nationally representative but are not matched pairs from the same schools, so any comparison between teachers and principals is a national-level pattern rather than a statement about any one school.
• Completion rates diverged sharply by role — 51.7 percent for teachers against 30.2 percent for principals. Weighting addresses some of this, but it does not remove nonresponse bias, and because the two populations were weighted separately the results cannot be pooled into claims about "all educators."
• The qualitative evidence is thin by design: 12 educators completed interviews out of 250 who consented (4 teachers and 8 principals), and the 31 young people were sampled purposively rather than representatively, so their quotes illustrate mechanisms the survey identified rather than how common those experiences are.
• The 74 and 69 percent cheating figures cover only educators who had experienced a dilemma and then described it in an open-ended response, coded by the research team. They are not a prevalence estimate of AI cheating in US schools, and the report does not present them as one. The five plays are likewise co-designed resources, not evaluated interventions: no outcome data is reported on whether calling them changes practice, behavior, or learning.

Citation

Tench, B., Weinstein, E., & James, C., with Starks, A., & Konrath, S. (2026). An AI Policy Isn't a Playbook: Educators Need Agency, Not Just New Rules. Cambridge, MA: Center for Digital Thriving.