What Teachers Actually Do With AI vs. What Administrators Say They Should

The Los Angeles Unified School District published its official AI guidance for teachers in September 2025. It runs 47 pages. It includes a decision tree for whether a given AI use case is “approved,” “requires review,” or “prohibited.” It distinguishes between “AI-assisted grading support tools” (approved with caveats) and “AI-generated summative assessment items” (requires curriculum committee review). It defines five categories of AI interaction and maps each to a corresponding set of disclosure requirements.

Almost no teachers in LAUSD have read it. The ones who have read it are the ones who are most afraid of using AI at all, because the document manages to make everything sound like a potential compliance violation.

Meanwhile, across the district, in classrooms that the document’s authors have never visited, teachers are using Claude to write differentiated versions of the same lesson for three different reading levels simultaneously. They’re feeding Claude a week’s worth of student writing samples and asking it to identify the three most common grammatical errors so they can plan mini-lessons. They’re using it to generate practice problems in six difficulty variations. They’re asking it to draft parent communication letters in Spanish when they don’t speak Spanish. None of this is in the approved use cases. All of it is making them better at their jobs.

This gap — between official AI policy and actual classroom practice — is the defining feature of AI adoption in K-12 education in 2027. And it’s not a Los Angeles problem. A survey I reviewed conducted by the RAND Corporation in late 2026, covering 2,400 public school teachers across 12 states, found that 68 percent of teachers use AI tools in their professional practice at least weekly. Among that group, 71 percent reported that their primary AI tools were not officially sanctioned by their district. A third of those teachers said they actively concealed their AI use from administrators, not because they were doing anything wrong, but because they didn’t want to deal with the policy review process.

Think about that for a moment. The teachers most willing to be early adopters of technology that is genuinely improving their practice are the ones most likely to be using it in a gray area, most likely to be doing it quietly, and most likely to be excluded from the official pilot programs that administrators point to when they want to show the school board that they are “responsibly integrating AI.” The responsible integrators are getting credit for caution. The actual innovators are working underground.

The divide maps, imperfectly but clearly, onto the divide between what AI is good at and what administrators think it’s good at. Administrators tend to think about AI in terms of student-facing products: tutoring software, adaptive learning platforms, things that can be procured through a vendor contract and measured in a dashboard. Teachers think about AI in terms of their own professional workflow: the 3 hours they spend on Sunday afternoon planning next week’s lessons, the 2 hours they spend writing report card comments, the constant low-grade stress of trying to communicate with 28 families with different languages and schedules. AI attacks the second category far more effectively than the first.

Jessica Morales teaches seventh-grade English at a Title I school in Phoenix. I spoke with her in March 2027, when her district was in the middle of a formal AI curriculum pilot that she had been excluded from because she hadn’t applied fast enough. She is using Claude daily. She uses it before school to generate discussion questions tailored to the specific reading her class is doing that day, questions that go beyond the textbook’s generic comprehension prompts and actually get at the interesting tensions in the text. She uses it during her prep period to diagnose patterns in student work. She uses it after school to draft communication to parents that she then edits and personalizes.

“The pilot program in my district,” she told me, “is about teaching kids to use AI in their writing process. That’s fine, I guess. But the thing that changed my teaching was using it myself, for my own work.” Her students’ writing scores, measured by the state assessment, improved 11 percentile points year-over-year. She attributes this to better-designed feedback cycles, which she can now do more consistently because the AI is handling the scaffolding work that used to crowd out the feedback. The district’s official pilot program, with its purchased software license and its usage dashboards and its quarterly report to the board, has not yet published outcome data.

This is the pattern everywhere. The teacher innovation is real, is improving practice, and is largely invisible to the institutional measurement apparatus because it’s happening in the workflow layer that schools have never systematically measured. Schools measure student outcomes. They have always measured student outcomes. They do not measure teacher productivity in any granular way, because measuring teacher productivity is politically fraught and methodologically difficult. So the AI-driven improvement in teacher productivity is real and nearly undetectable by the standard institutional metrics.

The administrators who have figured this out are the ones running the most interesting experiments. In Denver Public Schools, curriculum director Alicia Torres did something simple and radical in fall 2025: she held a series of informal “show me how you’re using AI” sessions for teachers, with explicit amnesty for anything that wasn’t officially approved. She learned more in those four sessions than in any vendor demo or pilot report. She then built the district’s actual AI guidance around what teachers were already doing successfully, rather than building it around what vendors were selling.

Denver’s guidance document is four pages. It has one decision tree. The decision tree has three nodes. The teachers in Denver are more likely to have read their district’s AI guidance than teachers in any other large district I know of, because it reads like something written for humans rather than for a compliance audit.

The contrast with most district policy is stark. The typical AI policy document in 2027 is written primarily to protect the district from liability, secondarily to satisfy board members who are worried about AI, and somewhere around fifth or sixth to actually help teachers use AI effectively. The audience is the district’s legal department and its communications staff. The outcome is a document that teachers ignore, which then becomes proof that “teachers aren’t engaging with AI governance,” which generates more policy, which teachers ignore more.

Higher education is not doing better, it’s just failing differently. At the university level, the governance problem is compounded by faculty governance structures that give individual professors significant autonomy over their own syllabi and assessments. This means AI policy at universities tends to be either extremely permissive (each faculty member decides their own rules, chaos ensues) or extremely restrictive (written by administrators who are terrified of academic integrity problems, ignored by faculty who find it unworkable).

The University of Michigan tried to thread this needle in 2025 by creating a “course-level AI transparency” requirement: every course syllabus must state explicitly what AI use is and isn’t permitted, and that statement must be made before the first class meeting. This is a reasonable idea. It gives faculty autonomy while ensuring students know what rules apply. The problem is enforcement: there’s no mechanism to actually ensure the statements are accurate, and no way to determine what happened in any given course if a dispute arises.

The places that are getting it right have accepted a counterintuitive premise: the goal is not to control AI use. The goal is to ensure that AI use produces genuine learning. These are different problems. Controlling AI use requires surveillance, which is expensive, unreliable, and corrosive to the teacher-student relationship. Ensuring genuine learning requires assessment redesign — making the assessment itself impossible to complete without the understanding you want the student to have. This is harder design work. It requires thinking carefully about what you actually want students to know and be able to do, rather than what’s easy to measure at scale. Most institutions have not done that thinking, and they show no signs of starting.

The irony is that the teachers who are using AI most effectively are often doing so in ways that make their own jobs harder in the short run. It is genuinely faster to generate twenty practice problems with AI than to curate twenty practice problems by hand. But the teacher who curates by hand knows, problem by problem, why each one is there and what misconception it addresses. The teacher who uses AI to generate them needs to evaluate them with the same critical eye, or the problems are just noise. Most teachers are doing the evaluation. A minority are accepting the output uncritically. You can tell the difference in the classroom, if you know what to look for.

What administrators mostly see is the output. AI saved time, or AI didn’t save time. AI improved scores, or it didn’t. They are not positioned to see the quality of the teacher’s engagement with the AI, the thoughtfulness with which the output was evaluated, the degree to which the AI amplified good pedagogical judgment versus substituted for it. The measurement gap is doing real damage, because without the ability to distinguish the good AI use from the bad, districts are making policy based on averages that obscure everything interesting.

The teachers who are doing this well know who they are. They talk to each other in the teacher Facebook groups and the Reddit communities that administrators don’t read. They share prompts and strategies and cautionary tales about outputs they caught before students saw them. They are building a practical knowledge base that is more useful than anything any vendor has published. And they’re doing it entirely outside the official structures, which will eventually catch up, probably around the time it’s no longer necessary.