Table of Contents
Teacher AI Lesson Planning Is 60-99% of Use: What It Means
Teacher AI lesson planning is 60-99% of usage in a 12-district study. What that means for your kid's assignments, and the small outcome data behind it.
The most-quoted number about classroom AI in 2026 is about teachers, not students. Across 12 Washington State school districts, teacher AI lesson planning accounted for 60 to 99 percent of all platform use. Not grading. Not chatting with students. Planning. That single statistic reframes the whole parent conversation: the AI in your child’s school life is mostly upstream of them, shaping what the assignment looks like before they ever see it.
Key Takeaways
- A University of Washington team tracked Colleague AI across 12 Washington districts from September 1 to December 31, 2025, and found lesson planning made up “60-99% of usage in every participating district.”
- Active users averaged 12 sessions, 3.7 hours total, about 18 minutes per session, and 143 messages over four months.
- Teacher counts grew from 1,438 in September 2025 to 2,891 by December 31 and 4,644 by March 31, 2026.
- The outcome analysis found reading gains of about 0.025 standard deviations per three hours of teacher AI use across 12,005 matched students, with math “not statistically distinguishable from zero.”
- Teachers serving more English learners used differentiation and student-profile features more; teachers with more special-education students used lesson-delivery features more.
What the study actually measured
The report is Generative AI in K-12 Classrooms: A Midyear Implementation Report, from AmplifyLearn.AI at the University of Washington in collaboration with Colleague AI, posted to arXiv with authors including Lief Esbenshade, Alex Liu, Michael Xiao, Zewei Tian, and Min Sun. It combines platform telemetry with administrative records across 12 districts ranging from a few thousand students to about thirty thousand, spanning rural, suburban, and urban settings.
That design is unusual and valuable. Most claims about teacher AI use come from surveys, which capture what teachers report. This one captures what they clicked. By the end of December, teachers had exchanged more than 412,304 messages with the platform.
The headline distribution: lesson planning was 60 to 99 percent of usage in every single district. Within teacher requests, the most common pedagogical topics were “Planning and Explicit Teaching,” followed by Assessment, Critical Thinking and Inquiry, and Student Profiles. The authors note that “non-educational conversations were rare,” which is a small but real finding given the assumption that teachers would use these tools for personal errands.
Adoption spread fast. From 1,438 teachers in September 2025 to 2,891 by December 31 (nearly 30% of teachers in participating districts) to 4,644 by March 31, 2026. There was “not a clear relationship between district size and participation rate.”
Use type by share: what the hours went to
| Use type | Share of usage | What it changes for your kid | Confidence |
|---|---|---|---|
| Lesson planning and explicit teaching | 60–99% across districts | The structure and sequence of the lesson | High, from platform telemetry |
| Assessment creation | Second most common topic | Quiz frequency and question quality | High |
| Critical thinking and inquiry prompts | Third | Discussion questions, open tasks | High |
| Student profiles, differentiation, accessibility | Referenced more by teachers with more English learners | Whether your kid gets a version at their level | Moderate; associational |
| Lesson delivery (generate image, generate interactive) | Used more by teachers with more special-education students | In-class materials and visuals | Moderate; associational |
| Non-educational use | ”Rare” | Nothing | High |
The two associational rows are the most interesting for families. Teachers serving higher shares of English learners referenced “Student Profiles, Differentiation & Accessibility” more frequently. Teachers with more special-education students were “more likely to use Lesson Delivery features such as Generate Image and Generate Interactive.” If your child is in either group, the AI is being pointed at their specific need more often, which is a plausible equity upside and not yet a proven one.
The outcome number, stated honestly
Here is the part most coverage either inflates or ignores.
For reading, the study found that “3 hours of teacher AI use was associated with a statistically significant increase of approximately 0.025 standard deviations,” across 12,005 students with matched beginning- and middle-of-year test scores. For mathematics, the association was “not statistically distinguishable from zero.”
The authors’ own framing: “preliminary correlation, and the small effect size is an encouraging signal.” They add that “teacher adoption was voluntary, and we may be observing a selection effect,” that only “several of the twelve participating districts” provided mid-year data, and that “some districts paused their AI implementation plans.” They also caution that “findings that link teachers’ use of Colleague AI to student characteristics should be interpreted as preliminary signals.”
What 0.025 SD means in practice: it is small. For scale, RAND’s 2017 Informing Progress evaluation of personalized learning across 40 schools for the Gates Foundation found treatment effects of about 0.09 in mathematics and 0.07 in reading, translating to roughly three percentile points, with only the math estimate statistically significant. So this effect is smaller than a personalized-learning effect that itself was called modest.
Small and real beats large and imaginary. But a parent should not expect a visible difference in their child’s reading level because their teacher used a planning tool.
Why teacher AI lesson planning is where the tools landed
There is a reasonable explanation for the 60 to 99 percent figure, and it is not that teachers are lazy.
Planning is the task with the highest ratio of writing to judgment. A teacher knows what the lesson needs; producing the artifact (the warm-up, three differentiated versions, the exit ticket) is typing. Grading is the reverse: low writing, high judgment, and it is also the task most restricted by policy. DC’s model AI policy, released September 1, 2026, places reviewing and grading student work in its “use with caution” tier requiring human final decisions, and bars AI from discipline, IEP eligibility, and teacher evaluation.
The tools followed. Anthropic’s Claude for Teachers, launched July 14, 2026, leads with a Learning Commons connector carrying academic standards for all 50 states plus connections to OpenSciEd and Illustrative Mathematics. Microsoft’s June 2026 release included Unit Plans in Teach for “fully developed, standards-aligned plans in minutes.” OpenAI’s teacher-facing program emphasizes lesson planning and classroom resources. Every major vendor aimed at the same task, because that is where the measurable time is.
Gallup’s data explains why: teachers using AI weekly save about 5.9 hours a week, and the top tasks were preparing to teach (37%), making worksheets (33%), and modifying materials for student needs (28%).
What to do at home
Look at the assignment, not the tool
Since planning is where the AI lives, the assignment is where you would see it. Ask yourself: does this worksheet target the standard, or does it look generic? Are there versions at different levels? Does the sequence build? Those are lesson-design questions, and they are now partly AI-assisted questions.
Check the standard alignment yourself
Most states publish their learning standards online. If an assignment seems off-target, look up the standard for that grade and unit. Tools like the Learning Commons connector exist precisely because alignment used to be hard; a misaligned assignment in 2026 is worth a polite question.
Ask whether differentiation actually shipped
The study found teachers with more English learners and more special-education students used differentiation features more. If your child needs a different level and the materials never vary, the feature exists but is not reaching them. Our guide on interleaving subjects versus blocking during homework covers what good sequencing looks like.
Watch for content errors, and report them once
AI-generated materials can contain plausible errors. If your child brings home a worksheet with a wrong date, a wrong formula, or a science claim that contradicts the textbook, tell the teacher once, factually, with the specific item. Teachers reviewing AI output want to know where it failed.
What not to do
Do not assume AI planning means less teacher effort per lesson; the study’s active users averaged 18 minutes per session, which is a working session, not a shortcut. Do not treat 0.025 SD as a reason to demand the tool or to ban it. And do not conclude your kid is interacting with AI because their teacher is; in this data, students were not the users. Our piece on the AI tutor in your kid’s classroom covers the student-facing side separately.
What to Watch For Over the Next 3 Months
- Week 4: You have looked at three assignments and can say whether they target the standard and whether they vary by level.
- Month 2 red flags: A factual error in generated material that nobody caught; identical assignments across two very different classes; feedback that reads like it was written about a generic student.
- Month 3 self-check: Ask the teacher what planning takes less time now and what they do with that time. A specific answer means the reinvestment is real. A vague one means the hours went to the teacher’s evening, which is not nothing.
Frequently Asked Questions
Does 60 to 99 percent lesson planning mean teachers aren’t using AI for grading?
In this dataset, mostly yes. Lesson planning dominated usage in every one of the 12 districts studied. That is partly preference and partly policy: guidance like DC’s model policy puts grading in a caution tier requiring human final decisions.
Is my child interacting with AI in this scenario?
Not through this tool. Colleague AI, like Claude for Teachers and ChatGPT for Teachers, is teacher-facing. Student-facing tools are a separate question, and whether your child has one depends on your district’s settings.
How big is the 0.025 standard deviation effect, really?
Small. For comparison, RAND’s 2017 study of personalized learning across 40 schools found about 0.09 in math and 0.07 in reading, roughly three percentile points, and that was described as modest. The authors of the Washington study call theirs “preliminary correlation” from a voluntary-adoption sample.
Does AI planning help kids with IEPs or English learners more?
The study found teachers serving more English learners referenced differentiation and student-profile features more, and teachers with more special-education students used lesson-delivery features like image and interactive generation more. That is an association, flagged by the authors as a preliminary signal, not a demonstrated benefit.
Should I be concerned that an AI wrote my kid’s lesson?
Concerned is too strong; curious is right. A lesson generated against your state’s standards and reviewed by a teacher who knows your child is not obviously worse than a photocopied packet from 2014. An unreviewed lesson is worse. Ask about the review step, which is the part that matters.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Esbenshade, L., Liu, A., Xiao, M., Tian, Z., Sun, M., Zhang, Z., Han, T., Lapicus, Y., & He, K. (2026). Generative AI in K-12 Classrooms: A Midyear Implementation Report. AmplifyLearn.AI, University of Washington. https://arxiv.org/abs/2605.16277
- Gallup / Walton Family Foundation. (2025, June 24). “Three in 10 Teachers Use AI Weekly, Saving Six Weeks a Year.” https://news.gallup.com/poll/691967/three-teachers-weekly-saving-six-weeks-year.aspx
- Pane, J. F., Steiner, E. D., Baird, M. D., Hamilton, L. S., & Pane, J. D. (2017). Informing Progress: Insights on Personalized Learning Implementation and Effects. RAND Corporation. https://www.rand.org/pubs/research_reports/RR2042.html
- Anthropic. (2026, July 14). “Introducing Claude for Teachers.” https://www.anthropic.com/news/claude-for-teachers
- Microsoft. (2026, June 24). “Microsoft’s new AI in Education Report highlights widespread adoption and increasing demand for support.” https://news.microsoft.com/source/2026/06/24/microsofts-new-ai-in-education-report-highlights-widespread-adoption-and-increasing-demand-for-support/
- DC Office of the State Superintendent of Education. (2026, September 1). “OSSE Releases AI Model Policy to Guide Responsible Staff Use in Schools.” https://osse.dc.gov/release/osse-releases-ai-model-policy-guide-responsible-staff-use-schools
- EdTech Innovation Hub. (2026, September 1). “OpenAI adds 55 systems to ChatGPT for Teachers.” https://www.edtechinnovationhub.com/news/openai-expands-chatgpt-for-teachers-to-55-more-us-school-systems
- OECD. (2026, September 8). “PISA 2025: Students’ reading and mathematics performance declined sharply across the OECD.” https://www.oecd.org/en/about/news/press-releases/2026/09/pisa-2025-students-reading-and-mathematics-performance-declined-sharply-across-the-oecd.html
