From the K–12 edition of September 19, 2026
UW/Colleague AI midyear report: WA teacher AI use correlated with +0.025 SD reading gains; lesson planning dominates
University of Washington AmplifyLearn.AI and Colleague AI (IES/NSF-supported) · K-12 multi-district teacher AI implementation research · Washington State, United States (12 school districts)
- Publisher
- Generative AI in K-12 Classrooms: A Midyear Implementation Report (arXiv PDF)
- Original publication
- March 2026 midyear report; arXiv posting April 8, 2026 (arXiv:2605.16277v1)
- Source retrieved
- 2026-09-20
- Event date
- 2025-12-31
What happened
Across 12 Washington districts, teacher use of Colleague AI grew from 1,438 teachers in September 2025 to 2,891 by end-December (412,304 messages), reaching 4,644 teachers by March 31, 2026 (+61%). Active users averaged ~12 sessions and 3.7 hours; Claire AI lesson-planning features comprised 60–99% of usage in every district. Teachers serving more ELL/SPED students referenced differentiation and accessibility features more often. Among 12,005 students with matched BOY–MOY scores, roughly three hours of teacher AI use associated with a statistically significant +0.025 SD reading interim gain; math association was not distinguishable from zero. Authors label the student-outcome link a preliminary correlation under voluntary adoption.
Why it matters
Shows where teacher-facing GenAI actually concentrates (lesson planning) and how modest correlational ELA signals—not math—appear in multi-district U.S. K-12 settings, informing SEA/LEA expectations for staff AI ROI and evaluation design.
Evidence and measured results
Opened arXiv PDF end-to-end (executive summary key findings, adoption tables, feature-mix figures, BOY–MOY correlation section). Cross-checked HTML mirror for continuous wording. Confirmed authors’ caution that only a subset of districts provided mid-year admin data and that findings linking use to student characteristics are preliminary signals.
Limitations and uncertainty
Correlational, voluntary adoption, first four months; vendor coauthor (Colleague AI); subset of districts supplied admin data; not an RCT; effect size small; math null. Why now: first unarchived multi-district WA teacher-AI↔student-score package contemporaneous with binding state AI rules and homework-learning-penalty evidence.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-20; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Districts lack credible, local evidence that teacher GenAI use connects to student learning beyond anecdotal time savings.
- Stakeholders
- Superintendents
- Chief academic officers
- CTOs
- Teacher leaders
- Board research liaisons
- Discovery questions
- Can you link teacher AI dosage to interim ELA/math without consumer tools?
- What share of AI use is lesson planning vs student-facing today?
- Will you accept correlational midyear signals or require an RCT?
- How will ELL/SPED differentiation prompts be governed?
- Value hypothesis
- Teacher AI that concentrates on lesson planning can show modest correlational ELA signals when dosage and interim scores are measurable—without promising math gains.
- Potential engagement
- 8–12 week midyear teacher-AI implementation study: telemetry, PD, and BOY–MOY association analysis with published caveats.
- Unsupported claims
- Causal student learning gains from Colleague AI
- Guaranteed math achievement improvement
- That +0.025 SD generalizes beyond these WA districts
Pre-sales engineering
Role takeaway
- Fit
- Fit for districts with interim assessment data and SSO-managed teacher AI; poor fit where AI use is unmanaged consumer chat.
- Architecture and integration
- Teacher IdP → enterprise AI workspace → feature telemetry warehouse → roster join for interim scores; student-profile fields optional and permissioned.
- Prerequisites
- Teacher SSO and roster data sharing agreement
- Interim assessment extracts (BOY/MOY)
- Feature-level usage logging
- Privacy review for student-profile prompts
- Constraints
- Voluntary adoption selection bias
- Vendor coauthorship conflict management
- Incomplete admin data from some districts
- Security
- Keep learner attributes and chats in enterprise tenancy; minimize student PII in teacher prompts; audit differentiation feature access.
- Proof of value
- Reproduce adoption counts and feature-mix shares
- Replicate +0.025 SD ELA association with district-held data
- Show math association remains null under same spec
- Document limitations slide matching authors’ caveats
Delivery
Role takeaway
- Work
- Onboard teacher cohort, enable planning-first workflows, collect dosage telemetry, run interim score joins, and brief leadership on correlational results.
- Dependencies
- Assessment office
- IT/IdP
- Union/teacher advisory
- Vendor DPA
- Ownership
- CAO owns instructional interpretation; CTO owns telemetry and privacy; principals own classroom practice.
- Skills and adoption
- PD focused on lesson-planning prompts and ELL/SPED differentiation; discourage treating AI outputs as student grades.
- Governance checkpoints
- DPA and telemetry plan approved
- Pre-registered analysis note
- Midyear public limitations briefing
- Go/no-go on expansion beyond planning
- Acceptance criteria
- Feature-mix report shows planning share by school
- ELA/math association table with confidence intervals
- Written correlational caveat in board materials
- No student-facing GenAI expansion solely on this signal
- Risks
- Selection bias inflates apparent gains
- Vendor marketing overclaims causality
- Student-profile leakage via differentiation prompts
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
District stacks should assume teacher copilots will be lesson-planning heavy; instrument feature telemetry (planning vs delivery vs differentiation) and link roster/SIS identifiers carefully before claiming student-outcome causal effects.
Governance
Who approves, reviews and stays accountable for outcomes?
Voluntary multi-district pilots need pre-registered outcome metrics and clear disclosure of vendor coauthorship; treat +0.025 SD reading association as a signal to evaluate, not a procurement warranty.
Security and privacy
What data, permissions and controls need testing?
Student-profile and differentiation prompts imply sensitive learner attributes flow into teacher AI sessions—require enterprise DPA, minimization, and audit of profile fields before scaling.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Teachers serving more ELL/SPED students already lean on differentiation features—pair AI enablement with accessibility PD and workload-aware adoption supports.
Procurement
What should contracts, pricing and exit terms secure?
Require usage telemetry, interim-outcome evaluation hooks, and disclosure of vendor research roles in RFPs for teacher copilots.
Operating model
Which teams own the service once it runs?
Treat teacher AI as a planning workflow with midyear evidence reviews, not a perpetual pilot without outcome joins.
Publication history
- 2026-09-19K–12 · Issue 144 resources
Stable resource ID: uw-colleague-ai-washington-midyear-2026