AI Integration as Instructional Design: Lessons from a Cross-Institutional Faculty Collaboratory in Teacher Preparation

Synthesis: The AmplifyGAIN Center convened 15 teacher educators from 13 U.S. institutions for a semester-long EPP Faculty Collaboratory, and 12 of them completed the design-implementation cycle that this white paper synthesizes. Its central claim is a reframing: integrating Generative AI into teacher preparation is an instructional design problem, not a tool adoption problem. Five themes carry the evidence, and each is a design lesson rather than a recommendation to adopt or prohibit. AI works when it supports professional judgment instead of supplying answers; disciplinary reasoning has to come first, because AI that arrives too early bypasses the analysis a methods course exists to build; critique of AI output must be written into the assignment, since candidates did not question outputs on their own; what counts as responsible use depends on the course; and integration costs faculty more design work, not less. The paper opens with a classroom episode it names the balloon popping effect, where candidates devalued feedback they had already judged useful once they learned AI produced it.

Key Findings

1. Adoption has outrun institutional guidance. In the 2026 survey wave cited here, 58% of U.S. math and science teachers reported using generative AI in teaching and 31% weekly, up from 50% and 21% in 2025, while only 11% said their district had formal guidelines and 58% had received no formal AI training.
2. Even the best-supported teachers are on their own. Among frequent users, 32% reported no district AI guidelines, only 18% reported established policies, and 61% had received no district-provided training.
3. Teacher concerns focus on students, and they are rising. The 2026 survey records creativity loss at 67%, over-reliance at 66% and academic integrity at 60%, each higher than the year before.
4. Position AI as support for professional judgment. Across the twelve implementations, faculty used it as a thinking partner, a critique generator or a rehearsal tool, with candidates keeping responsibility for evaluating, adapting and justifying decisions.
5. Disciplinary analysis has to come before AI, not after. Faculty who protected independent analysis found candidates could evaluate AI output more critically; when AI entered too early it bypassed the reasoning the course was designed to develop.
6. Critique must be assigned, because it does not happen spontaneously. Candidates did not audit AI outputs on their own, so requirements to compare, revise and justify generated content had to be built into the task itself.
7. Disclosure changes how output is valued. In the opening episode, candidates had judged feedback substantively useful and devalued it once AI authorship was revealed, with the content unchanged.

What the twelve implementations show

The collaboratory deliberately did not standardize practice: each faculty member designed an implementation for their own course, discipline and candidate population, spanning mathematics education, educational technology, assessment, multilingual education, clinical practice, elementary education and AI literacy, and institutions from R1 universities to regional comprehensives, an HBCU and urban public universities. That variation is what makes the convergence credible. Every implementation treated AI as part of the instructional design, from where in the sequence it appears to what candidates must produce for the instructor to see their reasoning. It is also why the paper rejects a single model: context decides what responsible use looks like, and programs need shared principles with course-specific boundaries rather than one policy applied across a curriculum. The faculty reported the work as design work, with clearer directions, modeled prompting and inspection, explicit boundaries and assessments that capture candidate thinking all needed after the first iteration.

Where the design work actually went

Three moves recur across the cases. The first is bounding AI use to a single activity with a clear instructional purpose, so the tool serves a step in the candidate's reasoning rather than standing in for it. The second is making the process visible, which means asking for prompts, inspections, revisions and justifications rather than only the final artifact, because a polished AI-assisted lesson plan tells an instructor nothing about whether the candidate can explain why it fits the learners and the standards. The third is treating AI literacy as an integrated dimension of professional preparation rather than a separate technology skill, which the paper contrasts with guidance centered on plagiarism prevention. Against those moves sit the exclusions the faculty converged on: no identifying student data, no AI-generated content sent into classrooms without review, and no AI grading or feedback standing in for the instructor's Feedback and judgment.

What this means for practice

• Program leaders. Start with a program-level vision and fund designated time for assignment redesign. Rigid, curriculum-wide prohibitions are named as the failure mode because they block course-specific design.
• Faculty. Sequence the disciplinary work first and put critique into the task: require candidates to compare, revise and justify AI output, then grade the reasoning, not the artifact.
• Accreditation and policy stakeholders. Treat the first year of integration as a pilot and pair any standard with faculty practice evidence, since the field still lacks consensus on what AI literacy means for educators specifically.
• R&D partners. Prioritize short, interpretable outputs that candidates can inspect and support the generate-inspect-critique-revise workflow rather than longer autonomous generation.

Limitations

• This is a white paper synthesizing twelve faculty-designed course implementations and their documentation, not a controlled study: no candidate learning outcomes were measured across the cases.
• The survey figures come from cited national surveys of U.S. math and science teachers, so they describe that population and not teacher preparation across all disciplines.
• Three of the fifteen convening faculty did not complete the design-implementation cycle, and the collaboratory's lessons reflect faculty who volunteered for the work, which is likely a more motivated group than a program average.

Connected Concepts

• AI Literacy
• Teacher AI Competency
• Teaching
• Academic Integrity
• Human AI Collaboration
• Scaffolding
• Workplace Learning
• Educational Development
• Generative AI
• Large Language Models (LLMs)
• Feedback
• Equity
• Educational AI Policy
• Change Management
• Professional Development

Connected Articles

• Preparing Pre-Service Teachers for Responsible Generative AI Use: Curriculum Implications for Ethics, Privacy, and AI Literacy — curriculum implications for preparing pre-service teachers to use generative AI responsibly
• Human-centered AI for teacher educators: Designing professional learning for critical AI literacy — professional learning design for critical AI literacy among teacher educators
• Pioneering Teacher Educators Navigating AI Integration in Pre-Service Teacher Preparation: Strategies and Challenges — how teacher educators navigate AI integration in pre-service preparation
• Writing the Rules for Generative Machines: Tensions and Entanglements in Preservice Teachers' Classroom AI Policies — tensions when pre-service teachers write classroom AI policy

Citation

Liu, A., Zeller, A., Traynor, A., Sarmiento-Quezada, B., Walkington, C., Holtz, E., Pai, G., Lee, H.-J., Sun, L., Bondurant, L., Burnett, S., Casey, S., Girtz, S., & Sun, M. (2026). AI integration as instructional design: Lessons from a cross-institutional faculty collaboratory in teacher preparation. AmplifyGAIN R&D Center, University of Washington College of Education. AmplifyLearn.AI Center Series.