Introduction. Educators report that the hardest part of integrating new technology into STEM instruction is knowing how to control it and understand its outputs. This finding recurs across a decade of the authors’ research, from agent-based modeling with NetLogo, through micro:bit programming, through NSA-funded GenCyber camps where most participants embedded cybersecurity in their classrooms, to statewide CS integration (WySLICE) and an NSF-funded RET site (WySTACK). Each intervention introduced a different technology; each exposed the same bottleneck. Generative AI amplifies this: LLMs produce fluent, plausible content requiring verification skills that most STEM educators have never been taught.
Methods. This study synthesizes data across multiple STEM integration PD cohorts. Pre/post self-efficacy measures from earlier WRNN sessions and GenCyber follow-ups provide baseline trajectories. An NSF-funded Noyce workshop series scaffolds a three-stage competency progression, comprising AI-assisted curriculum design, verification of AI-generated STEM content, and applied integration, all using AI tools within authentic disciplinary practice. Instruments capture self-efficacy, verification confidence, and implementation intent. University data (128 course experiences, nine semesters) contextualize the policy environments that educators are preparing students to enter.
Results. Across cohorts and technologies, verification-focused sessions produce the largest self-efficacy gains. Educators completing full progressions describe goals differently than single-session participants, framing AI as something students need to evaluate rather than a tool that saves planning time. University data reinforce this: engineering students in courses with no explicit AI policy report the lowest clarity and highest uncertainty about academic integrity.
Conclusions. Each new technology introduced to educators, NetLogo, micro:bit, GenCyber platforms, and LLMs, required the same scaffolding: guided adoption, structured verification, and disciplinary integration. Self-efficacy gains become most apparent at the verification stage, with novices most challenged by output evaluation. This technology-agnostic competency model, grounded in longitudinal evidence, offers PD designers a replicable framework for whatever classroom technology comes next.
