Pre-service teachers' agency during their interactions with generative AI while designing for learning - a process view

Synthesis: Krushinskaia, Elen, and Raes (2026) use process-level indicators of Learner Agency to examine 78 pre-service teachers randomly assigned to design learning materials with either a custom expert GenAI bot scaffolded on an instructional-design model or basic ChatGPT. Two of three agency indicators (length of interaction and self-generated prompts/ownership) suggested higher agency in the expert-bot condition, but a third indicator — collaborative Problem Solving — revealed reduced agency in both conditions, as participants often simply agreed with the bot's suggestions rather than meaningfully critiquing them.

Key Findings

• Using process-level indicators, pre-service teachers in the expert-bot condition interacted longer and generated more self-generated prompts, suggesting greater agentic engagement on those dimensions.
• A third agency indicator, collaborative problem-solving, revealed reduced agency in both conditions: participants frequently agreed with AI output instead of providing meaningful Feedback.
• Both the material constraints of the expert bot and the open-ended structure of ChatGPT impose distinct constraints on teacher agency.
• Prior knowledge about GenAI or instructional design moderates how much agency a teacher can exercise in either context.

What this means for practice

• Teacher educators. Set an explicit expectation that pre-service teachers critique and push back on AI suggestions; participants in both conditions tended to agree with the bot rather than give meaningful Feedback, the lowest-agency behavior observed.
• Teacher educators. Teach the affordances and constraints of specific tools — a custom design bot versus unrestricted ChatGPT — alongside tool use, since each imposed distinct limits on agency.
• Designers. Do not assume extra Scaffolding preserves agency; the expert bot raised interaction length and self-generated prompts but coincided with reduced collaborative problem-solving.
• Researchers. Use process-level indicators — interaction length, ownership, and collaborative problem-solving behaviors — rather than self-report, and state each indicator's validity caveats.

Limitations

• 78 pre-service teachers recruited from a single university were randomly assigned to the two conditions, which the authors note may limit sample diversity.
• The study captures a one-time interaction with a GenAI bot; repeated or longitudinal measures were not collected.
• The agency indicators are proxies with known weaknesses: interaction length may reflect verbosity rather than engagement, ownership was measured only at task-execution level, and the proactive/reactive classification of collaborative problem-solving behavior was the authors' own extension.
• Some utterances were difficult to categorize, yielding substantial but not perfect inter-rater reliability (κ = 0.70).

Citation

Krushinskaia, K., Elen, J., & Raes, A. (2026). Pre-service teachers' agency during their interactions with generative AI while designing for learning – a process view on Intelligent-TPACK. Computers and Education Open, 10, 100325.