Pre-service teachers’ agency during their interactions with generative AI while designing for learning – a process view on Intelligent-TPACK

• Pre-service teachers designed a lesson using an expert GenAI bot and ChatGPT. • Three indicators were used to explore their agency during interactions with GenAI. • Length of interaction indicated higher agency in the expert bot condition. • Ownership indicated higher agency in the expert bot condition. • Collaborative problem-solving behaviors showed mixed results and reduced agency. Agency is considered a core societal value, yet it remains unknown how human agency is affected in human-AI interactions. This study is among the first to use process-level indicators to examine teacher agency during interactions with generative AI (GenAI) in the context of instructional design. While instructional design is regarded as a key teacher role, research shows that teachers often avoid this cognitively demanding task. Today, GenAI offers new opportunities to support teachers in the design process. One approach involves using expert bots customized with one instructional design model, which scaffold the design process and offer ready-made components (e.g., learning goals). However, it is unclear how interactions with an expert bot versus basic ChatGPT influence teacher agency. In this research, we analyzed process data and qualitative feedback from an experiment involving 78 pre-service teachers randomly assigned to one of two conditions: (1) the expert bot condition, where participants interacted with a custom GenAI bot; and (2) the ChatGPT condition, where participants could request any type of support. Three indicators of agency were examined: (1) length of interaction, (2) ownership, and (3) collaborative problem-solving behaviors. The first two indicators suggested higher agency in the expert bot condition, as participants interacted longer and used significantly more self-generated prompts. However, the third indicator revealed reduced agency in both conditions. Instead of providing meaningful feedback on the bot’s suggestions, participants in the expert bot condition often agreed with its output. This suggests that the material contexts of ChatGPT and the expert bot both have unique constraints on teachers’ agency.