Abstract:

Most trainee teachers receive detailed feedback on their teaching only rarely. Weekly observed lessons are the norm in initial teacher training in England, but the gap between formal observations is largely unexamined. A trainee may teach thirty lessons between observations; only one will be seen, and mentor time for post-lesson discussion is, in most schools, limited. This paper describes a tool designed to address that gap, and examines the pedagogical and ethical questions it raises.

The tool takes audio from any lesson or lecture, transcribes it using OpenAI's Whisper, and submits the resulting transcript to a large language model (Anthropic's Claude Haiku) prompted by a structured knowledge base drawn from the ITT and Early Career Framework (ITTECF) and organised against the Teachers' Standards. The feedback identifies lesson phases, notes two or three strengths with specific ITTECF references, names areas for development with concrete suggestions, and closes with a single priority next step. Privacy is central to the design: audio is discarded immediately after transcription, no data is stored, no login is required, and trainees control what reaches the model. The feedback is caveated as AI-generated and unreviewed, and is presented as the starting point for a mentor conversation rather than a substitute for one.

Demszky et al. (2023) showed in a randomised controlled trial that AI-generated feedback on classroom audio increased teachers' use of high-quality questioning by 20%. Jacobs et al. (2025) found that AI tools could surface patterns in classroom discourse that post-lesson conversations rarely reached. Where studies have compared AI and human feedback directly (Nygren et al., 2025), AI proved more consistent and comprehensive in scope, while human feedback remained better attuned to the relational and contextual dimensions of teaching. The consensus is that AI and human mentoring are complementary, with AI most productively positioned as a prompt for professional conversation rather than a replacement for it.

Applied to a recorded lecture, the tool generates structured, ITTECF-referenced feedback covering strengths, developmental suggestions, and a single priority next step. A key limitation of the approach is also evident: a think-pair-share activity whose pair discussion was inaudible became, in the transcript, simply absent. The system acknowledged it could not evaluate what it could not hear. The tool works from language alone, and teaching is not only language.

Three questions structure the analysis. Does low-stakes, formative AI feedback support professional formation, or does it risk outsourcing the cognitive work of reflection that is itself part of learning to teach? What does it mean for the mentor relationship if trainees arrive at post-lesson conversations already holding detailed AI feedback? And what would constitute adequate evidence of effectiveness, given the privacy constraints the tool is designed to respect?

This is proof-of-concept work. The tool is in use with PGCE trainees at Roehampton; robust evidence of effectiveness beyond positive qualitative feedback is not yet available. Qualitative responses have been consistent: trainees describe the feedback as more detailed than mentor feedback and specific in useful ways. The work contributes to an emerging conversation about where AI can and cannot usefully serve as a pair of ears in teacher education.

Keywords:

Coaching, pre-service teachers, inservice teacher training, generative AI.