An AI-supported professional development course helped maths teachers teach ratios and proportional relationships, and their pupils subsequently performed better on an assessment of that material. The trial followed teachers into the 2021–22 school year.[1]
The teacher does the learning
Researchers including USC’s Yasemin Copur-Gencturk designed the programme around mathematical understanding and knowledge of how pupils learn. Teachers solved problems, interpreted student work and examined teaching. An intelligent tutoring system responded to their written explanations with feedback on concepts they still needed to address.[3]
How the professional development works
| Teacher activity | Purpose of the feedback |
|---|---|
| Explain the mathematical reasoning behind an answer. | Identify which expected concepts are present and prompt for missing ideas.Source 3 |
| Interpret a pupil’s work or error. | Develop knowledge of common misconceptions and ways to respond.Source 3 |
| Examine an instructional example. | Connect the learning goal, task and explanation to the understanding pupils need.Source 3 |
Experts prepared expected answers, likely misconceptions and hints. The system used them to guide exchanges and offer fuller explanations when needed. Teachers could receive feedback while working independently through an idea.[3]
The pupil result was specific and measurable
The pupil assessment study
| Measure | Value | Context |
|---|---|---|
| Allocation and follow-up | 70 teachers randomised in summer 2021; 53 teachers and 1,727 pupils included in the 2021–22 analysis. | |
| Learning assessed | Grade 6 and 7 ratios and proportional relationships, using a researcher-assembled assessment. | |
| Adjusted result | A positive effect of 0.18 standard deviations, expressed relative to control pupils’ baseline variation. | |
| Teacher commitment | The training group averaged 11.1 hours on the course. |
The US What Works Clearinghouse recorded the positive assessment finding in its review, while rating the study as meeting its standards with reservations. Its broader numbers-and-operations domain judgement was uncertain effects. That combination supports a carefully bounded conclusion about the measured content, without treating one assessment as proof of a general improvement across mathematics.[5]
What the classroom work showed
A companion paper examined classroom practice from the same recruitment and randomisation, so it provides another view of the programme rather than an independent replication. It analysed material from 52 teachers, with 51 in the mathematical-coherence analysis. Teachers submitted lesson tasks, examples of pupils’ work and explanations of instructional decisions; raters did not know their study group.[2]
The trained teachers selected tasks with greater potential to demand mathematical thinking and provided more coherent opportunities to develop conceptual understanding. However, there was no statistically significant difference in the demand maintained while teaching the tasks. The distinction is useful: choosing a rich problem and sustaining pupils’ reasoning through it are separate achievements. The classroom measures were based on submitted artefacts, and the analyses used directional, one-sided significance tests.[2]
The next version changes the conversation
A July 2026 design paper describes a later system using language-model agents within a curriculum prepared by mathematics educators. One agent filters relevance, another checks expected concepts, and a responder adapts the next authored hint. After repeated unsuccessful hints, a facilitator provides a more direct explanation. Teachers’ feedback on these exchanges is reviewed by content experts.[4]
Those newer design features should not be credited with the earlier pupil result: the trial used the preceding intelligent-tutoring approach. The continuity is the emphasis on substantial subject knowledge and carefully planned feedback. AI makes it possible to vary the conversation around that material. The educational opportunity is a course that teachers can complete in their own time while working actively through the ideas they will later help a class understand.[4]
Sources & further reading
- The impact of an interactive, personalized computer-based teacher professional development program on student performance: A randomized controlled trial Computers & Education · 23 Nov 2023 · Accessed 11 Oct 2026
- Improving Teaching at Scale: Can AI Be Incorporated Into Professional Development to Create Interactive, Personalized Learning for Teachers? American Educational Research Journal · 16 May 2024 · Accessed 11 Oct 2026
- A promising approach to scaling up professional development: intelligent, interactive, virtual professional development with just-in-time feedback Journal of Mathematics Teacher Education · 15 Dec 2023 · Accessed 11 Oct 2026
- Designing professional development in mathematics teacher education with AI as a facilitator: a roadmap grounded in high-quality professional development features Journal of Mathematics Teacher Education · 29 Jul 2026 · Accessed 11 Oct 2026
- Review of the impact of an interactive, personalized computer-based teacher professional development program on student performance Institute of Education Sciences — What Works Clearinghouse · Accessed 11 Oct 2026