The conference theme “From Tools to Teammates: Human-AI Synergy for Augmented Learning” captures something I have been sensing through my work this year: that the research is moving decisively past the question of whether AI can perform educational tasks, and into the far more important question of how humans and AI can learn, teach, and think together, and what is at risk when that partnership is designed without care.

Professor Vincent Aleven from Carnegie Mellon University opened the conference with the question, “How best to harness AI’s great potential to improve education?”. It is a question that sounds straightforward but was helpfully demonstrated to be more complicated through the keynote address. AI tutoring systems have substantial scientific evidence behind them. Research is starting to show that they can help students learn better than many other forms of instruction, and some evidence suggests they can reduce educational inequalities, and yet standardised test scores in the United States remain stagnant, including in K-12 mathematics, where AI-based tutoring is most widely used. How do we make sense of this? His answer, consistent with our findings and recommendations from the Rapid Literature Review, is that deploying an AI system is not the same as creating the conditions under which it can work. The smart classroom is a socio-technical ecosystem. It involves students, teachers, peers, and families, and AI will only reach its potential when we design human-AI interactions thoughtfully across that whole ecosystem.

AI augmenting Teacher tasks

Considering the presentations I attended across the week, the most prominent shift is from automation to augmentation. This is true for student use of AI, but was especially prominent in the area of teacher use of AI. The best current research is no longer asking whether AI can perform educational functions, it is asking how we design AI systems that preserve human judgement rather than eroding it. An example of this was a paper from Singapore’s Ministry of Education, which evaluated an AI-powered Lesson Collaborator Chatbot deployed across seven schools (Huang et al., 2026). Rather than generating lesson plans for teachers, the system was deliberately designed to pose reflective questions at each stage, inviting teachers to exercise their own professional judgement rather than simply accept AI output. 58% of participants used it monthly, it was shown to have demonstrated improvements in the quality of lesson design. It is a model of AI as coach rather than AI content producer. A closely related theme was the risk of teacher deskilling. When AI is too capable, too available, too convenient, teachers (like students) can stop thinking deeply. Systems that require genuine intellectual effort from teachers, while reducing their unnecessary cognitive load, are showing the most promise.

AI tutors

One particularly interesting paper was from Khan Academy reporting on the Khanmigo AI tutoring platform, now used by hundreds of thousands of K-12 students (Udeshi et al., 2027). Rather than seeing if AI can produce correct answers, the worthwhile research question was what configurations of an AI tutor produced the optimal student interactions. They acknowledge that AI can be thought of as a black box, we don’t really follow what occurs in the back end, but we can measure what the impact is on students. They ran more than 40 live experiments with real students over five months. Some of the changes that moved the needle were not the ones they expected. Making the AI’s maths-checking process produce more concise responses, under fifty words rather than verbose explanations, reduced student waiting time by nearly a third. Switching to a more compact input classification model, rather than the largest available, increased cognitive engagement by nearly twelve percent and reduced time to first response by fourteen percent.

But the more revealing finding is what did not behave as predicted. When the team improved the system’s ability to avoid giving away answers directly, something we would want from any good human or AI tutor, students found new ways to coax the answer out. The metric would improve and then gradually drift back as students worked out the new conversational patterns. Cumulatively across those experiments, next-item correctness improved by ten percent and cognitive engagement by fourteen percent.

This matters for schools. Deploying an AI tool is not the same as improving learning. The question is not whether students are using a system. The question is what is actually happening to their thinking as a result. Similarly, it emphasises the importance of measuring impact and interaction, rather than making assumptions about whether any initiative is good for students – something the Barker Institute is perfectly positioned to do to support learning at Barker.

AI-supported teacher professional learning has also emerged as a genuine research priority in its own right, something I was glad to see given our work at the Barker Institute that transcends research and professional learning. A multi-agent tutoring system from Michigan State University demonstrated that AI can replicate many features of quality in-person professional development at scale (Li et al., 2027). A paper from Cameroon explored how WhatsApp-based AI modules could deliver professional learning to teachers in low-resource, mobile-first contexts (Cannanure et al., 2026). Perhaps one of AI’s greatest contribution to schools may not be directly to students, but to the ongoing growth of the teachers who work with them.

The impact of the week

Attending AIED has clarified something I had already begun to sense from the Lakefield College School conversations last week: the Barker Institute’s research agenda, and specifically our Strategic Research Program on AI, Learning, and the Developing Person, is deeply aligned with where the best international scholarship is heading. The questions we are asking about how AI shapes young minds, and about what it means to cultivate distinctly human capabilities in an AI-saturated world, underpin every discussion that is taking place.

I am grateful to the AIED 2026 community and organizing committee. I am grateful that I was able to engage with researchers around the world, all with their own unique perspective on AI in education. I was affirmed that the Barker Institute approaches research with the rigour required for a global contribution, and that we are uniquely placed to be interacting with staff and students in a way that many academics at this conference could only dream of. Too often I was giving feedback to presenters about what this would actually mean for a school, though they were very grateful for the close-to-practice perspective.

In exciting news, AIED 2027 will be held closer to home – Adelaide! So, while sadly I will be unlikely to be able to make it back to Lakefield College School in 2027, I think a return to AIED will be on the agenda for next year. It will be another important opportunity to demonstrate to the international research community the important contribution of the Barker Institute through close-to-practice research.

References

Blanchard, E. G., Chen, G., Chi, M., & Isotani, S. (Eds.). (2027). Artificial intelligence in education: 27th International Conference, AIED 2026, Seoul, South Korea, June 27–July 3, 2026, Proceedings, Parts I & II (Lecture Notes in Artificial Intelligence, Vols. 16581–16582). Springer. https://doi.org/10.1007/978-3-032-29744-0

Cannanure, V. K., et al. (2026). Teacher professional development on WhatsApp and LLMs: Early lessons from Cameroon. In E. G. Blanchard, G. Chen, M. Chi, & S. Isotani (Eds.), Artificial intelligence in education: AIED 2026 companion proceedings (Communications in Computer and Information Science, Vol. 3032, pp. 249–261). Springer. https://doi.org/10.1007/978-3-032-29791-4_18

Huang, Z., Tan, S., & Chew, C. (2026). Teammate or tool: Teacher-AI co-design considerations. In E. G. Blanchard, G. Chen, M. Chi, & S. Isotani (Eds.), Artificial intelligence in education: AIED 2026 companion proceedings (Communications in Computer and Information Science, Vol. 3032, pp. 280–287). Springer. https://doi.org/10.1007/978-3-032-29791-4_21

Li, H., Yang, K., Chu, Y., Han, A., Meng, R., Copur-Gencturk, Y., & Liu, H. (2027). I-VIP: A LLM-driven multi-agent system for professional development of mathematics teachers. In E. G. Blanchard, G. Chen, M. Chi, & S. Isotani (Eds.), Artificial intelligence in education: 27th International Conference, AIED 2026, Seoul, South Korea, June 27–July 3, 2026, Proceedings, Part I (Lecture Notes in Artificial Intelligence, Vol. 16581, pp. 294–308). Springer. https://doi.org/10.1007/978-3-032-29744-0_20

Udeshi, T., Khazenzon, A., Khan, K., Breen, N., Corwin, R. J., DiGiano, C., Weatherholtz, K., & Zaluski, M. (2027). Methodologies for improving the quality of AI tutoring in K-12 education. In E. G. Blanchard, G. Chen, M. Chi, & S. Isotani (Eds.), Artificial intelligence in education: 27th International Conference, AIED 2026, Seoul, South Korea, June 27–July 3, 2026, Proceedings, Part II (Lecture Notes in Artificial Intelligence, Vol. 16582, pp. 201–214). Springer. https://doi.org/10.1007/978-3-032-29755-6_14

(To read the post about Week 2 of the trip, at the Lakefield College School AI Academy, Canada, click here).