3-2-1 on AI and learning

1. The Teaching Method That Can’t Fail by Barbara Oakley. Aside from the constructivism debate, Barbara highlights a deeper issue in education: falsifiability. Many claims about learning are framed in ways that make them impossible to disprove. In that sense, they are not even wrong.

2. Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schools by a team from Cambridge University Press & Assessment and Microsoft Research. They performed a randomised experiment with 405 students aged 14-15 in England, found that note-taking alone and note-taking combined with LLM use both significantly improved comprehension and retention compared to using the LLM alone. Students felt the LLM helped more when they used it directly, but objectively learned less. This is the classic perceived-vs-actual learning gap (it was also seen in the Penn paper). It highlights the importance of productive struggle and desirable difficulties. If I have the time, I’ll write up a full breakdown of the study.

3. Agentic Everything by Teddy Svoronos. A reflection from a Harvard educator on the new tensions in teaching now that we have AI agents which can not only complete assignments, but also “able to do the jobs that we have been training students for”.

1. We recently opened up Bloom AI for invite-only access after two years of co-design with educators, institutions, and learning scientists. Check out our launch video below. If you want early access, leave a comment and we’ll reach out.

2. We recently won a Grand Prize of the EdTechnical AI in Education Forecasting Competition. Track 3 asked the question: By the end of 2028, what percentage of high school students will spend more than 2 hours per day in school learning through AI-powered, personalized and/or gamified educational content that adapts to their individual interests and learning pace?

   Our forecast: by 2028, only ~2% of US high school students will spend >2 hours per day learning through AI-powered, personalized systems. We predict, with very high probability, that AI systems will be created that are good enough to support learning growth equivalent to high-quality human tutors. But getting to AI for 25-30% of the school day requires overcoming political and institutional barriers such as timetable redesign, assessment alignment, teacher workflow changes, and change management. If institutional constraints move faster than we expect, adoption could be significantly higher, but history suggests structural change in schooling is slow.

   
   Read the full essay here.

When using AI, how do you tell whether you’re actually learning rather than just getting to the answer faster?