In the 1980s, the educational psychologist Benjamin Bloom introduced the 'two-sigma problem' - that students who were tutored one-on-one using a mastery approach performed on average two standard-deviations (two-sigma) better than students educated in a more 'traditional' classroom setting. That research is often taken as a benchmark for how good an educational intervention might be (relative to a traditional classroom baseline). The problem, of course, is that one-on-one tutoring is not scalable. It simply isn't feasible for every student to have their own personal tutor. Until now.
Generative AI makes it possible for every student to have a personalised tutor, available 24/7 to assist with their learning. As I noted in yesterday's post though, it becomes crucial how that AI tutor is set up, as it needs to ensure that students engage meaningfully in a way that promotes their own learning, rather than simply being a tool to 'cognitively offload' difficult learning tasks.
One promising approach is to create customised generative AI tools, that are specifically designed to act as tutors or coaches, rather than simple 'answer-bots'. This new working paper by the LearnLM team at Google (and a long list of co-authors) provides one example. They describe an 'AI-augmented textbook', which they call the 'Learn Your Way' experience, which:
...provides the learner with a personalized and engaging learning experience, while also allowing them to choose from different modalities in order to enhance understanding.
Basically, this initially involves taking some source material, which in their case is a textbook, but could just as easily be lecture slides, transcripts, and related materials from a class. It then personalises those materials to the interests of the students, adapting the examples and exercises to fit a context that the students find more engaging. For example, if the student is an avid football fan, they might see examples drawn from football. And if the student is into Labubu toys, they might see examples based on that.
The working paper describes the approach, reports a pedagogical evaluation performed by experts, and finally reports on a randomised controlled trial (RCT) evaluating the impact of the approach on student learning. The experts rated the Learn Your Way experience across a range of criteria, and the results were highly positive. The only criterion where scores were notably low was for visual illustrations. That accords with my experience so far with AI tutors, which are not good at drawing economics graphs, in particular (and is an ongoing source of some frustration!).
The RCT involved sixty high-school students in Chicago area schools, who studied this chapter on brain development of adolescents. Half of the students were assigned to Learn Your Way, and half to a standard digital PDF reader. As the LearnLM Team et al. explain:
Participants then used the assigned tool to study the material. Learning time was set to a minimum of 20 minutes and a maximum of 40 minutes. After this time, each participant had 15 minutes to complete the Immediate Assessment via a Qualtrics link.
They then did a further assessment three days later (a 'Retention Assessment'). In terms of the impact of Learn Your Way:
The students who used Learn Your Way received higher scores than those who used the Digital Reader, in both the immediate (p = 0.03) and retention (p = 0.03) assessments.
The difference in test outcomes was 77 percent vs. 68 percent in the Immediate Assessment, and 78 percent vs. 67 percent in the Retention Assessment. So, the AI-augmented textbook increased student learning and retention by about 10 percentage points in both immediate learning and in the short term (three days). Of course, this was just a single study with a relatively small sample size of 60 students in a single setting, but it does offer some promise for the approach.
I really like this idea of dynamically adjusting content to suit students' interests, which is a topic I have published on before. However, using generative AI in this way allows material to be customised for every student, creating a far more personalised approach to learning than any teacher could offer. I doubt that even one-on-one tutoring could match the level of customisation that generative AI could offer.
This paper has gotten me thinking about the possibilities for personalised learning. Over the years, I have seen graduate students with specific interests left disappointed by what we are able to offer in terms of empirical papers. For example, I can recall students highly interested in economic history, the economics of education, and health economics in recent years. Generative AI offers the opportunity to provide a much more tailored education to students who have specific interests.
This year, I'll be teaching a graduate paper for the first time in about a decade. My aim is to allow students to tailor that paper to their interests, by embarking on a series of conversations about research papers based on their interests. The direction that leads will be almost entirely up to the student (although with some guidance from me, where needed). Students might adopt a narrow focus on a particular research method, a particular research question, or a particular field or sub-field of economics. Assisted by a custom generative AI tool, they can read and discuss papers, try out replication packages, and/or develop their own ideas. Their only limits will be how much time they want to put into it. Of course, some students will require more direction than others, but that is what our in-class discussion time will be for.
I am excited by the prospects of this approach, and while it will be a radical change to how our graduate papers have been taught in the past, it might offer a window to the future. And best of all, I have received the blessing of my Head of School to go ahead with this as a pilot project that might be an exemplar for wider rollout across other papers. Anyway, I look forward to sharing more on that later (as I will turn it into a research project, of course!).
The ultimate question is whether we can use generative AI in a way that moves us closer to Bloom’s two-sigma benefit of one-on-one tutoring. The trick will be designing it so that students still do the cognitive work. My hope (and, it seems, the LearnLM team’s) is that personalisation increases students' engagement with learning rather than replacing it. If it works, this approach could be both effective and scalable in a way that human one-on-one tutoring simply can’t match.
[HT: Marginal Revolution, for the AI-augmented textbook paper]
No comments:
Post a Comment