Report · Page 68

68
Center for Research on Foundation Models (CRFM)
2021], or even create personalised and adaptive learning experiences that tailor the learning process
to individual students’ needs and dispositions [Connor 2019].
Despite this potential, building technical solutions to effectively scale inclusively and quality of
education has proven to be exceptionally difficult. One particular challenge is that existing work
has focused on custom solutions to highly specific tasks for which large amounts of training data
has to be collected from scratch. Due to the difficulty and cost of creating large datasets, using
this approach to solve every educational task independently is fundamentally limited. Instead, is it
possible to create general-purpose approaches that are reusable across various tasks and subjects?
Foundation models have already started to boost the performance of some specific flagship tasks
in education. Recent examples include using MathBERT [Shen et al. 2021b] to power “knowledge
tracing" — the challenge of tracking a student’s understanding over time given their past responses —
and the “feedback challenge", where an algorithm has to interpret a student’s answer to a structured
open-ended task, such as a coding question [Wu et al. 2021e]. Can foundation models lead to even
more transformative changes in this domain? And what are the known and imagined risks of
foundation models applied to education? In this section, we first frame the conversation around
the ethical considerations. We then ground our discussion in two concrete tasks: (1) understanding
student misconceptions, and (2) improving student understanding through instruction.
3.3.1
Important concerns for centering foundation models in education research.
The future of AI for education is exciting, especially in the context of foundation models. However,
we caution the reader to be especially thoughtful about the impact of any AI research applied to
education.44 The goal of education are deeply interwoven with complex, long term social impact.
While we actively work to improve digital education, it is imperative that we put in substantial
thought to try and imagine the complexities of any disruption in this space [Piech and Einstein
2020]. Ethical challenges range from issues such as data bias, legal constraints, and the impact of
digital socialization. These issues are not unique to foundation models, but they are worth reflecting
on regularly as research makes substantial progress in AI for education. Reflection on impact is
especially important when research starts by asking “what can new AI technology afford?"
Many of the issues in §5.6: ethics apply to education. For example, as in many other domains,
small biases in foundation model training data could be hard to track down [Dixon et al. 2018;
Bolukbasi et al. 2016], but have important implications for equity of educational access. Moreover,
these systems may experience a high degree of “feedback", where the collected data continually
reinforces the model’s decisions. This issue of bias goes beyond what data is collected and in-
cludes concerns over the applications that researchers choose to work on. Below, we discuss other
education-specific issues. Many of the issues revolve around the question: “who benefits?" and for
whom is new technology created?
Removing teachers from the loop One of the goals of digital education, especially based on
AI, is to increase the productivity of the learning experience so that mor