ChatGPT Isn't the Only Way to Use AI in Education
Soon after ChatGPT broke the internet, it sparked an all-too-familiar question for new technologies: What can it do for education? Many feared it would worsen plagiarism and further damage an already decaying humanism in the academy, while others lauded its potential to spark creativity and handle mundane educational tasks.
Of course, ChatGPT is just one of many advances in artificial intelligence that have the capacity to alter pedagogical practices. The allure of AI-powered tools to help individuals maximize their understanding of academic subjects (or more effectively prepare for exams) by offering them the right content, in the right way, at the right time for them has spurred new investments from governments and private philanthropies.
There is reason to be excited about such tools, especially if they can mitigate barriers to a higher quality or life--like reading proficiency disparities by race, which the NAACP has highlighted as a civil rights issue. Yet underlying this excitement is a narrow view of the goals of education. In this framework, learners are individual actors who might acquire new knowledge and skills with the help of technology. The purpose of learning, then, is to master content--often measured through grades and performance on standardized tests.
But is content mastery really the purpose of learning? Naming reading proficiency as a civil rights issue likely has less to do with the value of mastering reading itself, and more to do with the fact that mastery of reading (or math, or other subjects) can help lay a foundation for what learning can unlock: breaking the intergenerational cycle of poverty, promoting greater self-awareness and self-confidence, and cultivating a stronger sense of agency over one's destiny and the destinies of one's communities. Content mastery is part of this equation, but making it the primary focus of education misses the fact that so much of a child's future is shaped by factors beyond the classroom. Critically, networks, or who children and their families are connected to, and how, matter for helping children prepare to live fulfilling lives. This is especially true for networks that cut across socioeconomic, demographic, and other lines. Indeed, a large recent study highlighted how social capital, defined as friendships across socioeconomic divides, can play a larger role in fostering intergenerational economic mobility than school quality (often measured by the test scores of students who go there).
Networks that connect parents to coaches to help them navigate their children's schooling can forge new support structures and trusting relationships between families and educators. Networks that connect students to role models and mentors can change the course of their academic and professional lives. A child's broader social context, in addition to the knowledge and skills they gain through school, matters deeply for their future outcomes. Left without intervention, however, real-world networks often form and evolve in inherently unequal ways. For example, patterns of preferential attachment can lead "the rich to get richer," excluding many from accessing connections that might improve their lives in important ways.
In practice, each AI needs an objective function that represents what it is optimizing for. Applications of AI for pedagogy and content mastery might optimize for "helping students get the highest possible score on a test." Fostering more inclusive network connections, however, is a more deeply rooted and structural type of change than improving test scores. Using AI to help cultivate these networks might do more for children's life outcomes than focusing on pedagogy and content mastery alone.
But some may argue that optimizing network connections is a more nebulous task than optimizing test scores. What, precisely, should the objective function(s) be?
One framework for exploring this may involve focusing on how the networks that children and families are enmeshed in form and evolve in the first place. In the context of schooling, this involves the wide range of policies that school districts design to determine which schools students can attend ("school assignment policies"), along with the practices families adopt when picking schools for their children under these policies. Such policies and practices have historically perpetuated harmful features like school segregation by race and socioeconomic status--which, despite nearly 70 years since its formal outlawing, continues to define public education in the US. Many scholars argue that demographic integration has historically been one of the most effective methods not only for enhancing the academic preparation of historically disadvantaged groups, but also for fostering greater compassion and understanding--say, an ethic of pluralism--across people from different backgrounds.
AI can help support the design of more equitable school assignment policies that foster diverse and integrated schools, for example, by supporting district-level planning efforts to redraw "school attendance zones"--i.e., catchment areas that determine which neighborhoods feed to which schools--in ways that seek to mitigate underlying patterns of residential segregation without imposing large travel burdens and other inconveniences upon families.
Existing researcher-practitioner partnerships--and some of my own research with collaborators Doug Beeferman, Christine Vega-Pourheydarian, Cassandra Overney, Pascal Van Hentenryck, Kumar Chandra, and Deb Roy--are leveraging tools from the operations research community and rule-based AI like constraint programming to explore alternative assignment policies that could optimize racial and socioeconomic integration in schools.
These algorithms can help simplify an otherwise cumbersome process of exploring a seemingly infinite number of possible boundary changes to identify potential pathways to more integrated schools that balance a number of competing objectives (like family travel times and school switching). They can also be combined with machine-learning systems--for example, those that try to predict family choice in the face of boundary changes--to more realistically estimate how changing policies might affect school demographics.
Of course, none of these applications of AI come without risks. School switching can be disruptive for students, and even with school-level integration, segregation can persist at smaller scales like classrooms and cafeterias due to curricular tracking, a lack of culturally responsive teaching practices, and other factors. Furthermore, applications must be couched in an appropriate sociotechnical infrastructure that incorporates community voices into the policymaking process. Still, using AI to help inform which students and families attend school with one other may spark deeper structural changes that alter the networks students connect to, and by extension, the life outcomes they ultimately achieve.
Changes in school assignment policies without changes in school selection behaviors amongst families, however, are unlikely to lead to sustainable transformations in the networks that students are tapped into. Here, too, AI may have a role to play. For example, digital school-rating platforms like GreatSchools.org are increasingly shaping how families evaluate and select schools for their children--especially since their ratings are often embedded across housing sites like Redfin, which can influence where families choose to live.
Some have argued that school-rating platforms, where ratings largely reflect test scores--measures notoriously reflective of race and income and not as indicative of how much schools actually help students learn--might have historically led white and affluent families to self-segregate into neighborhoods zoned for highly rated schools, creating a vicious cycle of residential segregation that reinforces patterns of school segregation and ensuing achievement gaps. A recent research project I did in collaboration with Eric Chu, Doug Beeferman, Rebecca Eynon, and Deb Roy fine-tuned large language models to explore how parents' open-ended reviews on GreatSchools might contribute to such trends. Our results showed that parents' reviews are strongly associated with school-level test scores and demographics, and not associated with measures of student progress, suggesting that parents who consult reviews to make schooling choices may be factoring demographics more than actual school effectiveness into their decisions.
GreatSchools continues to invest in new ratings schemes that seek to break these feedback loops and offer a more complete view of school quality--as Sisyphean a task as it may seem. What if platforms like GreatSchools also trained and deployed school recommender systems that simultaneously try to expose families to schools that satisfy their desires for their children (for example, rigorous course offerings, language immersion programs, compassionate and nurturing teachers) while also exposing them to schools "outside of their bubbles"--that is, quality schools they might not otherwise consider, perhaps because they have lower test scores, are in neighborhoods they wrote off before ever exploring, or something else? This multi-objective AI would not come without challenges of transparency and agency that accompany recommender systems deployed in other settings, but it could help spark new network connections that may not form otherwise.
These are just some examples, and they are not mutually exclusive with pedagogically focused applications. For example, while we likely lack the data to do this today, looking ahead, AI might help determine which students would benefit the most from which tutors--those who can not only help bridge learning gaps but also serve as relevant sources of mentorship, guidance, and inspiration. And expanding our focus in AI for education to include networks will not absolve us of the fairness concerns and other risks that existing deployments of AI continue to pose. Designing new applications of AI calls for careful and thoughtful exploration, especially as we as a society continue to respond to our rapidly changing AI landscape with a dynamic blend of fear, hope, concern, awe, and wonder. Of course, as in life itself, all of these emotions are important. Harnessing them to foster more inclusive network connections for the next generation of learners may be our most meaningful response of all.