K–12 Education in the Age of AI

By Shuchi Grover

The Henry and Bryna David Award—sponsored by the Division of Behavioral and Social Sciences and Education (DBASSE) at the National Academies of Sciences, Engineering, and Medicine and Issues in Science and Technology—honors a leading researcher who has drawn insights from the behavioral and social sciences to inform public policy. The recipient of the 2024 award, Shuchi Grover, is director of artificial intelligence and education research at Looking Glass Ventures. She delivered her Henry and Bryna David Lecture, “K–12 Education in the Age of AI: The Role of the Social Sciences in Shaping Learning Designs for a Transformative Technological Era” on October 10, 2024.

---

As powerful AI tools begin to affect our lives, what should learners and teachers understand in order to use them in ways that are empowering and equitable? What role can policy, research, and education stakeholders play to ensure that AI tools are leveraged to serve our goals of educating the next generation of citizens and problem solvers? And most importantly, what are enduring lessons from the social sciences in guiding and envisioning socio-technical learning systems in the age of AI?

Watch the lecture and follow along with the transcript below.

Transcript

My abstract promised a lot, and I must deliver, so my lecture is going to be a bit of a juggling act—a four-part juggling act. I’ll share my journey into becoming an interdisciplinary scholar. We’ll talk about AI education, and AI in education, and finally discuss key ideas from the social and learning sciences that can guide us in this endeavor. There’ll be some looking back and looking ahead, and zooming in to specific projects and zooming out to look at the big picture.

A. Journey and background: Becoming an interdisciplinary scholar

For background instead of showing you a boring timeline, I figured I’d tell my story through pictures. I was born in 1967 in a small town in India, the youngest of three daughters, to amazing, hardworking, and supportive parents. (Today would have made them happy and proud). I’m proud that ours was a family devoted to education. My mother was a teacher and both my sisters have been lifelong schoolteachers. While most people might remember 1967 as the Summer of Love, I like to think of it as the year Star Trek debuted, showcasing multiple computation devices, Seymour Papert birthed Logo, a language to teach kids about programming and powerful ideas, and ‘Shakey the Robot’ was created at Stanford Research Institute, or SRI, an organization that I would join 47 years later. I’m a true child of the computing age! I have always harbored a strong STEM identity, unlike everyone else in my family. When I was 14 and ranked second in my town in the National 10th-grade Board exam, I told a newspaper that interviewed me that I would go to the Indian Institute of Technology (IIT) or else do a Bachelor’s in Science and later become a computer scientist. Well, in 1984, I left home—not for IIT, but for BITS Pilani, a premier school of engineering and science in the desert of Rajasthan that was founded in the 60s with funding from the Ford Foundation and the curriculum guidance from MIT. (This was an article about BITS Pilani in the New York Times in 1979).

1984 was a momentous year in computing. It heralded the start of the worldwide personal computing revolution. Many will remember it from the famous Apple commercial. What it meant for me was that by the time I completed my prerequisites, I was part of the first cohort coding in the C programming language on personal computers, as opposed to punch cards on a PDP-11. 1984 also signaled the start of a sinister gender divide in the world of computing. As a student majoring in physics and computer science (CS), I was no stranger to the phenomena. Well, while girls comprised about 10% of the entire cohort that joined in 1984, I was the only girl in my physics cohort and amongst a handful in CS. 30 years later, when I conducted research with middle school students in the ‘Computers’ elective in Palo Alto, California, sadly the gender distribution looked much the same.

After a brief stint in software engineering, I decided to pursue further studies in CS and came to Case Western Reserve for a PhD program. My advisor was George Ernst, who had worked in AI, having earned his PhD with Allen Newell (of Newell and Simon fame) at Carnegie Mellon. I learned Lisp and did coursework in expert systems, but it was a project at the intersection of computing and education funded by IBM that changed the course of my life. In 1992 I joined the team on the Library Collection Services (LCS) project that aimed to leverage digitized music score sheets to build client-server multimedia applications for students of music. (Those adjectives sounded super cool at that time!) I developed an enduring love for Bach’s Brandenburg Concertos and Mussorgsky’s Night on Bald Mountain—the music that we were using for our pilot software. I remember my very first experience in the LCS lab with Mosaic, the first web browser. It was an exhilarating time. The internet was coming to life, and we were in the slump of an AI winter. I decided to drop out of the PhD program with only a master’s to join the booming software industry. (Given where AI is now, I do sometimes wonder about that decision).

But the LCS experience stayed with me. Perhaps it was the impetus to delve deeper into technology and computing in the service of education, or maybe it was the fact that I had a family full of teachers. But following the dotcom bust in 2001, I joined Harvard for a master’s in Technology, Innovation, and Education, engaging for the first time in readings on education and learning theories. I then devoted my time to working with children on robotics and programming, mostly in informal settings, and on preparing teachers for 21st-century teaching in India and also around the world through Harvard’s Wide World online program. When I finally enrolled for the PhD in Learning Sciences and Technology Design at Stanford in 2009 with Roy Pea as my advisor, we were in the thick of the social media boom while NSF was laying the groundwork for computer science and programming becoming formal topics of study in K–12. So, given my background and experiences, computational thinking (CT) and K–12 CS education became a natural area of focus.

My PhD qualifying paper from 2011 that I subsequently published in collaboration with Roy Pea, became a go-to paper for CT (computational thinking) with a whole lot more context for K–12 education than Jeannette Wing’s original seminal 2006 paper. (The paper has about 3,500 citations now, so I am proud of that work). Soon after I graduated in 2014, NSF launched a new program called STEM+Computing Research, which encouraged integrating CT into STEM learning. I won two large grants at SRI in short order. In January 2016, “CS for All” officially became a national movement with CT skills explicitly called out in President Obama’s remarks. I was part of a team that conducted a workshop that was a prelude to the NSF “CS for All: Research Practice Partnership” program that launched later that year. Since then, I have had about a dozen NSF grants in this area. It was also the age of big data. NSF’s Big Ideas in 2017, responded to the oncoming tsunami of AI and machine learning, and spurred research into AI and preparing a future workforce working at the human technology frontier. Three of these Big Ideas in particular (Harnessing the Data Revolution, Work at the Human-Technology Frontier, and NSF INCLUDES) inspired NSF research in AI in education, K–12 CS education, STEM+C Integration, data science, and broadening participation in computing.

And so here we are now, with AI dominating the global consciousness, especially with the arrival of ChatGPT, and AI developers seem to be getting Nobel Prizes in physics and chemistry too! We may finally be realizing the long-held dream of augmenting human intellect with intelligence amplifiers that visionaries like Doug Engelbart of SRI had written about 60 years ago—“The term ‘intelligence amplification’ seems applicable to our goal of augmenting the human intellect in that the entity to be produced will exhibit more of what can be called intelligence than an unaided human could.” (Engelbart, D. (1962). A conceptual framework for the augmentation of man’s intellect. Menlo Park, CA: Stanford Research Institute). I shared all this background not only to give a sense of my long and winding journey to becoming an interdisciplinary scholar, uniquely positioned to contribute to CT and K–12 CS, and AI education but also to underscore that we’ve been through this rodeo before, many times. We’ve now been through many transformative technological areas, such as the arrival of personal computing and the Internet in schools and classrooms. In addition to learning from those experiences, we have close to a century of research in education psychology, about 40 years of research in the learning sciences and computer-supported collaborative learning, about 30 years of research in bringing technology to classrooms, and about 20 years of research looking at how to bring technical topics like computer science and programming into K–12 schools. So when we think about our challenges ahead, a tabula rasa is not one of them. But like always, there is hype and rush to transform education, and this time it’s with AI. As folks at TeachAI (a global consortium on whose advisory board I serve) believe, doing nothing is not an option. A recent Brookings report also suggested that just like access to the internet and computers was in the past, who gets to use and understand AI is going to fuel the next digital divide. While that may be true, I think rapid deployment without careful thought could also engender a new kind of digital divide.  

B. Teaching AI in K–12 (AI Education)

K–12 education in the age of AI comprises two distinct aspects. The first is AI in education, which is patently about using AI in the teaching and learning process. The second—AI education—also encompasses two aspects—the first is about foundational AI literacy, which encompasses basic knowledge of AI, how it works, how to use it, including the risks involved, and the second is about learning the more technical concepts of AI and machine learning, so that students have an understanding of what’s going on under the hood, and they are empowered to create with AI. I see this distinction as analogous to digital literacy versus computer science education. In this talk I’ll focus on this second aspect of AI education. A lot of what I’m going to share here is drawn from a paper that was published earlier this year that also won best paper at ACM’s Flagship CS education conference (SIGCSE) in Portland. It’s not a meta-analysis, but rather a synthesis of what’s happening, and I’ll share some key ideas here. When it comes to AI in K–12, we have to address it all: a foundational literacy, and then in CS classrooms—teaching AI with data science alongside CS, integrating AI learning with other STEM subjects and language arts and social studies, and finally perhaps as an advanced elective in high school. We’re still figuring out the depth of experience for each of right now, because technical understanding of machine learning, for example, requires fairly advanced mathematics knowledge. But I do believe that AI and CS education efforts need to be intertwined (Grover, 2024). They are intertwined and they need to be dovetailed. And the last 10-15 years have seen an enormous amount of progress in CS education research. This is the nation’s report card, so to speak, on CS. It’s the percentage of high schools in each state that offer computer science. The national average stands at about 60%. Even though it was perhaps one of the few issues that has enjoyed bipartisan support in Congress in the current and past administrations, progress has been steady but slow. And it has not been even. There are divides by gender, race, and ethnicity to access by geography and school size. Hence the last few years have seen a redoubling of efforts to focus on issues of equity.

But that’s not all that we’ve learned. There are lessons from K–12 CS that can guide us to as we move forward with AI, and I see them as falling in five buckets. One is that we need to attend to crucial skills beyond just CS concepts and programming, and this includes socio-emotional skills and computational thinking. Then there’s integration—for students to see applicability and understand the use of CS in the real world, it needs to be taught in context. And we need to attend to conceptual learning through leveraging connections, community backgrounds—the socio-cultural and socio-political aspects. Then there’s plurality of pedagogies—no single approach works for all contexts and learners and topics. There’s unplugged, constructionist programming, and game-based approaches. All of those are valid depending on context. And teacher preparation, the most important of all. No effort to introduce computer science or AI will succeed without preparing teachers, co-designing with them, and also building on their teaching experiences.

I’ll give a glimpse of one project that involved integrating CT into pre-K STEM settings, where we made a very conscious home and school connection. Even though it was in the context of preschool learning, there are ideas here for primary grades as well. The project website is hosted by Digital Promise—it describes our research. We were guided by frameworks on early learning, early STEM learning, computational thinking (CT), and also by culturally responsive teaching, leveraging families’ funds of knowledge through connecting activities done in school and in the classroom, and supporting activities done at home (mostly Hispanic families that we were working with—parents, siblings, uncles, aunts, and grandparents). We co-designed mobile apps and unplugged play with teachers, parents, and caregivers. Not surprisingly, we saw significant improvements from pre-to-post for children who participated in the home plus school condition relative to the comparison condition.

So, coming to AI, it has been a major priority area for policymakers and consequently NSF research, these past 8-10 years. Around 2018, an NSF-funded project kickstarted AI in K–12 with the creation of AI4K12.org that led the development of national guidelines for teaching and learning about AI in K–12 school settings. One of their first efforts was to articulate these big five ideas of AI that we need to be teaching—Perception: Computers perceive the world using sensors. Representation and reasoning: Agents maintain models and representations of the world and use them for reasoning. Learning: Computers can learn from data. (And boy, can they learn!) Natural interaction: AI developers strive to create agents that interact naturally with humans. Finally, the big one—Societal Impact: AI can impact society in both positive and negative ways. They have created a Big Five Ideas poster which has been translated into many languages.

And then in addition to content and disciplinary concepts, it’s recognized that, as before, students also need to be learning computational thinking, which, as we’re all aware, comprises problem-solving approaches that transcend programming environments. It’s more relevant than ever now when AI can actually produce code. So students need CT and “code sense” rather than coding skills, per se. Over time, conceptualizations of CT have evolved and expanded and been articulated from Wing (2006) to Grover and Pea (2012) to Tissenbaum, Sheldon, and Abelson (2019) moving from CT to computational action, and finally, Kafai and Proctor making the case for including situated and critical framing in addition to the cognitive Incidentally, Work It Out Wombats is a new PBS kids’ show that engages little learners in computational thinking. It’s a project partly funded by the NSF, and I had the pleasure of serving as an advisor on that.

However, AI, and ML especially, have ushered in a new paradigm of computing, which is more data-driven. Where the old CT involves step-by-step algorithmic thinking approaches, machine learning algorithms are defined by data and a definition of success, which is essentially a confidence level rather than 100% accuracy, as it used to be. It’s more inductive rather than deductive, more black-boxed and non-deterministic. This chart came out from the work of Matti Tedre and his team (Tedre et al., 2020) in the generation AI project in Finland, a large project on which I serve as an advisor. (I seem to be advising a lot of efforts). These shifts necessitate a relook at CT as well. Students now need to learn both old rule-based CT, which still powers a lot of computing, but now they also need to understand data-driven and probabilistic computation that drives machine learning, and building on Kafai and Proctor, we also have situated dimensions (which include context), and critical, ethical, and justice-centered framings that prompt students thinking about ethics and risks and benefits (Kafai & Grover, 2025). You can’t think about computing, computation, and automation without thinking about all these things. There are a couple of examples of research that showcase these expanded notions of CT.

And then there’s the growing prominence of data science and data literacy, learning of data practices. It’s becoming a key competency given its importance in machine learning especially. It’s a topic of an ongoing National Academies consensus study. ‘Data literacy’, as we know, involves understanding of collecting, cleaning, analyzing, visualizing, and interpreting data. But there’s also a move to extend it to ‘data agency’, which emphasizes people’s ability to not just understand data, but consciously and actively control information flows. Think about the opt-out thing that keeps popping up. How should people respond to those kinds of things? What kind of data do you want collected about yourself? And then there’s ‘data equity’, which advocates for shared responsibility and fair data practices that respect and promote human rights, opportunity, and dignity. And maybe these are separate right now, but might become an expanded notion of data literacy. I think that’s where we’ll go.

Data equity leads me to this last critical aspect of learning AI, which is ethics. Today, ethics is being taught with much more urgency than was the case in computing education—critical interrogation of bias embedded in AI systems that creep in at various stages of the AI development and deployment cycle. And these are baked into almost all current curricular efforts. New frameworks such as Responsible AI and Tech Justice from the Kapor Center, and this brief on Ethical and Just AI provide guidance for education practice, and research.

In terms of curricula, the last five years have seen a rush to design curricula, many through research efforts funded by the NSF, and they leverage the many freely available tools and activities that allow students to explore machine learning and neural networks, etc. Many of them are designed by the creators of AI itself. These are from Google, others by Microsoft. Then there’s a plurality of pedagogies for teaching AI and ML. There’s exploration of pre-designed models and tools (like I just showed), digital non-programming interactives, constructionist sense-making, unplugged activities, using APIs for classification, lifting the hood through programming models, integrating movement and cultural practices, socially relevant AI apps co-designing ML applications and ethics focused curricula, and there’re citations for all of them in Grover (2024), if you want to see, for all the research projects.

I’ll delve into one quick example from one of my projects called ‘CS Frontiers’ (csfrontiers.org) that aims to engage high schoolers, especially females, in emerging topics such as machine learning and AI, cyber security, the Internet of Things, and distributed computing. Basically, all the topics that surround students these days but they don’t get to engage much through current CS curricula. These are my fabulous collaborators at North Carolina State and Vanderbilt. These are the graduate students who have worked on this project. And there’s an incredible team of teacher partners that have piloted our curriculum in summer camps and their classrooms. Special shout out to April Collins, who’s been using it as a year-long curriculum in her class. These are essential ideas and philosophies that underpin the CS Frontiers units—engage underserved groups (especially girls), engage in CT and data science, project-based learning, collaborative activities, and inter-disciplinary learning. The unique piece here is that we are bringing questions and data from other disciplines into AI classrooms.

We used NetsBlox (netsblox.org), which is an extension of the Snap! block-based programming environment. Basically there’s one block that they have added that essentially opens up Snap! to the internet, unleashing a world of possibilities by allowing you to access real world data sets and web services through leveraging APIs. And this ‘Call’ block, for example, allows you to access data on air quality and COVID and hurricanes and NASA and ocean data. And all this is just under the Science services. Similarly, there’s climate and New York Times and Google Maps and movie database and image stuff and traffic and Twitter and what not. And the best thing is that it extends to user-created data sets, too, with your own data. So, think of how contextual students’ inquiry can be. So, this example shows how you can call the climate service, which gets carbon dioxide levels from the Mauna Loa Observatory in Hawaii and passes it to a chart drawing service. And voila, you have a visualization in seconds of actual carbon dioxide levels from 1960 to 2020. Here’s an example where you access data services for weather and overlay on Google Maps. Similarly, for kids that are interested in astronomy, you could be displaying and navigating a sky map. This is a student project where students used an online movie database to create a trivia game. And finally, charting COVID-19 data from the Johns Hopkins database was particularly popular when we ran the research during COVID. All of these projects show how students can engage in data science and data practices of not just pulling in data, but also pre-processing and visualizing and working with abstractions and such.

The AI and ML module has as its core ethics in ML and biases examining those perpetuated from preexisting data sets, et cetera. An example activity, for example, is this Sentimental Writer where students call the Parallel Dots API, which is a natural language processing API. And it basically classifies sentiment in positive, negative, or neutral, or can classify hate speech or sarcasm and things like that. And students’ final projects of choice included examining biases of New York Times articles, or the sentiments of Taylor Swift’s song lyrics. And how do we integrate the ethics of AI? For example, students play with the Google Quick Draw app. They examine how the machine has learned to classify objects from thousands of images, and then we go to the GitHub pages and examine the data set and find that Google captured a bunch of data and discuss how people’s data gets collected without their explicit permission. Is that right or wrong? And in some cases, maybe it doesn’t make a difference. But this discussion is very topical and pertinent, given the news from just a couple days ago. That Meta is training its models on users videos and images from the smart glasses without explicit permission. Here are some papers and publications from CS Frontiers if you’d like to read more about NetsBlox, the project curriculum.

I just want to quickly mention one more research effort in which we’re integrating AI learning with cybersecurity learning for high schoolers. I mentioned this because in this project, we actually worked to lift the hood on how machines learn, how classifiers and decision trees are coded, how optimization works by repeatedly evaluating and the performance of the model, and making tiny tweaks, which is basically gradient descent. And we believe that lifting the hood allows students to get a better sense for how biases can creep into machine learning, not only through training data and deployment, but through the optimization process itself. And the pedagogical approach includes learning in context, making connections with cybersecurity scenarios, games and guided exploration, scaffolding, sensemaking, and collaboration. These are some of the publications so you can take a look.

In light of these shifts, the field sees these new needs and is basically working to reimagine CS pathways. The CS education community and CSTA are revising CS standards. And so there’s a lot going on to sort of, you know, keep up with all of this. Teach AI is also releasing briefs to guide CS students, and I’m involved in this work as well.

C. Teaching & Learning with AI (AI in Education)

 So, coming to teaching and learning with AI. It’s worth noting that AI and education is not new at all. The goal of this graphic is not to go through it, but to underscore that AI in education has progressed in parallel with the progress of AI itself. And through the years NSF and other federal agencies have funded many of these efforts. Currently the five National AI Research Institutes set up by NSF for a total investment of about 100 million represent the state of the art in research on AI augmented learning and education. The primary focus across these institutes is to advance AI driven innovation, to improve human learning and education, and address the grand challenge of education for all. Although each institute has specific research objectives, they all share a common goal to improve student learning outcomes, particularly addressing disparities impacting students from historically marginalized communities.

Now, coming to the long history of AI and ED, I think the most prominent AI efforts in recent memory took off in the 80s and 90s, and were centered on cognitive tutors and intelligent tutoring systems. They were initially designed on the behaviorist model of learning, where individualization was basically just a matter of individual pacing. But as computational true tools have become more sophisticated, there’s been increasing adaptivity. It has pawned new subfields at the intersection of computer science and education like Education Data Mining, Learning Analytics, AI in ED, Learning@scale. And the clickstream of big data from digital environments coupled with computational analytics, have afforded more adaptivity and personalization. There’s increasing attention now to (a) socio-emotional factors in learning, (b) multimodal analytics—where you’re looking at learner cues besides log data such as gaze and gesture and facial expressions, speech, (c) supporting not just individual learners, but collaborative learning, (d) increasing support for teachers, teacher dashboards, and combining teacher knowledge with predictive analytics. I think all tech basically comes around to working in the grammar of school or finding ways to do that. Increasingly, human centered and inclusive designs as focused on educational equity has increased.

But these have been slow to scale. They require adoption in classrooms, planning, funding, stakeholder buy-in, professional development, which is distinct from what happened with generative AI. It was not adopted. It arrived as Klopfer and Reich and others (Klopfer et al., 2024) at MIT have said. Generative AI came directly into the hands of students and teachers to use as they wished. It’s more general purpose as opposed to the old “narrow AI”. And it’s embedded in many tools that students and teachers already use, like Google Classroom or Gmail and things like that. And NSF has come out with targeted calls for proposals to basically examine teaching and learning with generative AI in particular. So, Jeremy Price and I recently authored a brief for CADRE, which is currently under review, the DRK–12 NSF Resource Center (Price & Grover, 2025). And we examined some recently funded efforts from the NSF. And here’s a sampling of how researchers are envisioning generative AI use. So, for example, LLM chatbot to assist teachers in developing culturally responsive curricula using natural language prompts; having a math LLM generate technology, enhanced formative assessments, which require coding at the backend. (that’s my project with Corinne Clarkson); “digital assistants” which tailor lesson plans according to teacher goals and student characteristics; AI image generators that create images that are not available otherwise; an AI instructional assistant that engages other students in classroom while a teacher is working with individuals and groups in STEM classrooms; giving instructors feedback on their dialogue in instructional practices; classifying students’ expertise in scientific explanations aligned to NGSS; “human in the loop” NLP to automatically score long-form responses in formative check-ins in a curriculum; and generative AI and LLMs that provide teachers real-time feedback during teacher PD.

I’ll talk a little bit about my project, which is centered on formative assessments in high school math classrooms, and on giving teachers the ability to create their own interactive auto-gradable isomorphic technology-enhanced assessments, which are more engaging and you can generate seemingly infinite versions of them. They typically require WeBWorK. It’s a format in which you can do these kind of mathematical things and it requires coding at the backend (Perl code), so teachers were literally left out of the equation (no pun intended) of creating their own assessments. So, the creators of Edfinity—an NSF SBIR funded assessment platform and Homework System by Looking Glass Ventures—trained an LLM called ALICE, which takes in English language prompts and generates an interactive assessment along with hints and a solution. (This is the code, and these are the English language prompts. This is the hint and solution, and these are sort of the interactive assessments). So, we conducted research in Indiana and Illinois with high school teachers that are part of the dual enrollment program of Indiana University. We collected data—(a) pre-post surveys, (b) teacher logs of prompts, prompt revision, their feedback, their opinion on the problem quality, (c) in-depth teacher interviews, and (d) a corpus of over 500 problems.

And the themes from teachers open-ended survey responses talked about efficiency, ease of use, the challenges and learning curves with prompt engineering (that was a big one), teacher empowerment attitudes toward AI. Overwhelmingly positive responses from teachers about the experience. But it wasn’t exactly even. In-depth interviews revealed that there were several factors that impacted their experience, and their answers to whether they would use a tool like ALICE depended on the (a) teaching context, (b) their attitudes towards AI—the ones that were not particularly excited? Turns out they’re not particularly excited about software or AI in general. (c) learning curve with respect to prompt engineering, (d) whether or not they viewed generative AI as a thought partner, (d) alignment with their teaching philosophy, their business as usual, (e) utility and functionality, and (f) student engagement and interaction—the ones that were really excited actually had students suggesting prompts or topics for prompting or students analyzing the quality of the problem, et cetera. And I share this because in essence, this encapsulates what will make LLMs useful and successful in classrooms. It’s attending to these issues. Each of these represents a vector along which AI systems need to be aligned, but currently are not. We’ll revisit this issue of alignment later.

Here are some more considerations or questions for LLM use in classrooms that have emerged from the brief (Price & Grover, 2025):

• Can an LLM be fine-tuned, augmented customized for a specific topic or context?
• How might we design tools to give teachers agency and control to do this sort of last mile of customization and prepare them to do this?
• What pedagogies and cognitive models of student learning are encoded in the LLM? There’s no one size fits all in learning.
• Whose language and worldviews are privileged? How should biases be addressed?
• How can we promote rather than hinder culturally relevant designs? (It’s already been shown that there are racial and cultural biases in image generators and speech recognition.)
• Which model should be used for which task? Not all LLMs answer prompts the same way. Models are untested yet, and they misinform.
• How can we leverage things that we know from learning theory and research, for example, learning progressions?
• How can situational awareness of the classroom be incorporated and leveraged?
• How can we center teachers, learners, and the social culture of the classroom and the communities in which learning happens?

And when it comes to the holy grail of personalized learning, there’s always been a hope that it’ll provide real time personalized support for student learning, engage in conversations, etc., and it feels like we are closer than ever to that one-on-one tutoring and Bloom’s two sigma promise. That we can prepare learning materials and pathways that are customized to each learner—an IEP for every student! Empower students by offering diverse learning opportunities. Maybe have students creating their own materials since we have such creative generative AI, foster connected and interest driven learning across pace and time, engage in culturally appropriate pedagogical practices, especially for students from non-dominant cultures, support students with disabilities and other specialized learning needs and supercharge teacher capacity and address teacher shortage.

But there are challenges—

• ML models could misdiagnose or misclassify learners by overlooking important factors,
• Exacerbate existing inequities and biases towards learners from different cultural and racial backgrounds. This is possible because the training data (the internet) is so skewed to the global North, to Western English-speaking societies, and they may not be appropriately responsive to wide ranging needs of diverse student populations.
• They may also track or pigeonhole students into limiting paths based on biased data. (Think of Ray McDermott (2001)’s “acquisition of a child by a learning disability” and mislabeling of learners).
• Learner isolation, lack of collaborative learning,
• A corrosion of the social climate of the classroom, and loss of human relationships with teachers and peers, which is key to learner wellbeing.

So, to me, an ideal AI in education system needs to have situational awareness. It needs to be built with research on teaching and learning, and attend to issues of ethics and justice, which is a great segue to the last portion, portion of the talk, AI powered sociotechnical systems and how the social and learning sciences can guide us.

D. Socio-technical systems designs: Guidance from the social & learning sciences

So, to understand what students need to learn and the kinds of skills they need to develop, I’ve found the “deeper learning” framework to be particularly useful in my work. If you’ve seen my past talks, you may have seen this before. Deeper learning basically draws attention to the fact that all learning experiences need to be designed to attend to, not just disciplinary competency and knowledge (cognitive skills), but intrapersonal skills such as metacognition, self-regulation, learning to learn, persistence, identity, mindsets, and the interpersonal, which are communication, collaboration, respect for the other and such. And the focus on cognitive skills needs to be not just on content knowledge, but also problem-solving habits of mind of the discipline and such. When we learning scientists set out to design social technical systems (STSs) of learning, we usually draw on some combination of these many theories that come from these great books of ours. And it’s from these that we design for project or problem based learning, or immersive and embodied learning or learning through visualizations and simulations emerge. Any learning scientists in the room, I don’t know if I’ve sort of missed any of the key ones out, but I think these are the ones that I thought about.

But to understand human learning at its core, these amazing scholars (Nasir, Lee, Pea, & de Royston, 2020) who authored the handbook on the ‘Cultural Foundations of Learning,’ put it very succinctly and well. Human learning and development “are cultural processes in which language and linguistic structures shape, identity and knowledge,” “are tied to place and the natural world,” “are fundamentally relational and evolve over the life course,” and “fundamentally involve knowledge systems and must be understood within axiological systems of values and ethics.” These systems are tied to and give shape to power, politics, and societal structures. So, from these pearls of wisdom, I’ll distill three key ideas about learning, that you can take away, that must shape the design of social technical systems. Learning is social and cultural, it’s situated and contextual, and it’s deontic or values-driven.

And so, what this means in terms of considerations and features for social technical systems is that they need to be contextually relevant, culturally responsive, tailored or adaptive to individual learner as well as collective needs, foster and enhance agency, support collaboration, provide cognitive scaffolding, afford metacognitive supports, offer real-time feedback and assessment, and be ethically and responsibly—designed with respect and dignity for individuals and recognizing issues of positioning and power, and maybe be cross-disciplinary and integrative (because we know that that makes for more authentic learning). And if we are lucky, they should be scalable and sustainable too.

In the interest of time, I’ll discuss only a couple of ideas in more detail. The first is this idea of scaffolding. Basically it’s situations “in which the learner gets assistance or support to perform a task beyond his or her own reach if pursued independently when “unassisted” (Wood et al., 1976. p. 90), and fading is an intrinsic component of the scaffolding framework. (Pea, 2004), which means as learners become more competent, you need to provide less and less support. It’s very key. And perhaps the best paper on scaffolding is this one by Roy Pea, which presents these two axes that afford various levers and kinds of support for learning the social and the technological And this is also related to this other influential work by Roy, which is distributed intelligence (Pea, 1993). And this is a theory that’s increasingly relevant as we design for human AI teaming in classrooms. And I found it to be very useful as a theoretical framework in my grant proposals as well. Distributed intelligence refers to the perspective that intelligence is not best depicted as a trait of individual minds, but as distributed across people, artifacts, materials, and digital tools available in sociotechnical systems.

Talking about culture, I feel funds of knowledge, it’s something I had talked about briefly in the pre-K CT project. I want to talk a little bit about that. Funds of knowledge are basically historically developed and accumulated strategies or bodies of knowledge that are essential to functioning and wellbeing. Every culture has funds of knowledge. And I mention this because there’s a tendency to view indigenous and non-Western cultures as somehow inferior. But we are all enriched when we view funds of knowledge as an asset. We create identity-safe classrooms that support the education of the whole child. This relies on teachers understanding the view and experiences that children bring to the school from their homes and communities. And this means developing practices like culturally sensitive listening that capitalize on funds of knowledge that are abundant in students, households, and communities rather than a deficit based orientation. And this benefits all students, not just students from non-dominant cultures.

To summarize, all AI powered sociotechnical technical systems must be context sensitive, supporting and adapting to specific learning environments with specific needs of learners, and offering relevant supports and tools. They need to be socially aware; they need to be culturally competent, provide cognitive scaffolding, support, collaboration, serve as collaborative agents, be sustainable, accessible, and affordable and responsible and trustworthy. That’s a tall order. There’s a reason why EdTech usually fails. And such work requires interdisciplinary expertise in learning research, social sciences, the technical fields, AI, CS, data science, ethics and such.

This last point is particularly important in the age of AI— trust in AI-powered systems. Trust is a firm belief in the reliability, truth, or ability of someone or something. And it turns out, unsurprisingly, that about two out of three people are wary about trusting AI systems around the world. And social sciences, research on trust and morality (the two are connected), provides guidance on what AI systems need to be to be deemed trustworthy. Voiklis et al. analyze the “trust fall” between humans and ask the tantalizing question about whether you would fall if it were a robot rather than a human behind you. The human calculus is basically around issues of (a) competence, (b) reliability, (c) sincerity, (d) integrity, and (e) benevolence. You fall only if you’ve answered ‘yes’ to all those questions. And they have this framework that encompasses this experience-based trustworthiness that we just talked about, (the five factors), and also an identity-based trustworthiness (likability or affinity). And these issues of trust have a bearing on our personal interactions with bots. In fact, there’s a danger that we may trust them too much. Try, as we might, it’s hard to not anthropomorphize these technologies and interact with them as we would with other humans. This idea was presented many years ago by Stanford’s Byron Reeves and Clifford Nass in their (2003) book ‘Media Equation.’ And these issues of not enough or too much trust becomes increasingly relevant as we move from text-based interactions to speech-based, and finally, humanoid interactions.

On the policy front, the Department of Education’s Office of Technology Policy has used the White House Blueprint of an AI Bill of Rights and the EU equivalent to produce a set of recommendations for AI teaching and learning, and the key takeaways from that are human-in-the-loop AI, AI as a tool to augment human capabilities, think electric scooter rather than robotic vacuum, and always center teachers “ACE”. And recently in July (2024), they came out with this other guide, which is about aimed at EdTech developers, and it’s about building tools for classrooms. They encourage co-design with educators and balancing responsibility with innovation. TeachAI.org is actively working with states to produce resources and guidance to help policy makers, especially at the state level, on how to bring AI tools to classrooms. As of September of this year, 24 states have published AI guidance already, and many of the 44 education state agencies participating are creating guidance.

I’ll conclude on a philosophical note addressing the third of those (three foundational aspects of learning)—that education and learning is deontic. It’s values-driven. Deontology is a kind of normativity that binds people with respect to their actions. Deontology is (the ought) where ontology is (the is), and human culture is a fusion of these, embodied in the reality that functions to bind together a human community (Martin Packer and Michael Cole, The Handbook of the Cultural Foundations of Learning, 2020). And learning at its very core is driven by core human values for what the aims of education are. There are moral and ethical commitments to education that education demands. They focus on the broad idea of the aims and purposes of education and the accompanying duty to action. So, deontic logic is essentially the logic of ethics. And because the deontic is so fundamental, a primary issue in AI is alignment with human values. This book by Brian Christian (The Alignment Problem) is a must read. He says that there is an alignment problem when systems learn by example rather than by being explicitly programmed. For example, model parameters are tuned based on an optimization algorithm—Who or what are they optimizing for? Who defines that objective function? Who decides how the models should be trained and on what data? Are these systems learning what we think they’re learning? Can we trust them to do what we think they’re going to do? Why the rush to push LLMs into classrooms? Who benefits from the widespread use of AI? Who suffers? What is the real cost? Is it worth the cost?

Because today we have LLMs that have automated everything, including our worst instincts. They’ve encoded all the “isms”— racism, sexism, ableism. We’re all aware of just how pernicious AI can be, perpetuating the biases that exist in our societies in ways that were not always so apparent in computing before AI. Communities of color are more likely to be negatively impacted by new technologies, and too often technologies are designed and deployed without the inputs of the very people they harm the most. Beyond issues of bias, there are unimaginable dangers of cyber bullying, sexual exploitation, and loss of privacy that these technologies pose. This one of the Meta smart glasses being just the latest example. Not to mention the catastrophic environmental damage AI technologies are wreaking. What happened to our climate crisis and pledges to save the planet? We’re so caught up in addressing the what and how, that we’re not stopping to ask why and whether or not we ought to use these tools, or maybe when, maybe not never, but when? Tools that pose so many risks and dangers. It feels like the tail is wagging the dog right now. And this is not to say that AI is to be avoided. It seems inevitable because, as Allan Collins said, “the nature of education must inevitably adapt to the nature of work in society.” But there has never been a silver bullet in education, and AI won’t be. We’ll be wise to heed these words from a scholar who articulated it better than I can—“The path of educational progress more closely resembles the flight of a butterfly than the flight of a bullet.” (Philip Jackson, 1968). Computers can do what they do well, but what they can do well may not be the best for students development, learning or instruction.

And so, I’ll close with a few thoughts:

• Recognize that education and learning are very complex. Be guided by social and learning sciences and designing these systems.
• We need to develop an infrastructure for building capacity to thoughtfully deploy AI—Teacher capacity. District level expertise. State-level capacity. Bridging the digital access, use, and design divide. This is from NETP 2024. It’s hard to believe that this is a need even in 2024. And research that develops the science of human-AI co-evolution.
• Slow hybrid growth informed by research may be a way forward.
• Reimagine the study of AI. I’ve been thinking about this, AI and social sciences. They need to be taught in a very integral way. Designers of AI and AI tools for education must have knowledge of and formal training in the social sciences. And if they resist, create a new discipline, if necessary.
• Privilege humans and their wellbeing above all else.

Thank you. I’d like to acknowledge my research collaborators over the years, and NSF, whose grants have supported most of my research. All these folks have shaped my thinking on all of this. Thank you.