Title: 2025 AI+Education Summit: Harnessing AI to Understand and Advance Human Learning Resource URL: https://www.youtube.com/watch?v=JA76MG1yCVc Publication Date: 2025-03-03 Format Type: Video Reading Time: 47 minutes Contributors: Michael Frank;Emma Brunskill;Victor Lee;Patrick Gittisriboongul; Source: Stanford HAI (YouTube) Keywords: [Artificial Intelligence, Education, AI Literacy, Generative AI, Child Development] Job Profiles: Data Scientist;Artificial Intelligence Engineer;Learning and Development Specialist;Innovation Manager;Research and Development Manager; Synopsis: In this summit moderated by Patrick Gittisriboongul, educators Victor Lee, Emma Brunskill, and Michael C. Frank explore how AI can accelerate educational innovation, model child learning, and redefine AI literacy in schools. Takeaways: [AI systems like DeepMind now outperform 90% of elite math students, yet U.S. educational innovation lags decades behind in pace., Current generative AI struggles to simulate how real students learn over time, limiting its utility for authentic educational modeling., Simulated educational "experts" using AI can help generate and evaluate new learning content, though alignment with human educators remains imperfect., Emerging child-focused AI models trained on first-person toddler data show promise but highlight significant limitations in scaling visual learning., The concept of "AI literacy" is highly fragmented, spanning critical use, understanding of algorithms, and developer-level competencies, making unified curriculum design challenging.] Summary: AI is progressing at a rapid pace, with systems like DeepMind’s outperforming top international math students in elite competitions. Yet educational research and innovation remain sluggish, hindered by the need for long-term studies and delayed outcomes. Emma Brunskill discusses how recent advances in generative AI could be repurposed to simulate and optimize educational tools, such as using AI to judge and improve math worksheets. Despite promising early results in AI-human alignment on educational task evaluation, simulations still fall short in modeling student learning progression realistically. Mike Frank presents a complementary vision of using AI models trained on actual child experiences—collected through wearable cameras and observational studies—to understand cognitive development. His lab focuses on language input and developmental benchmarks to train AI systems that mimic human language learning. These models show early developmental similarities, particularly in how weaker models resemble four-year-olds and stronger models approach adult behavior. He emphasizes the importance of child-centered, open-access datasets that support scientific modeling across diverse global populations. Victor Lee expands the conversation to AI literacy, revealing its nebulous and inconsistent interpretations across student, teacher, and policymaker communities. From personalization to algorithmic bias, students' understanding of AI varies wildly, while teachers express diverse needs in professional development—from integrating AI tools to teaching core AI concepts. Lee argues for a broader, flexible definition of AI literacy that includes critical, user, and developer perspectives. With new AI education legislation emerging, he calls for a shared language to ensure clarity and relevance in curriculum design. All three speakers agree on the urgent need for nuanced discussions around scaling, implementation fidelity, and data ethics. They caution against decontextualized solutions and advocate for research that respects the real-world complexity of schools, teachers, and students. The session concludes with a call to prioritize skills like healthy skepticism, learning-to-learn capabilities, and information source awareness to prepare students for an AI-native future. Content: ## Introduction At the 2025 AI + Education Summit, a panel of three researchers—moderated by the chief technology officer of a large public school district—explored how artificial intelligence (AI) can accelerate learning, illuminate child development, and shape AI literacy. The discussion began with an overview of notable AI achievements, proceeded to case studies in cognitive science, and concluded with a framework for defining and teaching AI literacy. A subsequent question-and-answer segment addressed human skills, scaling challenges, data considerations, and essential competencies for an AI-infused future. ## Accelerating Educational Innovation with Generative AI ### AI Achievements in Competitive Mathematics Recent advances in generative AI have demonstrated capabilities once deemed out of reach. In mid-2024, DeepMind introduced an AI system that performs at approximately the 90th percentile in the International Mathematical Olympiad—a competition that convenes the world’s top high school mathematicians. This system now outperforms roughly 550 elite students on the most challenging problems. Until recently, large language models struggled even with basic arithmetic, making this rapid progress especially striking. ### Simulating Student Learning at Scale Educational innovation, by contrast, often unfolds over decades. To explore whether AI can speed up this process, researchers attempted to create simulated learners. Initial experiments used large language models (LLMs) to emulate static student performance on algebra problems; these LLMs proved surprisingly accurate at predicting a single post-lesson outcome but failed when simulating incremental learning—often leaping from not knowing linear equations to mastering calculus after one instruction. Shifting focus, the team then employed generative AI to simulate expert judgments rather than student cognition. One LLM generated multiple versions of a math worksheet, while another evaluated their expected effectiveness. Through an agentic workflow—with separate “critic” and “creator” agents—the system iteratively refined worksheets. To validate these AI judgments, researchers recruited 90 experienced educators to perform pairwise comparisons. The AI’s predicted post-test scores correlated strongly with human rankings, demonstrating promise as an ideation tool, though notable disagreements underscored the need for further refinement before full optimization. ## AI Models as Theoretical Tools for Child Development ### Capturing Naturalistic Language Input A second research strand uses AI to model child development. To train and evaluate these models on authentic data, investigators developed a lightweight, head-mounted camera rig—built from a GoPro Hero and customizable with familiar “Croc charms”—to record toddlers’ day-to-day interactions. Families across the United States contributed over 1,000 hours of video capturing home and preschool environments. Following rigorous consent procedures, these recordings will be shared openly with qualified researchers under strict privacy safeguards. ### Evaluating Learning Outcomes in AI and Children Researchers then train “mini‐GPT” language models on the BabyView video and audio data to examine emergent linguistic capabilities. Early findings show that these child-trained models acquire grammatical structures more readily than standard vision-language models learn to categorize real-world objects. To evaluate cognitive development, the team also designed tablet-based vocabulary assessments that present children with multiple-choice tasks ranging from simple to complex words. AI models are judged on both accuracy and error patterns, comparing their response distributions to those of children aged five to twelve. Building on these efforts, the Levante Project provides an open, international framework for longitudinal developmental assessments. Partnering with philanthropic foundations and research laboratories worldwide, the platform offers standardized tasks—such as inhibition control, Theory of Mind, matrix reasoning, and vocabulary—in multiple languages. Deidentified data from diverse populations are archived centrally and made immediately accessible, enabling both cognitive scientists and AI researchers to evaluate models against a broad spectrum of human development. ## Defining AI Literacy for All Stakeholders ### The Ambiguity of “AI Literacy” and Bias The panel’s third presentation addressed the challenge of defining AI literacy. The term occupies a conceptual space between casual consumer knowledge and professional developer expertise. For example, students often equate algorithmic bias with personalized recommendations or default settings in video platforms, or conflate it with statistical bias. Such divergent interpretations highlight the lack of a common vocabulary. ### Teacher Professional Development and Co-Design A recent survey of over 1,000 K–12 teachers revealed three primary professional development needs: (1) how to use AI as an instructional tool, (2) how to teach students about AI, and (3) basic understanding of AI concepts. In ongoing co-design workshops, educators collaborate with researchers to develop free, discipline-based curricula. Computational analysis of recorded sessions shows distinct emphases: some teachers prioritize practical tool integration and responsible use, while others focus on the mechanics of AI algorithms and career preparation. ### Existing Frameworks and a Proposed Meta-Framework Several AI literacy frameworks are already in circulation. One uses a fruit-themed decision tree to guide middle school understanding, while another high school standard enumerates reasoning tasks—classification, prediction, search, logical deduction, and statistical inference—and corresponding algorithmic approaches (e.g., neural networks versus symbolic search). To reconcile these diverse perspectives, the panel proposed a simple meta-framework comprising three roles: user, developer, and critic. Each role captures different knowledge and skill emphases, offering flexibility for curriculum designers and policymakers. ### Legislative Initiatives In California, Assembly Bill 2876 mandates the integration of AI literacy into future curriculum frameworks across subject areas, providing a legal impetus for defining core competencies. A corresponding bill is under consideration in the U.S. House of Representatives. The panel stressed the urgency of reaching consensus on essential AI literacies to guide equitable and responsible implementation. ## Q&A Highlights ### Unique Human Skills Beyond AI When asked which human capability AI cannot replicate, panelists cited **intentionality**—the purposeful drive to create or appreciate beauty—and **authentic affect**, such as genuine love. They noted that throughout history, many traits once thought uniquely human (e.g., translation, empathy) have eventually been matched or surpassed by machines in controlled settings. A cognitive science perspective further argued that human uniqueness emerges from complex, flexible interactions among emotion, social context, and embodied experience. ### Scaling Educational Innovations Panelists identified **implementation fidelity** as a key barrier to scaling AI-driven instructional tools. Pilot studies often ensure ideal conditions—dedicated time, technical support, on-task behavior—but real-world classrooms require attention to teacher training, district technology infrastructure, procurement processes, and policy frameworks. In scientific modeling, the challenge lies in moving from average performance to representing the **variability** of learners across home and classroom environments. ### Data Challenges and Opportunities Educators, parents, and technologists differ on data access. While open, deidentified data accelerate research, privacy concerns justify stricter controls on sensitive information. Panelists recommended **stakeholder-driven governance**: broad data should be readily available under minimal restrictions, whereas more sensitive records require formal agreements that protect participants while permitting valuable research. ### Preparing Students for an AI-Native World In rapid-fire closing, panelists recommended that schools prioritize: - **Healthy skepticism**: Developing critical awareness of AI outputs and limitations. - **Meta-learning**: Cultivating the ability to learn and adapt continuously in a fast-changing technological landscape. - **Source evaluation**: Teaching students to assess the credibility and provenance of information, whether human-generated or AI-produced. These competencies, panelists agreed, will equip the emerging generation to navigate and shape an increasingly AI-infused society.