Stanford CS547 HCI Seminar | Winter 2026 | Does GenAI Work in Education?

Key Concepts

• AI in Education is Complex: Generative AI’s impact on learning is inconsistent, necessitating a shift from if it works to how to make it work effectively.
• Knowledge Engineering is Paramount: Deep understanding of the cognitive requirements of a task is crucial for building effective AI-assisted learning tools.
• Cognitive Fidelity Matters: AI tutors should accurately model human cognitive processes and anticipate common misconceptions.
• Cognitively Aligned Interfaces are Key: AI should support, not replace, cognitive effort, offering assistance at an optimal level between full automation and complete manual work.
• Generalizable Process, Not System: While specific AI systems may not transfer across domains, the underlying knowledge engineering and interface design process is broadly applicable.

Challenges & the Current Landscape of AI in Education

Current educational environments face significant challenges, including limited personalized support in large classes (e.g., University of Michigan courses exceeding 2,000 students), a widening skills gap (the World Economic Forum estimates 6/10 workers will need training by 2027, yet only 50% will have access), declining academic performance (as reported by the National Center for Education Statistics), and inadequate preparation of medical professionals (92% of general surgery residents feel unprepared for independent practice). Despite the growing adoption of generative AI by students (70-85% using it for homework), research on its effectiveness yields mixed results. Studies from Harvard (AI tutor outperforming in-class learning in physics) and Stanford (AI suggestions improving human tutor effectiveness in K12 math) show potential benefits, while research from UPenn (AI access hindering learning when removed) and MIT (AI essay writing reducing brain activity) demonstrate negative effects. AI-generated hints also suffer from quality issues, with 30-50% exhibiting problems.

The Shift to Knowledge Engineering & Cognitive Tutors

The research focuses on moving beyond simply asking if AI works in education to understanding how to make it work effectively. This necessitates a focus on knowledge engineering – concerted efforts to understand the cognitive requirements of a task – to achieve “higher cognitive fidelity” in AI tutoring, feedback, and practice. This approach is rooted in “Bloom’s two sigma challenge,” which demonstrates that one-on-one tutoring is two standard deviations more effective than traditional classroom instruction. Early examples of AI-powered personalized tutors, known as cognitive tutors, have shown two times better learning outcomes than traditional algebra curricula. These tutors rely on cognitive models composed of production rules – “if-then” statements – that mimic expert thought processes. Cognitive fidelity refers to the degree to which an AI tutor’s model accurately reflects the cognitive processes involved in a task, including anticipating common misconceptions (e.g., providing different feedback for answers of 12, 24, or 10 when finding the least common denominator).

Case Studies & System Development

The research involved developing and testing AI-assisted learning systems in various contexts, including an Econ 101 course at the University of Michigan, teaching hospitals, and randomized trials in Harvard physics and Stanford K12 math classrooms. Specific systems developed include the “Feedback Writer” (used in the University of Michigan study, resulting in a 0.5 effect size improvement in revised draft quality – equivalent to moving a student from the 50th to the 70th percentile) and the “Note Copilot” (with three versions: Automated AI, Intermediate AI, and Minimal AI, the Intermediate AI version yielding the highest quiz scores). The development of the Feedback Writer involved iterative rubric refinement with instructors, AI-generated feedback based on rubric criteria, and TA review/modification.

Generalizability & Interface Design Principles

The research argues that while specific AI systems may not be directly transferable across different domains, the underlying knowledge engineering process is broadly applicable. A key finding is the correlation between the quality of rubric items – specifically, their targetness and concreteness – and the quality of AI-generated feedback. The more detailed and unambiguous the rubric items, the higher the quality of the feedback. Effective AI assistance also requires cognitive align interface design, positioning AI as an intermediate step between full automation and complete manual effort. This involves supporting core cognitive processes, such as encoding (initial comprehension and processing of information) and storage (retention of information), while maintaining “necessary cognitive engagement” for interpretation. AI-generated summaries, for example, should facilitate more efficient interpretation, not replace it. Identifying these core cognitive processes and designing interfaces to specifically support them is a generalizable approach.

Conclusion

The research underscores that successful integration of AI into education requires a deliberate and cognitively informed approach. Simply deploying AI tools is insufficient; a robust knowledge engineering process, focused on understanding and modeling human cognitive processes, is essential. Furthermore, designing interfaces that align with these processes and provide assistance at an optimal level – supporting rather than supplanting cognitive effort – is crucial for maximizing learning outcomes. The focus should shift from asking if AI works to how to make it work effectively, leveraging the power of AI to enhance, not replace, the human learning experience.