Wharton Study: AI Help Can Undermine 'Productive Struggle'
Research from Wharton cautions that on-demand AI assistance can erode a student's “productive struggle,” the effort required to cement long-term learning. The study found that learners who received immediate AI-generated answers were less likely to persist through difficult tasks, resulting in weaker mastery. The findings suggest AI tutors should be calibrated to provide hints and scaffolding rather than direct answers, allowing space for independent problem-solving.
- The Wharton study, co-authored by Hamsa Bastani, found that high school students using a basic GPT-4 model for math practice performed 48% better during the sessions but 17% worse on a later exam compared to a control group. A second group using a "GPT Tutor" with built-in safeguards performed 127% better during practice and scored about the same as the control group on the exam. - Reinforcement learning (RL) frameworks can create adaptive learning paths by treating educational content modules as "actions" and student performance as "rewards," optimizing the sequence for individual learners. Systems can use a student's rating on a task (e.g., a score from 0-10) to determine the difficulty of the next exercise, providing lower-level problems for scores of 3-5 and higher-level ones for scores of 9-10. - Knowledge Tracing is a critical component of AI tutors, using models like Bayesian Knowledge Tracing (BKT) or Deep Knowledge Tracing to model a student's understanding of concepts in real-time. BKT uses four main parameters: the initial probability of knowing a skill, the probability of learning it with practice, the chance of making a mistake on a known skill (slip), and the chance of guessing correctly on an unknown skill. - For early literacy in K-3, speech recognition technology is a viable tool for providing assessments and feedback on pronunciation and reading fluency. However, developing Automatic Speech Recognition (ASR) for young children is challenging due to developing articulation and physiology; one hybrid ASR engine for preschooler speech reported a word-error rate of 40%. - Multi-armed bandit (MAB) algorithms can be used to personalize content sequencing by balancing exploration (trying new content) and exploitation (using content known to be effective). A contextual MAB approach can use a student's prior knowledge state as the "context" to select the next best learning action, such as a video or a practice question. - Designing UX for young children requires simplified navigation with minimal menus, large buttons (60x60 to 80x80 pixels), and visible, recognizable icons to reduce cognitive load. For emerging readers, large fonts (24pt), high contrast, and audio feedback for actions are crucial for accessibility and engagement. - To ensure AI safety for young learners, it's critical to use tools compliant with regulations like COPPA, which limits data collection for children under 13. Educators should establish clear rules, such as allowing AI for brainstorming but not for final answers, and teach children not to input personally identifiable information. - Wharton professor Ethan Mollick, a notable researcher in this area, advocates for a "co-intelligence" model where AI acts as a partner to augment human capabilities. He suggests that while AI tutoring shows immense potential for personalized instruction, it cannot fully replace the broader social and practical functions of a classroom environment.
