Teachers Want Co-Design for Classroom AI

Reinforcement learning is being adapted for adaptive education to allow for real-time content changes based on a learner's progress and performance. This approach moves beyond static course structures, with systems applying deep reinforcement learning to personalize the learning path and make interactions more effective for each student. A significant challenge in building AI tutors for young children is the performance of Automatic Speech Recognition (ASR). Error rates for ASR systems are substantially higher for children's speech compared to adults—in some cases by as much as 340% when using models trained on adult speech. Even with models trained specifically on children's speech, error rates can be over 170% greater than those for adult systems. The variability in children's growing vocal tracts and their developing language patterns make it difficult for standard ASR to perform accurately. To model a student's evolving knowledge, Deep Knowledge Tracing (DKT) uses neural network architectures like LSTMs and transformers. These models can identify complex, non-linear patterns in a student's response history to predict performance on future interactions without needing constant retraining. The field has evolved from earlier psychometric and Bayesian methods to now include graph neural networks and attention mechanisms for greater accuracy. Effective phonics instruction for early readers is systematic and explicit, starting with high-utility letter-sound correspondences and progressing to more complex relationships. Research supports beginning with a few high-frequency letters and staggering those that are visually or auditorily similar, like 'b' and 'd'. A multi-sensory approach that combines seeing, hearing, and saying the sounds helps to reinforce these connections. For content recommendation within an adaptive tutor, multi-armed bandit (MAB) algorithms offer a framework for balancing the exploration of new content with the exploitation of proven material. Algorithms like Thompson sampling and Upper Confidence Bound (UCB1) can dynamically adjust recommendations to optimize for engagement and learning outcomes. This approach treats each piece of content as an "arm" and aims to maximize the cumulative "reward," such as correct answers or completed lessons, over time. Protecting children's privacy is paramount and legally mandated through regulations like the Children's Online Privacy Protection Act (COPPA) and the Family Educational Rights and Privacy Act (FERPA). This requires obtaining verifiable parental consent before collecting data from children under 13 and implementing strict security measures like end-to-end encryption and data minimization. Companies must be transparent about when AI is being used and how it makes decisions. A study involving over 10,000 adolescents found a strong link between reading for pleasure at an early age and better performance on cognitive tests measuring verbal learning, memory, and speech development. These individuals also exhibited better mental well-being, with fewer signs of stress and depression. Brain scans revealed that early readers tended to have larger total brain areas, particularly in regions critical for cognitive functions. The career path for a senior individual contributor (IC) in engineering focuses on technical leadership rather than people management. This involves setting technical direction, mentoring other engineers, and being accountable for the success of technical initiatives. The emphasis shifts from personal output to multiplying the impact of the entire team through technical leverage and enabling others to do their best work.