LAUSD's High-Profile AI Edtech Bet Goes Bust

The architect of the failed LAUSD AI deal was Superintendent Alberto Carvalho, who championed the $6 million contract with the Boston-based startup AllHere. The project, an AI chatbot named "Ed," was publicly launched in March 2024 but the project collapsed within months and never fully launched. The FBI investigation is reportedly focused on Carvalho himself, not the school district, and centers on financial issues surrounding the contract. A key figure in the controversy is Debra Kerr, a Florida-based consultant with long-standing ties to Carvalho, who helped broker the deal. Kerr later filed a claim for $630,000 against the now-bankrupt AllHere for unpaid commissions related to the LAUSD contract. The FBI also raided a Florida property owned by Kerr in connection with the investigation. The company at the center of the deal, AllHere, had a background in automated text messaging to reduce student absenteeism, not sophisticated AI. Before its collapse, the firm's founder and CEO, Joanna Smith-Griffin, was indicted on charges of defrauding investors. A whistleblower from the company also alleged that AllHere took shortcuts with security and data privacy, potentially violating its contract with LAUSD. For adaptive learning systems to succeed where "Ed" failed, robust machine learning is critical. Techniques like Bayesian Knowledge Tracing and Deep Knowledge Tracing are used to model a student's understanding in real-time, predicting performance and identifying knowledge gaps to personalize instruction. These models analyze patterns in student responses to tailor the learning experience, a level of sophistication the LAUSD chatbot likely never achieved. Reinforcement learning (RL) offers a powerful mechanism for optimizing educational content delivery. By framing the problem as a multi-armed bandit, an RL agent can explore different teaching strategies or types of content ("arms") and learn to exploit the ones that yield the best student outcomes, maximizing engagement and learning over time. This data-driven approach adjusts dynamically to student interactions, a far cry from a simple Q&A chatbot. Speech recognition for early learners, a key feature for any K-3 reading tutor, presents unique challenges. Children's voices are higher-pitched and more variable, and systems trained primarily on adult speech often struggle, leading to high word error rates. Overcoming this requires fine-tuning models on diverse datasets of children's speech to ensure equity and accuracy, a crucial step for providing effective, real-time feedback on pronunciation and fluency. Successful AI reading tutors like Amira Learning and Plabook demonstrate the potential of this technology when properly implemented. These platforms use speech recognition to listen to children read aloud, providing interactive tutoring and assessing skills down to the phoneme level. Studies have shown such tools can significantly boost reading fluency and comprehension, offering a stark contrast to the LAUSD's failed project. Ultimately, designing AI for children necessitates a "safety-by-design" approach, with strong guardrails for younger learners. This means robust content filters, no emotional manipulation, and strict data privacy to protect children's data. For AI to be a scaffold for learning rather than a crutch, human judgment and age-appropriate design must always be central, a lesson underscored by the LAUSD's cautionary tale.