Study finds AI boosted homework scores 18%—then tanked exam results 20%

The investable signal is not that generative AI hurts learning; it is that consumer AI products in education can monetize convenience while degrading the outcome users actually pay for. That creates a product-design split: tools that generate answers become low-retention utilities, while tools that enforce practice, assessment, and progress tracking gain pricing power. In public markets, that favors outcome-oriented platforms and hurts any edtech model whose core value proposition is offloading work rather than improving mastery.

The near-term catalyst is policy and procurement, not earnings. Over the next 1-3 months, school districts, parents, and test-prep buyers are more likely to react to score-decay headlines by tightening AI rules or shifting budget toward supervised instruction; over 6-18 months, that can widen the valuation gap between guided-learning names and homework-assist names. The main falsifier is evidence that AI is being embedded as a scaffold inside curricula and lifts standardized scores, which would shift the narrative from substitution to augmentation.

The contrarian miss is that this is probably not a broad anti-AI trade. The more likely equilibrium is guardrailed AI with human-in-the-loop design, which would concentrate value in incumbents that can blend compliance and pedagogy. That argues for relative rather than outright bearish exposure: avoid extrapolating a niche learning failure into a blanket short on the AI complex.