With or Without

AI Assistance Reduces Persistence and Hurts Independent Performance by Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A Bakker, Rachit Dubey
Preprint, April 7, 2026

Even for readers overwhelmed by the research and thinkpieces on AI, this recent research article offers a nuanced, pragmatic approach to thinking about AI. Grounded in three fascinating experiments, this research provides causal evidence for two key consequences of AI use for learning tasks: reduced persistence and impairment of unassisted performance. In other words, the studies showed that the use of a “short-sighted collaborator” like AI led participants to stop trying, to lose motivation, and to perform less competently once AI assistance was removed. While the research supports the view that AI can be an impediment to durable learning, the scholars don’t give in to despair; rather, they use their findings to prompt programmers and system designers to “[rethink] how AI systems are built to collaborate with humans” by “[optimizing] for long-term human capability and autonomy.” After describing their research, the scholars explore some existing conceptual frameworks, including “cognitive offloading,” and conclude that AI systems “represent a new kind of cognitive scaffold.” They then explore the unique affordances of this technology – including shifting users’ “reference point for how long a task should take” and the erosion of “productive struggle” – and what those affordances mean for system design. It’s this analysis that educators and school leaders need to help students develop the cognitive structures necessary for averting AI’s more pernicious effects. The scholars finish with a flourish, a hopeful call for “the field to think about optimizing not just what people can do with AI, but what they can do without it.”

‍