I’ve been arguing for a while that AI in education is a pedagogy problem, not a tool problem. The tool gets the headlines. The pedagogy is where the costs and benefits actually show up. Liu, Christian, Dumbalska, Bakker, and Dubey (2026) just published a preprint that gives this argument new teeth. Their three randomized controlled trials, with 1,222 participants, provide what the authors describe as the first large-scale causal evidence that AI assistance impairs unassisted performance and reduces persistence.
What the Study Did
The authors ran three experiments. Experiment 1 had 354 participants solve fraction problems either with GPT-5 available in a sidebar or without. Experiment 2 replicated the design with 667 participants and tighter controls. Experiment 3 extended the same setup to SAT-style reading comprehension with 201 participants. The pattern held across all three.
In every experiment, AI-condition participants first worked with AI assistance, then had the AI removed and were tested on the same kinds of problems. The control condition did the same problems without AI throughout.
The Cognitive Cost of Quick Answers
The authors found two things, both striking.
First, when the AI was taken away, AI-condition participants performed worse on independent test problems than control participants who had never used AI.
Second, AI-condition participants were also more likely to give up. The skip rate, the percentage of problems they refused to attempt, was higher across all three studies. As Liu and colleagues put it, “People do not merely become worse at tasks, but they also stop trying” (p. 2).
The Experiment 2 sub-analysis carries the most consequential finding for teachers. The authors split AI users by self-reported usage: 61 percent used AI to get direct solutions, 27 percent used it for hints or clarifications, and the remaining 12 percent didn’t use it at all. The performance and persistence losses were concentrated entirely in the direct-solution group. Participants who used AI for hints performed comparably to controls.
So it’s not AI use that breaks persistence. It’s the type of AI use.
Why This Connects to Earlier Work
The finding aligns closely with Fan et al.’s (2025) work on metacognitive laziness, where students using AI showed reduced self-regulation in writing tasks. It also extends Kosmyna et al.’s (2025) “Your Brain on ChatGPT” study from EEG-based observation to behavioral causation. Liu and colleagues cite both. What they add is the causal claim. The earlier evidence was largely correlational or based on small samples. The new RCTs move that needle.
The authors also propose a mechanism. The argument is that “AI removes the productive struggle through which people develop not only accurate knowledge but accurate self-knowledge” (p. 10). When students offload the hard part, they don’t just lose the answer-finding skill. They also lose the metacognitive calibration that tells them what they can do on their own. That calibration is what sustains persistence under difficulty.
Where I Agree and Where I’d Push
I find this paper persuasive on the central claim. The triple replication across two cognitive domains, the brief exposure window, and the dose-response pattern in the usage-type analysis all hold up. Combined with Gerlich’s (2025) work on cognitive offloading and Bastani et al.’s (2025) classroom findings on unrestricted AI, the empirical base for AI-induced deskilling is starting to feel solid.
Where I’d push: the studies are short. A 10 to 15 minute lab session is a useful proof-of-concept, but it doesn’t tell us much about what sustained classroom AI use does over a semester or a year. The “boiling frog” hypothesis is suggestive, not yet measured. The next generation of these studies needs to be longitudinal and naturalistic.
I’d also flag that the design implications the authors draw are debatable. They argue that Socratic AI modes and reduced use time are “band-aids.” That may be true. But the sub-analysis in Experiment 2 told a different story. Students who used AI for hints performed similarly to controls. Only students who used it for direct answers showed the deficit. That’s not a band-aid finding. That’s evidence that how teachers structure AI use is exactly the variable worth working on.
Which brings me back to where I started. AI in education is a pedagogy problem. This is the strongest paper I’ve read this year making the case that the pedagogy is doing real cognitive work, whether we design for it or not.
References
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122 //
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544 // https://medkharbach.com/metacognitive-laziness-and-ai/
- Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006 //
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing tasks. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance [Preprint]. arXiv. https://arxiv.org/abs/2604.04721