The AI-in-education conversation has been stuck on cheating for two years now. Detection tools, integrity policies, disclosure forms. These have dominated the framing. A new report from Lodge and Loble (2026) at the University of Technology Sydney argues we’ve been looking at the wrong risk. The deeper problem isn’t that students might cheat on an essay. It’s that AI may interfere with the cognitive processes that build the knowledge and critical thinking education is supposed to produce.
The report runs 35 pages and grounds itself in cognitive load theory, self-regulated learning research, and the recent empirical work on AI and learning. It synthesizes the field carefully. What makes it valuable is the framework it offers teachers and policy makers for thinking about AI in schools without falling into either the booster or doomster camp.
The Cognitive Risk of AI in K-12
The central argument: “the true educational risk of AI is not simply that students will use it to cheat on an essay. The far more profound risk is that AI may fundamentally interfere with the cognitive processes of knowledge construction and verification, the very processes that build the long term memory stores and subsequent skills upon which the majority of critical thinking depends” (p. 12).
The argument reorients the entire discussion. The cheating question is about academic integrity. The cognitive question is about whether the next generation actually develops the capacity for independent thinking. Those are different problems, and the report makes the case that we’ve been focused on the smaller one.
The conceptual move that does the most work is a two-part separation. Beneficial offloading happens when a student uses AI to manage extraneous load (grammar checks, syntax cleanup, finding examples), which frees working memory for the harder work. Detrimental offloading, which the authors call outsourcing, happens when a student asks AI to do the intrinsic cognitive work itself. The first supports learning. The second short-circuits it.
The Performance Paradox
The report’s empirical centerpiece is what it calls the “performance paradox.” Students using AI often look like they’re learning. Their immediate task performance goes up. Their durable, long-term learning goes down. The clearest evidence comes from Bastani et al.’s (2025) study of nearly 1,000 high school math students, which I’ve covered elsewhere on the blog. The students with AI access solved problems well during the work, but when the AI was removed, their learning gains evaporated. The scaffolded performance didn’t translate into durable knowledge.
This is the part of the report I find most useful. It explains why the systematic reviews showing positive effects of AI on learning and the studies showing harm don’t actually contradict each other. They’re measuring different things. The positive studies measure performance with AI present. The negative studies measure learning after AI is removed. Lodge and Loble argue that the durable learning is what schools actually need to deliver.
Metacognitive Laziness Compounds the Problem
The second mechanism the report emphasizes is what Fan et al. (2024) call metacognitive laziness. Self-regulated learning, the planning and monitoring and revision that drives deep learning, itself takes cognitive effort. Students seeking efficiency will offload that effort to AI when given the chance. With sustained use, the metacognitive muscles atrophy. I’ve covered Fan et al.’s work in a previous post on AI and self-regulated learning, and the Lodge and Loble report places it in the broader framework cleanly.
The compounding factor is what they call the illusion of competence. AI produces fluent, confident output. That fluency feels like learning. The brain mistakes ease of processing for depth of understanding. Combine fluency, efficiency-seeking, and the illusion of competence, and you get what the report describes as a vicious cycle that erodes the student’s actual knowledge base.
My Take
The report’s strongest move is naming what it calls a “metacognitive equity gap.” The students best positioned to benefit from AI are those who already have strong domain knowledge and self-regulation skills. The students most vulnerable to detrimental offloading are those already experiencing disadvantage. Unstructured AI use widens existing equity gaps. The report calls this the Matthew effect with AI: the cognitively rich get richer, the struggling fall further behind.
This is the line I’d press hardest on if I were a policy maker reading this report. Not because the analysis is wrong, but because the obvious response (more structured AI use in schools) requires resources, teacher capacity, and curriculum reform that disadvantaged schools tend to lack. The equity argument cuts both ways.
I leave you with this interesting quote: “the educational imperative is not to protect students from a world where cognitive (and metacognitive) offloading is the norm, but to prepare them for it” (p. 29).
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
- Lodge, J. M., & Loble, L. (2026). Artificial intelligence, cognitive offloading and implications for education. University of Technology Sydney. https://doi.org/10.71741/4pyxmbnjaq.31302475