AI Is Improving Student Performance. It Is Also Destroying Student Learning.
A new report explains the difference, and why it is the most important distinction in AI education.
The struggle is where the learning lives. Remove the struggle and you remove the learning. AI removes the struggle beautifully.
In this post I will:
Explain the ‘performance paradox’ identified in a major new report on cognitive offloading.
Show why AI fluency creates an illusion of competence that erodes critical thinking.
Argue that the distinction between beneficial and detrimental offloading is the concept every educator, parent, and manager needs to understand right now.
Student grades are going up. Student learning is going down.
Those two statements describe the same process, measured at different timescales. A new report from the University of Technology Sydney, authored by Professor Jason Lodge and Professor Leslie Loble, names this the ‘performance paradox’ and provides the most comprehensive evidence base I have seen for why unstructured AI use in education is causing serious, measurable harm.
The report draws on Cognitive Load Theory and a substantial body of recent research to make a deceptively simple argument. When students use AI to manage the tedious parts of a task (checking grammar, reformatting a reference list, brainstorming counter-arguments), it frees their working memory for the hard thinking: structuring an argument, evaluating evidence, synthesising sources. Lodge and Loble call this beneficial offloading. The extraneous load is removed. The intrinsic cognitive work remains.
When students use AI to bypass the hard thinking itself (“write me an essay about the causes of World War I”), the entire schema-construction process collapses. The generation of ideas, the retrieval of knowledge, the analysis of connections, the synthesis of an argument: all of it is outsourced. Lodge and Loble call this detrimental offloading. The student is no longer learning. They are completing a task.
The distinction sounds obvious. The evidence for its consequences is not.
The performance paradox in practice
In a large randomised experiment involving nearly 1,000 high school maths students, Bastani and colleagues (2025) found that access to an AI assistant appeared to improve problem-solving during the study period. But when the AI was removed, students’ independent performance collapsed. The scaffolded gains did not transfer into durable knowledge. The AI had boosted performance while undermining learning.
A quantitative study of 350 participants by Ejaz and colleagues (2025) found a significant negative correlation between frequent AI use and critical thinking skills. The same study found a strong positive correlation between higher cognitive load and better critical thinking, providing direct empirical support for the desirable difficulties framework. The lower cognitive load reported by frequent AI users was not beneficial. It was a signal that essential intrinsic processing had been bypassed.
The pattern is consistent across studies. Short-term task performance improves. Long-term learning is harmed. And the students most at risk are the ones who need the learning most.
Fluency is the trap
The report identifies a vicious cycle that should concern anyone working with AI, not only in education.
AI produces text that is coherent, confident, and articulate. Lodge and Loble describe this as ‘fluency on demand.’ This fluency serves as a misleading metacognitive cue. The ease of processing (how smooth the output reads) is mistaken for the depth of learning (how much the student actually understood). Research has long established this effect with other media; for example, people who watch a polished lecture video consistently overestimate how much they retained.
Fan and colleagues (2024) call the result ‘metacognitive laziness.’ The convenience of AI undermines engagement in the self-regulatory processes (planning, monitoring, revision) that are essential to learning. The student abdicates their metacognitive responsibilities to the tool. Not because they are lazy in any moral sense, but because AI makes the rational decision to outsource feel costless.
The cycle runs like this: the student seeks efficiency. The fluency of AI output creates an illusion of competence. The illusion triggers metacognitive laziness. The laziness leads to more outsourcing. The outsourcing erodes the student’s actual knowledge base, making them more dependent on the tool and less able to judge its output in future.
Zhang and Xu (2025) call this the paradox of self-efficacy: AI use increased students’ confidence while intensifying their dependence. Lee and colleagues (2025) found the same pattern in knowledge workers: greater confidence in the tool was directly linked to reduced critical thinking in the user.
The new equity gap
The report’s most sobering finding is about equity. The cognitive risks of detrimental offloading are not distributed equally. Students who already possess strong domain knowledge and metacognitive skills are better placed to use AI beneficially: they offload the right things and keep the hard work for themselves. Students without those skills are more susceptible to detrimental offloading. They fall prey to the illusion of competence and bypass the learning they need most.
Lodge and Loble call this the ‘Matthew Effect’ for an AI-mediated world (ordinarily expressed as the idea that the rich get richer and the poor get poorer). The students who can already think critically get better at it. The students who cannot fall further behind. Left unstructured, AI widens the gap.
What learning a language taught me about cognitive load
From 2010 to 2012, I lived in Japan as a scholar of the Daiwa Anglo-Japanese Foundation, investigating the relationship between science and theatre. I was also trying to learn Japanese. For months I made the same mistake: obsessively memorising grammar rules from textbooks, drilling conjugation tables, filling notebooks with characters I could reproduce but not use. My performance on written exercises was fine. My ability to hold a conversation was non-existent.
Real progress began when I stopped memorising and started speaking to people. Badly. Slowly. With constant errors. The friction of live conversation, the embarrassment of getting it wrong, the effort of retrieving a word under pressure rather than reading it from a page: that was where the learning lived.
Do not test me on my Japanese now. It feels like a lifetime ago. Which is itself the point. Skills atrophy without constant use and practice. The cognitive effort I put in during those conversations built something durable. The grammar tables did not.
Lodge and Loble’s report gives this a name. The textbook drilling managed extraneous load (formatting, presentation) while letting me avoid the intrinsic load (actually thinking in Japanese, producing language under pressure). The conversations forced the intrinsic work. The discomfort was the learning.
Why this matters beyond the classroom
This report is framed for K-12 education (i.e., schooling for children between the ages of 5 and 18). Its implications reach further. Every professional using AI to complete cognitive work is making the same choice Lodge and Loble describe: offloading extraneous load (formatting, first-pass data cleaning) or offloading intrinsic load (deciding what the data means, structuring the argument, evaluating whether the conclusion holds).
The question the report leaves you with is not whether AI is good or bad for learning. It is whether the person using it knows the difference between the parts of a task that can be safely outsourced and the parts where the struggle is the point.
Most do not. That is the gap this report names.
Go slow.
This post gives you the argument. The Slow AI Curriculum gives you twelve months of structured practice in knowing the difference: which cognitive acts to keep, which to delegate, and how to tell one from the other under pressure. Monthly live seminars, CPD accreditation, a community of 225+ educators, researchers, and professionals working through this together.
Another strong piece, Sam.
What lands most is the distinction between offloading what surrounds the learning and offloading the learning itself. That is the seam many institutions still have not named clearly enough.
The “performance paradox” is especially useful because it explains why so much current evidence feels contradictory: better output, weaker retention; smoother task completion, thinner understanding. And the equity point is the turn of the screw. Students with strong prior knowledge may use AI to accelerate. Students without it may be nudged into a form of false mastery.
Feels to me like this also reaches beyond education. More and more professional settings are now facing the same question: what can be safely delegated, and where is the struggle actually the point?
“Remove the struggle and you remove the learning.” That really is the whole thing. We had a morning with a tangled ball of yarn. I told my daughter, “There’s no way around it, you just have to sit with it and work on it until it’s untangled.”
I love learning about your time in Japan: “The conversations forced the intrinsic work. The discomfort was the learning.” I can only imagine how difficult it would be to learn a language like Japanese, which uses characters, and to have to make conversation.