Every now and then, a research paper lands hard. Not necessarily because it says something entirely new. Perhaps more often, it matters because it gives shape to something classroom teachers have already been sensing – but been unable to clearly articulate – in the messy middle of their practice.
That’s my take on why David Strömberg, Victor Lei and Yanhui Wu’s The Generative AI Learning Penalty: Evidence from Chinese Secondary Education is important. It is a 2026 CEPR discussion paper based on 30 months of data from 26,811 Chinese students in Years 7–12. The authors examine the relationship between generative AI use, homework productivity, closed-book school exams, and high-stakes entrance examination results.
Their headline finding is stark. They argue that AI use increased homework scores and reduced homework completion time, but lowered exam performance.
That headline should make every secondary teacher, curriculum leader, and school administrator take pause. The data set was large. The time frame was large. The cultural context and the concept of ‘homework scores’ a bit different to my own. Yet, the paper is a ‘red flag’ – a flashing warning light.
We should take note. Not panic. Not respond with knee jerk reactions.
Pause. Think. Act.
Before we go any further, let’s sent the message straight. This paper does not buy into false binaries and tell us that AI is “bad”. It does not tell us to ban it. It does not tell us to retreat into nostalgic, pre-digital schooling.
It tells us a few very important things.
The first one is something that, no doubt, we suspect but haven’t quite interrogated enough:
When students use AI to bypass the cognitive effort through which learning is formed, they may look more productive while becoming less prepared.
That is a very different problem.
It echoes the echoes the findings of Natalya Kosminya et al (2025) at MIT about the dangers of cognitive debt.
It echoes the concept of the illusion of mastery explored by Lodge and Loble (2026).
In a way, it mixes the two ideas together and ties their findings to study and examination performance.
And for those of us teaching senior secondary students in Queensland – particularly Year 12 students preparing for QCAA external assessments – it raises some urgent questions about the type of homework that is expected of students, the ways in which they study, the formative assessments and checks for understanding that happen along the way within our teaching processes, students’ engagement in the productive struggle of learning, and other kinds of “at home” learning we still expect students to complete.
The productivity-learning split
The paper identifies a sharp divergence between task productivity and learning outcomes.
Students using AI completed homework more quickly. Their homework scores improved. On the surface, things looked better. ‘Homework scores’ rose by 18 per cent and completion time fell by 30 per cent. Yet monthly closed-book exam scores fell by 20 per cent within six months.
Without unpacking the specifics of the Chinese context, what matters here is that:
(1) The work appeared to have improved,
BUT
(2) The learning was weakened.
The authors point out that, in pre-AI classrooms, homework completion was never a perfect indication that learning was happening. Teachers probably have always known that students could copy answers, rush through work, rely too heavily on peers, or complete tasks mechanically. But generative AI changes the scale, speed, and invisibility of the problem.
A student can now produce accurate-looking homework (study notes and materials?) with limited cognitive friction (productive struggle) taking place.
Our students can generate explanations, summaries, practice responses, flashcards, plans, arguments, and even polished paragraphs without necessarily engaging in the mental work those products are supposed to represent. (See: The Climb, the Scoreboard, and Assessment Integrity (Part 1 – Some lessons from Miley Cyrus and Bill Walsh for AI-disrupted classrooms))
I often rant about the performativity of schools. I now have a new rant.
We need to adjust our pedagogy, our teaching practices, our assessment processes and expectations in ways that demand authentic and rich engagement by students. We need to be explicitly teaching students how to safeguard their learning in an age of AI. Without doing so, the easy convenience of AI for students may make study itself performative.
And, in systems where there are high-stakes closed book external examinations, the exam may be the moment where the truth is revealed.
I’m thinking of school leaders having a moment a little like when the crowd realises that, indeed, ‘the emperor has no clothes’. I suspect the authors of this article make well feel like the little boy who has seen the truth all along.
For school leaders and senior teachers, this matters because QCAA external assessments – including the Year 12 Modern History EA – remain closed-book, time-bound, individual, high-stakes demonstrations of student knowledge, skill, judgement, and fluency. Students, obviously, will have no access to AI in those examinations.
Any student who has outsourced the struggle may not discover the cost of taking the easy convenience of outsourcing struggle to AI until the day of their results come in.
That will be way too late.
Students are poor judges of their own learning
The authors also make the point that students are often not be reliable judges of whether they are learning. This is not because students are foolish. It is because learning often feels bad before it becomes good. Learning is challenging. Its effortful.
Coming up with ideas without access to notes and teacher provided scaffolds feels ‘exposed’ to risk.
A difficult source feels frustrating.
A first attempt at an essay hypothesis feels clumsy.
Attempting a response to tough a practice examination feels like a mountain to climb.
A student wrestling with a hard homework question may interpret that struggle as evidence that they are not learning, when in fact the struggle may be the very place where learning is happening.
Students need to know that thinking on your own is ‘scary’ and ‘risking being wrong’ is tough.
This is especially true if you have been over-scaffolded by those who have misapplied the ideas of Seller’s Cognitive Load Theory in the name of the ‘science of learning’. If you’ve constantly been supported by scaffolds and structures and not had them strategically and thoughtfully removed as expertise grows, learning is not optimised. (See my thoughts on the importance of guidance fading and the expertise reversal effect in The Strawman Problem: What the binaries of the EI vs Inquiry Debate Get Wrong).
AI in study can provide a constant easy tutor – a way out of the struggle.
We need to maintain an appropriate productive struggle, the challenge of learning.
The authors note that students who fall behind may not recognise that their use of generative AI is contributing to a decline in their learning. The deterioration, the believe, is gradual. The product of the study – the completed “homework” – still looks good. The task still gets completed. The ‘got all their homework right’ when it was corrected in class.
The student still feels productive.
But a closed-book exam tells a different story.
This is where generative AI becomes especially seductive. It removes the unpleasant feeling of not knowing. It gives the student fluency, speed, and completion. It makes the work feel smoother.
But smoother is not always better. In fact, the journey of learning being smooth may be the warning sign.
As teachers, we need to reinforce the importance of helping students understand the difference between:
- feeling like they are learning, and
- actually being able to retrieve, apply, explain, and judge under pressure.
That distinction now needs to be taught explicitly. In the AI age, metacognition is not an optional extra.
The slow accumulation of loss
The researchers found that regular exam results deteriorated within six months, while the full negative effect on high-stakes entrance exams took about two years to emerge. Kosmyna et al (2025) referred not to ‘cognitive debt’ alone but to the “accumulation of cognitive debt” – over time!
If students begin outsourcing the struggle of learning in their homework or home study in Year 9, for example, the real cost may emerge in Year 12.
If they begin outsourcing source analysis, argument construction, historical explanation, or essay planning in Year 10, the external assessment may become the place where the accumulated deficit finally becomes visible.
This is one reason schools may not yet be reacting strongly. The evidence of harm may be arriving slowly.
Teachers may see some concerning patterns. A student’s homework looks better than their classwork. A written paragraph seems oddly polished. A student can produce an answer at home but struggles under timed conditions. Their vocabulary is stronger on the page than in conversation.
But these are often interpreted as isolated issues. The Strömberg, Lei and Wu paper suggests we may need to read these patterns more carefully with a professional curiosity.
Curiousity growing from a position of academic care.
Homework outsourcing as the central mechanism
The authors identify a particularly important pattern. The learning losses were concentrated among roughly 80 per cent of AI users whose behaviour was consistent with homework outsourcing: unusually short homework completion time combined with high homework scores. By contrast, AI users who spent similar time on homework as non-users experienced small learning losses. That distinction is critical.
The issue is not simply “AI use”.
The issue is how AI is used.
There is a world of difference between a student who uses AI to:
- test their understanding,
- generate a counterargument,
- receive feedback on a draft,
- practise retrieval,
- compare explanations,
- improve a study plan,
- interrogate their own reasoning,
and a student who uses AI to:
- generate the answer,
- avoid reading,
- skip planning,
- bypass drafting,
- complete homework quickly,
- remove uncertainty,
- escape productive struggle.
There can be beneficial cognitive offloads and detrimental cognitive offloads. We need to maintain the friction of learning – the productive struggle.
This connects strongly with what I have been trying to articulate through the Bubble and Burner Model.
Students need to Think First.
They need to engage with some intellectual struggle – to encounter the problem – to be fully present to their work before AI enters the learning process. They need to engage their brains, generate some of their own ideas, come up with some questions, make an attempt, retrieve what they know, test an idea – experience the friction, struggle with the climb – before they draw upon AI as a relational presence that supports, questions, challenges, extends, or clarifies.
The Bubble and Burner Model frames the teacher as a regulator of cognitive friction in the classroom and argues that AI use should be sequenced according to learner readiness, task purpose, and the need to preserve effortful thinking. Students must be given the mindset and the skills to be that regulator of cognitive frictions in their own space when their teachers are not present – at home.
Which students are most affected?
The paper finds that the ‘learning penalty’ – the costs of accumulated cognitive debt to AI – were larger for younger students, high-achieving students, and boys. Again, this should give us pause.
The high-achieving student is not immune.
In fact, high-achieving students may be especially tempted by efficiency. They often carry heavy academic loads. They may be balancing multiple subjects, leadership responsibilities, co-curricular expectations, part-time work, family commitments, and the quiet pressure of ATAR calculation.
For some of these students, AI may appear to offer a rational bargain:
I already understand this.
I just need to get the homework done.
This task is not worth my time.
The AI can help me move faster.
Sometimes they may be right.
Not every task deserves deep engagement. Some homework is poorly designed. Some study tasks are low-value. Some assignments are compliance exercises dressed up as learning. We should be honest about that.
But students are not always good judges of what is worth the effort. They may underestimate the value of the tasks they are set. They may misread fluency as mastery. They may interpret ease as effective learning.
They may avoid the very struggle that would strengthen them.
This is where Lodge and Lobel’s distinction between beneficial and detrimental offloading becomes important. Students need to learn when offloading supports learning and when it corrodes it.
That judgement will not emerge by accident.
It must be taught.
Adaptation is possible – but not automatic
The final finding in the paper is cautiously hopeful. The researchers found suggestive evidence that learning losses reduced over time, with the estimated learning penalty from five months of AI exposure falling from around 25 per cent in early 2023 to around 16 per cent by June 2025. They suggest that students or teachers may be gradually adapting, while substantial barriers remain.
The story is not doom. Students and teachers are adapting but its not adaption by design. It’s so far been a reactive, haphazard, uneven response… This research paper suggests that AI education cannot be reduced to “hope”. It needs to be resourced – and initially timetabled to core staff who support students NOW.
Schools need to systematically respond to the warnings we are receiving.
Students will not automatically become wise users of AI simply because AI is available to them. Teachers will not automatically develop effective AI-infused pedagogy because the technology is in their hands, in the hands of their students, and in their classrooms. Schools will not automatically protect learning because policy documents are written.
We need pedagogical clarity. Such clarity might start with some reflections on what we value in education ( The Castlereagh Statement may be an excellent starting point for this), reflecting on what we already see of value in our established subject-specific pedagogies, and by reflecting on the three principles Alison Bedford, Petrea Redmond and I articulated in our work on generative AI in education. These are that teachers must ‘step up’ to the mark – and be supported and resourced by their leaders and systems – to…
- Teach students how to use GAI tools.
- Teach to promote discernment and critical thinking.
- Teach for the whole human.
These principles are explored in multiple posts in this site and are at the heart of my research. The Strömberg, Lei and Wu paper reinforces those principles.
We cannot take a ‘head in the sand’ approach to the changes upon us. We cannot ‘sleepwalk into the future’. Students need to know how learning works. They need teachers to help them. They need to understand why effort matters. They need to recognise when a tool is helping them think and when it is thinking instead of them.
Why schools may not be reacting quickly enough
One of the most interesting parts of the article is the authors’ discussion of why students, parents, teachers, and administrators may not respond quickly to the learning penalty. They suggest many ideas.
The aggregate effects emerge slowly.
Teachers usually see students in one subject. A decline that looks dramatic across all subjects may look less unusual within a single classroom.
Students and parents may not connect the decline to AI use because the deterioration is gradual.
And students may misinterpret the effort of active learning as a sign that they are learning less, not more.
That final point is deeply important.
Students often prefer the feeling of fluency. They like tasks that feel smooth. They like notes that look complete. They like answers that arrive quickly. They like the emotional relief of “done”.
But learning often feels less comfortable than performance. In a VUCA world, we cannot afford to let students believe that the absence of difficulty is the same as growth.
Productive struggle and the purpose of difficulty
The phrase productive struggle can be misused. It should not mean abandoning students to confusion. It should not mean glorifying stress. It should not mean making learning harder than it needs to be.
Productive struggle is purposeful friction. It is the stretch between what students can do now and what they are becoming able to do.
AI can support that stretch. But it can also remove it. That is why the teacher remains essential. The teacher adjusts the heat. The teacher knows when to scaffold and when to step back. The teacher knows when the student needs a model, a nudge, a question, a worked example, a peer conversation, a retrieval cue, or time to wrestle.
In the Bubble and Burner Model Model, this is not a minor teacher move. It is the heart of the work. The model positions the teacher as a regulator, calibrator, and ethical boundary-setter who mediates cognitive friction rather than simply enabling technological access. That may be one of the most important teacher roles in the AI age.
Not content deliverer. Not AI detector.
A regulator of meaningful effort.
Disrupted History 
You must be logged in to post a comment.