The "Cognitive Offloading" Paradox
New research shows that offloading learning tasks to AI can improve - rather than erode - human thinking and learning
Hey folks 👋
If you work in L&D right now, there’s a good chance you’ve heard people asking the question: what happens when learners lean on AI too much?
It’s a fair worry: the weight of evidence from the last eighteen months has pointed in one direction: cognitive offloading — letting AI do the mental work so you don’t have to — appears to erode critical thinking, reduce engagement, and weaken retention. The message has been consistent and increasingly loud: limit AI use, or pay the price.
However, a new study just complicated that picture significantly — with important implications for how we design AI-supported learning.
Let’s dive in.
The Case Against “Cognitive Offloading”
Before I get to the new research findings, it’s worth understanding just how strong the existing evidence is — because this isn’t a straw man I’m about to knock down. This is a real and well-documented concern.
Over the last 18 months, a series of studies have converged on the same finding: the more people use AI in the learning process, the less sharp their thinking becomes.
A 2025 survey of 666 participants across age groups and educational backgrounds found a significant negative correlation between frequent AI tool usage and critical thinking abilities, with cognitive offloading as the mediating mechanism. Younger participants were hit hardest, demonstrating higher rates of AI dependence and lower critical thinking scores (Gerlich, 2025).
Then came the brain scans. An MIT study used EEG to measure neural activity while 54 participants wrote essays using ChatGPT, Google Search, or no tools at all. The ChatGPT group showed the weakest neural engagement. When they were later asked to write without AI, they showed reduced brain connectivity and couldn’t recall their own earlier work. The researchers called it “cognitive debt” (Kosmyna et al., 2025).
The theoretical picture sharpened in 2025–26. Favero et al. (2025) warned that cognitive offloading undermines learning outcomes unless the mental effort that’s freed up gets redirected towards other meaningful tasks.
Then, in March 2026, Lodge & Loble went a step further, arguing that cognitive offloading isn’t inherently harmful to learners — what matters is whether it’s beneficial or detrimental, and the difference depends entirely on what happens with the freed-up cognitive capacity.
So over the course of 2025 and into 2026, the field was starting to move beyond “AI is bad for learning” toward a harder question: when is it bad, and when might it actually help? But the empirical evidence to answer that question — across a large sample, across cultures, with a clear mechanism — didn’t exist yet.
This is exactly what makes new research by Wang & Zhang (2026) so important.
The Rise of the “Offloading Paradox”
In March 2026, the International Journal of Educational Technology in Higher Education published a study that went beyond the question “does offloading hurt?” and asked a harder one: when students form genuine partnerships with AI — treating it as an intellectual collaborator rather than a passive tool — what actually happens to the way they think and learn? Specifically, do two cognitive responses — critical evaluation of AI outputs (what the researchers call cognitive vigilance) and strategic delegation to AI (cognitive offloading) — compete with each other, or can they coexist?
Based on previous research, Wang and Zhang hypothesised that cognitive offloading would hurt transformative learning. They expected the familiar story: delegation reduces cognitive struggle, struggle is where learning happens, therefore delegation undermines learning.
The study — 912 students across China, Europe, and the United States, using a three-wave time-lagged survey design that measured partnership orientation first, cognitive strategies two weeks later, and learning outcomes two weeks after that — found something more interesting than a simple reversal.
When students scored higher on partnership orientation with AI, two cognitive responses activated simultaneously. A partnership orientation predicted both increased vigilance (β=0.335, p<0.001) and increased offloading (β=0.351, p<0.001). In turn, both vigilance (β=0.437, p<0.001) and offloading (β=0.333, p<0.001) independently predicted what the researchers measured as transformative learning — the extent to which students questioned long-held assumptions, fundamentally changed how they understood their subjects, and re-evaluated the way they think. Not surface-level satisfaction. Deep shifts in perspective and understanding. The pattern was structurally consistent across all three regions, with some variation in path strength (Wang & Zhang, 2026).
The paradox isn’t just that offloading helped. It’s that the same partnership orientation that made students delegate more also made them more critical — and both made them learn more deeply.
But the relationship between offloading and learning depth wasn’t linear. A post-hoc analysis revealed a U-shaped curve (β-quadratic=0.102, p<0.001). Think of it as three zones:
Zone 1 — No offloading. The learner does everything manually. AI isn’t part of the process. They carry the full cognitive load: reading every source, writing every draft, organising every dataset. Learning happens, but it’s slow and capacity-constrained. There’s no freed-up bandwidth for higher-order reflection because every minute is spent on execution.
Zone 2 — Scattered, half-hearted offloading. The learner uses AI for a bit here and there — fixing a sentence, checking a fact, tidying a paragraph. This is where most current AI use in learning sits, and it’s the worst zone. The learner is still carrying almost all of the cognitive load, but now they’ve added the friction of managing the AI: deciding what to ask, evaluating whether the output is useful, switching between their own work and the tool. More effort, no meaningful benefit. This is what the negative studies measured.
Zone 3 — Committed, strategic offloading. The learner delegates entire categories of substantive work to AI: all the source summarisation, the full first-pass literature review, the complete data organisation. The cognitive savings are large enough to genuinely free capacity — and that freed capacity gets invested in the work AI can’t do: critiquing frameworks, questioning assumptions, constructing original arguments, making judgement calls. This is where the paradox kicks in. This is where transformative learning lives.
The “U-curve” identified by Wang & Zhang suggests that by adding overhead without freeing capacity, Zone 2 is worse for thinking and learning than Zone 1. Meanwhile, Zone 3 is dramatically better than both because delegating entire categories of substantive work frees enough cognitive capacity for learners to invest in what AI can't do — questioning assumptions, critiquing frameworks, constructing original arguments.
Put another way, in Zone 3, learners cross a threshold into positive ‘cognitive reallocation’. As one student in the study put it: “You have to offload significant, meaningful tasks to get a real learning benefit.”
The researchers call this the offloading paradox. Offloading can lead to better quality thinking and deeper learning — but only under two conditions:
Learners delegate enough to AI to genuinely free cognitive capacity, and
That freed capacity gets deliberately invested in higher-order work — questioning assumptions, critiquing frameworks, constructing original arguments (the kind of thinking AI can’t do for you).
Designing those two conditions isn’t a technology challenge — it’s a new and complex learning design challenge.
Designing for Cognitive Offloading
From the research, six principles emerge for how we should — and shouldn’t — encourage learners to use AI during the learning process:
Principle 1: Offload to AI substantially, or not at all.
Scattered, small AI assists produce the worst learning outcomes — full cognitive load plus coordination overhead, no freed capacity. Entire categories of substantive work delegated completely produce the best outcomes — genuine cognitive space for higher-order reflection. Half-measures are worse than no AI at all. (Wang & Zhang, 2026).
Principle 2: Frame AI as a partner, not a tool.
When learners treat AI as an intellectual collaborator, two productive cognitive pathways activate simultaneously: they become more critical of AI outputs AND they delegate more strategically. When they treat AI as a passive tool, neither pathway activates. The framing determines whether offloading is productive or mindless. (Wang & Zhang, 2026)
Principle 3: Build verification into the workflow, not the preamble.
Critical evaluation of AI outputs is the single strongest predictor of deep learning and the most underdeveloped skill among current AI users. “Don’t trust AI” is a warning. “Check one claim before you use this” is a habit. Design the second, not the first. (Wang & Zhang, 2026)
Principle 4: Make the learner think first, AI second.
Producing an answer — even a wrong one — builds stronger retention than receiving a correct one passively. When AI generates first, the learner becomes an editor. When the learner generates first, they’ve already done the effortful retrieval that builds durable knowledge. Never reverse the sequence. (Slamecka & Graf, 1978; Karpicke & Blunt, 2011)
Principle 5: Use AI to identify errors, not fix them.
The productive struggle of correcting your own mistakes is what builds the skill. When AI corrects for you, it removes the effort that produces learning. When AI flags that an error exists without supplying the fix, the learner must retrieve the correct form themselves. The correction is the learning opportunity — don’t let AI do it. (Bjork, 1994)
Principle 6: Assess without the scaffolding.
Without at least one unassisted assessment at the same difficulty level, you’re measuring the human-AI system, not the learner. The gap between AI-assisted performance and independent capability is significant and well-documented. If scores hold when AI is switched off, you built competence. If they drop, you built dependency. (Xu et al., 2026 — g=0.751 assisted vs g=0.369 independent, across 35 studies)
The Cheat Sheet
Here’s the design cheat sheet — a practical summary of what to offload to AI and what to protect to optimise for learning gains:
Concluding Thoughts: AI, Thinking & Learning
For the last eighteen months, the dominant story in our field has been simple: AI makes learners lazy. The evidence pointed one way, the headlines reinforced it, and the instinct to restrict, limit, and worry felt justified.
That story was never wrong, but it was incomplete because it only measured one condition: what happens when learners use AI without design. With no guidance on what to delegate to AI and no structure for what to do with the capacity freed by AI, offloading erodes thinking and learning.
Alongside the world of Lodge, Loble, and Favero before them, what Wang and Zhang’s recent findings add is the other half of the picture. When AI use is designed intentionally, cognitive offloading doesn’t replace thinking and learning: it creates more room for it.
Students who delegated the most to AI didn’t learn less — they questioned their assumptions more deeply, shifted their perspectives more fundamentally, and re-evaluated how they think. The mechanism isn’t mysterious: free up cognitive capacity on the low-order work, and humans will invest it in higher-order reflection. But only if someone structures that investment.
The emerging picture isn’t “AI is good for learning” any more than it was ever “AI is bad for learning.” The key message emerging from the research is that AI is a design variable — and like every design variable, its impact depends entirely on the decisions made by the person who shapes the interaction.
Happy innovating!
Phil 👋
PS: Want to keep up to date with the latest research on how humans learn? Check out my monthly Learning Research Digest.
PPS: If you want to learn more about how AI impacts our work, join me and a community of people like you on the AI Bootcamp for L&D.
