ep. 89. Cognitive Offloading to AI: The Peril and the Promise
6 min read
Anthropic’s recent 81,000-person study has been getting a lot of buzz, from well-founded methodological critiques to its heartbreaking quotes on the realities of AI making bosses richer, while employees face more work.
A finding that is less widely discussed is that educators were 2.5 to 3 times more likely than average to report witnessing cognitive atrophy firsthand.
Cognitive atrophy, per Anthropic’s study: “Over-reliance causing skill loss, intellectual passivity, students bypassing learning, [and] critical thinking decline.”
Even assuming methodological limitations, this is still a noteworthy finding. That’s because it’s another data point in a growing body of literature on how AI is negatively impacting cognitive capabilities.
One well-cited example is a 2025 study by Michael Gerlich, which found that people who used AI tools more frequently scored lower on critical thinking assessments. This effect was more pronounced in younger participants.
I know what you’re thinking. That was 2025, lifetimes ago in genAI land. Andrej Karpathy hadn’t even coined the term vibe coding yet when that study was published.
What do we know about AI’s effects on cognition in 2026, and how should we interpret the finding from the Anthropic study?
The Peril and the Promise
A March 2026 report from the University of Technology Sydney (UTS) synthesizes the latest available evidence on AI and cognitive offloading.
Wait, what? how did we get from cognitive atrophy to cognitive offloading?!
Cognitive offloading is the use of an external resource to reduce the mental demands of a task. You do it when you write a grocery list or use a calculator. It’s not inherently good or bad.
It becomes a problem when you offload the cognitive work that would have built or maintained a skill, and the skill degrades as a result. That’s cognitive atrophy.
Caveat alert: The UTS report focuses on the failure to build capability in the first place rather than the loss of existing capability, but the core offloading mechanism is the same.
The report distinguishes between beneficial and detrimental offloading:
Beneficial offloading: Frees up mental resources for higher-order work (aka managing extraneous cognitive load for my cognitive load theory fans). Think using AI to handle grammar so you can focus on the argument.
Detrimental offloading: Bypasses thinking itself. It eliminates the desirable difficulty that leads to building long-term knowledge schemas. This looks like taking and running with your first vibe coded prototype output you generate without assessing the UX quality.
The Peril
When you hear about AI eroding critical thinking, you’re hearing about detrimental offloading. Importantly, this happens during unstructured AI use.
To illustrate, consider this randomized study of nearly 1,000 high school students where students were assigned to one of three groups:
A standard GPT-4 chat interface with no constraints.
A GPT Tutor that included guardrails, with hints instead of answers, teacher-designed prompts, and instructions not to reveal solutions.
A control group with no AI access.
Both AI groups performed better during practice. But when access was removed, students who used the unguarded interface performed 17% worse than the control group. The guardrailed tutor eliminated this penalty. The unguarded tool had become a crutch, not a scaffold.
A separate study found a similar pattern in a larger study of AI-assisted peer feedback. In a randomized controlled experiment with 1,625 students across 10 courses, students received AI prompts that flagged low-quality feedback and encouraged revision.
After four weeks, the AI assistance was removed for one group. Those students immediately reverted. Their feedback became shorter, more generic, more repetitive, and less relevant to the material under review. They had relied on the prompts rather than internalized the standards.
A second group that received self-monitoring checklists in place of AI performed better than the no-support group, but still fell short of those who kept AI assistance. The tool had substituted for judgment rather than built it.
This pattern has a name: the performance paradox. AI improves output on the immediate task, but the underlying capability that would have developed through doing the work does not.
Output up, capability down.
Performance paradox: When AI boosts a student’s performance on an immediate task while simultaneously diminishing durable learning. - From Lodge & Loble (2026)
Making matters worse is our tendency to accept AI-generated output by default. An LLM’s responses are like listening to someone who doesn’t know what they’re talking about speak with confidence. The fluent delivery makes you believe it has to be right.
Researchers have called this metacognitive laziness: the more fluent the AI output, the more you’re inclined to accept it without critically engaging. Over time, you become less able to judge whether the output is actually good.
To be clear, this isn’t your fault. It’s not some failure of cognitive willpower. It’s a consequence of how these tools are designed by default, and how they’re integrated into workflows.
The Promise
Detrimental offloading isn’t inevitable.
Research consistently points to one principle: the impact of AI on learning depends on whether the interaction is structured or unstructured. When students use AI without pedagogical scaffolding, they trend toward outsourcing thinking to AI. When the interaction is deliberately designed, outcomes improve. Three design choices can shape the outcome.
Structuring what gets offloaded.
The first design lever is making the boundary between productive and harmful offloading explicit.
A 2025 study tested this with 240 university students learning English essay writing. One group was taught to delegate “lower-order” writing tasks to AI (brainstorming, grammar, co-revision) while keeping higher-order work (analysis, evaluation, reflection) for themselves. The other group received traditional instruction.
Over 12 weeks, the AI group showed significantly greater critical thinking gains. Importantly, the intervention didn’t restrict AI use. It made the distinction between what to offload and what to protect explicit, and built it into the workflow.
Structuring when the human thinks.
The second design lever is including a metacognitive pause into the interaction itself.
Several studies cited in the UTS report found that integrating non-optional prompts into AI environments, questions that required users to predict, reflect, or assess their understanding before receiving or acting on AI output, led to deeper engagement and improved self-regulated learning. The critical word is non-optional.
One paper found that adding self-reflection prompts alongside AI assistance didn’t improve outcomes, because the convenience of the AI overshadowed them. The metacognitive pause has to be structural, not suggested.
Structuring what role AI plays.
The third design lever is reframing what AI does in the interaction.
The default prompt bar positions AI as an answer oracle, which invites passive outsourcing.
Researchers have tested alternatives: AI that feigns confusion and asks clarifying questions (the “cognitive mirror”), AI that poses Socratic retrieval questions rather than providing answers, and AI that requires the user to evaluate and correct its output before proceeding. Each reframes the human as the active agent. Each produced better learning outcomes than unstructured use.
The “So What?”
AI eroding critical thinking is a design problem.
The default prompt bar positions the human as a requester and the AI as an oracle, and the research is clear about where that leads.
The good news is that every study showing harm also shows the harm is reversible through structure. The interventions that worked didn’t restrict AI use. They redesigned the interaction: what gets offloaded, when the human reflects, and what role the AI plays.
This matters most for K-12 students. They are building foundational knowledge. They have developing metacognitive skills, not established ones. The UTS report identifies a compounding equity risk: students who already have strong domain knowledge and self-regulation will use AI to accelerate their learning, while those without those skills are most susceptible to detrimental offloading. Left unstructured, AI widens the gap.
If you’re building AI interfaces, how might you:
Design interactions that afford thinking rather than replace it?
Build in structure around what gets offloaded to AI?
Create non-optional moments of reflection before the user acts on output?
If you liked this episode, you’ll love my book. Designing Automated Futures is coming this year with Rosenfeld Media. 📣 Sign up to be the first to know about new book releases, sales, and events.
⏪ Recent Episodes
ep. 88. AI Is An Accelerant. What Are You Feeding It?
ep. 87. What AI Can (and Can’t) Automate
ep. 86: ServiceNow’s Real Moat: Context Orchestration for AI
📖 Good Reads
Task analysis for AI agents: CMU Prof Dan Saffer walks through how this classic method from the 1960s helps you decide how to automate.
AI-moderated research needs a framework: Sr UX researcher Constantine Papas makes the pragmatic case that researchers needs to move past methodological debates and lead by defining routing logic, quality standards, and governance for these new tools.
Agentic AI Untangled: KPMG framework to navigate decisions to “build, buy, or borrow” AI agents.
That’s a wrap 🌯 . More on UX, HCI, and strategy from Sendfull in two weeks. Stay tuned for a new format, where I spotlight how practitioners are deciding what to automate with AI and what to keep human.
To me, the issue is not “AI causes cognitive offloading and should therefore be banned,” but rather “AI companies have shown they are not addressing cognitive offloading.” Soon, I think, it will be “AI companies cannot be trusted to make responsible products and therefore should be regulated.” Social media companies externalized their responsibility and are kind of being held accountable. But the effects of whole generations of cognitively illiterate people is terrifying enough to require regulation. The companies are not doing this.
The design question is 100% on point--and the writing tool you're describing is a genuine improvement over "here's your essay," right? But I keep thinking about how making the boundary visible isn't the same as making it legible. For a student who doesn't yet have the perceptual equipment to know what an "argument" worth articulating looks like, "write your two sentences first" might function as a compliance step rather than a judgment exercise--something to clear before getting to the formatting help. Seems like the scaffold works beautifully for the student who already has partial formation: it names what they need to do, and then gets out of the way. .. but for the student who lacks that formation, I'm betting the two-sentence exercise itself surfaces the gap--which means the tool ends up triaging the formation problem, rather than solving it. That might actually be the strongest design goal: not replacing prior formation, but making visible where it's absent. Which hands the student back to something slower, probably. Thanks for pressing this, Stef--the exchange clarified something for me, too.