Cognitive Offloading: Surprising Risk When Minds Handle AI

Cognitive offloading, handing mental work to something outside your head, is as old as the shopping list. What is new is a technology built to take on the thinking itself. In an essay published by The Verge on 5 October 2026, education writer Benjamin Riley argues that our minds are not equipped for that, and that the AI industry’s habit of treating brains as computers is why it keeps misjudging the damage. His image for the result is memorable: AI is “a cognitive version of a hot dog”, tempting in the moment and harmful as a regular diet.

The argument arrives as evidence and policy are moving fast. A working paper tracking 26,811 Chinese secondary pupils found homework scores rose with AI while exam scores fell. Norway has told schools that children aged 6 to 13 should, as a rule, not use AI. New York City and Los Angeles, the two largest school districts in the United States, have paused most student AI use for this school year.

This article explains Riley’s case, the neuroscience behind it, what researchers mean by cognitive offloading and why AI changes the stakes, the strongest evidence for and against, how schools are responding, and what businesses should do to keep their own people’s thinking sharp.

Table of contents

What Riley Argues About Cognitive Offloading
The Brain Is Not a Computer: Why It Matters for Cognitive Offloading
From Stone Tools to Schools: Culture as a Feedback Loop
What Cognitive Offloading Means, and What Is New About AI
The Evidence on Cognitive Offloading and Learning
Schools Push Back Against Cognitive Offloading
Where the Case Against Cognitive Offloading Overreaches
Managing Cognitive Offloading at Work
Cognitive Offloading FAQ
References and Further Reading

What Riley Argues About Cognitive Offloading

Riley’s essay, headlined “Our minds aren’t equipped to handle AI”, makes a three-part case: the brain is not a computer, human thinking grew out of feedback with the physical and social world, and AI weakens that feedback by doing the effortful part for us.

The cognitive hot dog

The central metaphor is food. Our hunter-gatherer ancestors evolved to seek out scarce fatty foods; agriculture then made those foods plentiful, and the same instinct now harms us. “AI poses a similar sort of danger to our cognitive capabilities, by clogging our capacity to develop the knowledge we need – in our heads – to navigate the world,” Riley writes. “It’s a cognitive hot dog.”

He is careful about the dose. “The occasional hot dog won’t harm anyone, but it sure will if it becomes a regular meal at lunch or dinner. The same is true for AI.” His target is cognitive offloading as a habit, not occasional use.

Who is making the argument

Riley runs Cognitive Resonance, a “think and do” tank on human cognition and generative AI, whose newsletter describes its mission as “building human knowledge to halt AI hype”. He discloses in the essay that he has given informal advice to Schools Beyond Screens, a group that pushed for the classroom AI bans he praises. That does not weaken the evidence he cites, but readers should know where he stands.

The Brain Is Not a Computer: Why It Matters for Cognitive Offloading

Riley’s deepest claim is about models of the mind, because the model you hold decides what you think AI can safely replace.

Input, computation, output

The idea that brains are computers traces back to Alan Turing. Perception supplies input, the mind computes, and action is the output. The AI industry largely shares this view: Riley quotes Google DeepMind’s Demis Hassabis calling the brain “a biological approximation to a Turing machine”, and Elon Musk saying “people should just think of the brain as a biological computer.”

Riley grants that the model “has been productive”, since it guided computing from adding machines to neural networks. But he notes that John von Neumann, a founder of computer science, doubted it could capture the “exceptional complexity of the human nervous system.”

Catching a fly ball: the feedback-control view

The alternative comes from Paul Cisek, a neuroscientist at the University of Montreal, who builds his model of the brain from evolutionary history rather than from computers. In his view, brains are feedback-control systems: we act to change what we sense, not just react to it. The philosopher John Dewey described the mind as a circuit in much the same way more than a century ago.

Riley’s example is an outfielder catching a fly ball. The computational model implies the player subconsciously calculates velocity and gravity. The feedback-control model suggests a simple rule: keep the ball in the same place in your visual field and move to keep it there. Engineers will recognise the shape of the idea: control systems correct themselves using feedback, and reinforcement learning trains AI by having it act and observe the results.

Question	Computational model	Feedback-control model
What is thinking?	Processing input into output, like an algorithm	Acting to keep control of what we sense
Catching a fly ball	Subconscious calculation of the ball’s path	Move to keep the ball steady in view
How it was built	Reverse-engineered from what minds produce	Built up from evolutionary history
Where learning happens	Inside the processor	In loops between body, world and other people
What AI looks like	A faster processor to plug in	Something that can cut the loops learning depends on

The difference is not academic. If thinking is computation, cognitive offloading to a faster computer looks like pure gain. If thinking is a set of control loops we build by using them, offloading the effort can stop the loops from forming at all.

From Stone Tools to Schools: Culture as a Feedback Loop

Riley extends Cisek’s model from biology to culture, arguing that our most important feedback loops are social, and those are the loops heavy cognitive offloading most threatens.

Imitation, language and writing

Hundreds of thousands of years ago, our ancestors learned to imitate each other, passing on practices such as chipping a stone tool. Imitation led to spoken language, settled farming, writing and, soon after, formal education, along with institutions such as markets, law and democracy. Riley admits that “each claim I’ve just made is contestable”, but the direction is the point: each step extended human control over the world, and each depended on people learning from one another.

Why education is the institution at risk

“No other species’ brain and body develops over such an extended period as ours,” Riley writes, and schools are how cultures pass knowledge between generations. That is why he treats education as the clearest case of cognitive offloading going wrong. “We’ve spent thousands of years building institutions and norms for learning from and communicating with each other, and over the course of less than a decade, Big Tech companies have systematically worked to dismantle them.”

What Cognitive Offloading Means, and What Is New About AI

Researchers had a name for this long before chatbots, and the history helps separate real risks from panic.

From calendars to chatbots

In a 2016 paper in Trends in Cognitive Sciences, Evan Risko and Sam Gilbert defined cognitive offloading as “the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand.” Writing an appointment in a calendar counts. So does using a calculator or tilting your head to read rotated text. Most cognitive offloading is sensible: it frees attention for harder work.

Critics of AI scepticism often point to this history, along with Plato’s supposed opposition to writing. Riley has heard the Plato argument so often that he answered it in a separate piece, noting that Plato was himself a writer. His real reply is about degree: “never before have we developed and broadly deployed something so explicitly intended to supplant human thinking.”

Cognitive automation and cultural technologies

Two framings Riley borrows make the distinction sharper. François Chollet, a former Google engineer, calls AI “cognitive automation”: “encoding human abstractions in a piece of software, then using that software to automate tasks normally performed by humans.” And a 2025 paper in Science by Henry Farrell, Alison Gopnik, Cosma Shalizi and James Evans argues that large language models are best seen “as a new kind of cultural and social technology, allowing humans to take advantage of information other humans have accumulated.”

A calendar stores a fact you already understood. A chatbot can produce the understanding for you. That is the step from ordinary cognitive offloading to what researchers now call cognitive delegation, and it is why learning is where the costs show first.

The Evidence on Cognitive Offloading and Learning

Riley’s essay links to a fast-growing body of research. The strongest studies share a pattern: AI makes the work look better while the learning underneath gets worse.

Study	Sample	Key result
Stromberg, Lei and Wu, The Generative AI Learning Penalty (SSRN, June 2026)	26,811 pupils in a county in central China over 30 months	Homework scores up 18%, homework time down from 64 to 45 minutes, monthly exam scores down 20% within six months
Rismanchian and colleagues, Faster Completion, Less Learning (arXiv, May 2026)	3.2 million ALEKS maths learning interactions over ten years	Study time on AI-susceptible problems fell 26.9% for college students over eleven quarters; odds of a correct proctored answer fell 25%
Kosmyna and colleagues, Your Brain on ChatGPT (MIT Media Lab, 2025)	54 students writing essays with ChatGPT, search or no tools, with EEG	ChatGPT users showed weaker engagement and recall when later writing unaided, which the authors call cognitive debt
Lee and colleagues, Carnegie Mellon and Microsoft Research (CHI 2025)	319 knowledge workers, 936 real examples	Higher confidence in generative AI was associated with less critical thinking
Gerlich, Societies (January 2025)	666 participants, surveys and interviews	Frequent AI use correlated with lower critical thinking, mediated by cognitive offloading

The Chinese homework study

The most striking evidence is the working paper by David Stromberg of Stockholm University and Victor Lei and Yanhui Wu of the University of Hong Kong. Following 26,811 pupils for 30 months, they found that generative AI raised homework scores by 18% and cut completion time from 64 to 45 minutes, yet monthly closed-book exam scores fell by 20% of the baseline mean within six months. High-stakes entrance exam scores fell by 24% for the zhongkao and 18% for the gaokao among pupils who had used AI for two years or more.

Fall in exam scores after AI adoption, by subject (Chinese secondary pupils)

Social sciences: 27%

STEM subjects: 22%

English: 17%

Chinese: 9%

Subject figures as reported by The Star’s coverage of the Stromberg, Lei and Wu paper; bar widths are the percentages on a 0 to 100 scale. Most users were not using AI as a tutor: by June 2025, 81% of established users showed signs of outsourcing their homework. That is cognitive offloading of exactly the kind Riley warns about, and the strongest pupils lost the most, with top performers down 24% against 16% for lower achievers.

Three million maths problems

A second study, by Sina Rismanchian and colleagues, used ten years of data from the ALEKS learning platform, covering 3.2 million learning interactions. After ChatGPT’s release, time spent on problems that AI can easily solve fell by 26.9% among college students over eleven quarters, by 31.3% among high schoolers and by 9.0% among middle schoolers, with no detectable change for Grade 5. The drop vanished entirely under proctoring, which rules out simple efficiency gains. The authors call the pattern “cognitive surrender”.

Brains, essays and critical thinking

Smaller studies of cognitive offloading point the same way. MIT Media Lab’s 2025 “Your Brain on ChatGPT” experiment recorded brain activity while 54 students wrote essays and found weaker engagement and recall among those who had relied on ChatGPT. Carnegie Mellon and Microsoft Research surveyed 319 knowledge workers and found that greater confidence in AI went with less critical thinking. Michael Gerlich’s study of 666 people linked frequent AI use to lower critical thinking scores, with cognitive offloading as the mediating factor and younger participants most dependent.

A tipping-point model

A September 2026 preprint by Ricard Solé, David Krakauer and colleagues models large language models as “a cognitive virus” spreading through populations. Their model shows how social pressure can create tipping points, after which “small increases in adoption can trigger rapid population-level shifts toward persistent dependence.” Riley draws on their prescription of “cognitive immunization”, which preserves “unaided problem solving, verification, critical discussion, [and] periods of deliberate disengagement” from AI.

Schools Push Back Against Cognitive Offloading

The policy shift against cognitive offloading in classrooms, which Riley welcomes, has happened in a matter of months.

When	Who	What
27 May 2026	American Federation of Teachers	Ten-point “Devices down, eyes up, hands-on” plan: no student-facing AI in elementary grades and no AI chatbots for under-16s
19 June 2026	Norwegian government	Pupils in grades 1 to 7, aged 6 to 13, should as a rule not use AI; cautious supervised use from ages 14 to 16
2 September 2026	New York City schools	One-year moratorium through 8th grade covering nearly 600,000 students; ChatGPT, Claude and companion chatbots blocked for all grades
September 2026	Los Angeles Unified School District	One-year pause on student generative AI on district devices for about 378,000 students; teachers keep access

Norway, New York and Los Angeles

Norway’s prime minister, Jonas Gahr Støre, said AI use increases the risk that young children skip important steps in their education; the country banned smartphones from schools in 2024. In New York, Mayor Zohran Mamdani was blunter: “The tech industry wants us to believe that AI in early education is not only inevitable, but that it is necessary. We do not see it that way.” We covered the detail of the NYC ban on student AI use when it was announced. Los Angeles disabled the Gemini chatbot, AI Mode in Google Search and AI assistants in Google Classroom and Docs on student accounts.

Students themselves are part of the pushback. Riley cites the Oberlin Luddite Club, whose members typed an open letter to their college president on a typewriter more than 70 years old: “As you embark on your year of AI, we’ll embark on our own year of self-actualization.”

The industry is moving the other way

At the same time, AI companies are pushing deeper into education. Riley quotes OpenAI’s vice-president of education, Leah Belsky, calling ChatGPT “the world’s largest learning platform” in a July 2025 podcast. In July 2026 Anthropic launched Claude for Teachers, giving verified US K-12 educators free access to premium Claude capabilities, with sign-ups open until 30 June 2027.

The picture is not simply teachers against tech. The American Federation of Teachers, which wants student-facing AI out of elementary schools, is also working with Anthropic on a privacy “gold standard” for Claude for Teachers. The emerging line is AI for teachers’ preparation and administration, with far more caution about putting it in front of children.

Where the Case Against Cognitive Offloading Overreaches

Riley’s essay is a polemic, and a fair reading should note where the evidence is less one-sided than the hot-dog image suggests.

Purpose-built tutors can work

The best counter-evidence comes from Harvard. In a randomised trial published in Scientific Reports in June 2025, Greg Kestin and colleagues found that 194 physics students learned more than twice as much from a custom AI tutor as from an excellent active-learning class, in a median of 49 minutes. The key word is custom: the tutor was built to make students do the thinking, using the same teaching principles as the classroom sessions. Riley says AI tutors are “already failing”, but this result shows design can matter as much as the technology.

The Chinese study draws the same line from the other side: its authors separated pupils who outsourced homework to AI from those who used it as a tutoring aid, and most fell into the first group. Cognitive offloading depends on how a tool is used, not only on whether it is used.

The computer metaphor still earns its keep

Riley admits the computational model “has been productive”, and many neuroscientists would say feedback control and computation are complementary descriptions rather than rivals. His evolutionary story is also, by his own account, contestable in its details. The practical case against heavy cognitive offloading in education does not depend on settling the philosophy of mind; the exam-score evidence stands on its own.

Managing Cognitive Offloading at Work

The debate is about schools, but cognitive offloading happens in every organisation adopting AI. The people most exposed are those still learning a job: juniors, apprentices and anyone moving into a new field.

Decide what must stay in people’s heads

Not every task needs to build expertise. Formatting a report or summarising a long thread is sensible cognitive offloading. Diagnosing a fault, drafting the first version of an analysis or checking a contract clause is how people become good at their jobs. Map your teams’ work against that distinction as part of your AI strategy, rather than leaving each person to decide alone.

Type of task	Offload to AI?	Why
Formatting, transcription, routine summaries	Yes	Low learning value, easy to check
First drafts of analysis by junior staff	Not at first	The drafting is how judgement develops
Fault diagnosis and troubleshooting	After an unaided attempt	Builds the mental model needed when AI is wrong
Reviewing AI output	No	Verification is the human’s job and the skill to protect
Training and certification tasks	No	Assessment must measure the person, not the tool

Build in deliberate disengagement

The cognitive immunisation research recommends “periods of deliberate disengagement” from AI. At work that can mean AI-free first attempts for trainees, regular problem-solving sessions without assistants, or rotating staff through tasks they would normally hand to a tool. The Rismanchian study’s proctoring result is a reminder that effort reappears when the shortcut is removed.

Make verification the job

The Carnegie Mellon and Microsoft study found AI shifts critical thinking towards verifying, integrating and overseeing AI output. That is a skill, and it can be trained and measured. Treat checking AI work as a named responsibility with time allowed for it, especially where errors are costly, rather than an afterthought squeezed into a faster workflow.

Cognitive Offloading FAQ

What is cognitive offloading?

Cognitive offloading is using something outside your head to reduce the mental effort a task requires, such as a calendar, a calculator or, increasingly, an AI assistant. Researchers Evan Risko and Sam Gilbert defined it in 2016 as “the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand.”

Is cognitive offloading always bad?

No. Most cognitive offloading frees attention for harder work. The concern with AI is that it can take on the effortful thinking that builds knowledge and skill, so people get the output without the learning.

What does Benjamin Riley mean by a “cognitive hot dog”?

Riley compares AI to junk food: appealing in the moment, harmless occasionally, but damaging as a regular diet. He argues AI clogs our capacity to build the knowledge we need in our own heads.

What is the evidence that AI harms learning?

A study of 26,811 Chinese pupils found AI raised homework scores by 18% while monthly exam scores fell by 20% within six months. An analysis of 3.2 million ALEKS maths interactions found college students’ study time on AI-susceptible problems fell 26.9% over eleven quarters.

Can AI help people learn?

Yes, when it is designed for learning. A Harvard randomised trial found a purpose-built AI physics tutor produced more than double the learning gains of an active-learning class. The difference is whether the tool makes the learner think or does the thinking for them.

Which places have restricted AI in schools?

Norway has said pupils aged 6 to 13 should generally not use AI. New York City and Los Angeles have paused most student AI use for the 2026 to 2027 school year, and the American Federation of Teachers has called for no student-facing AI in elementary schools.