aiX Weekly — AI in Higher Education (September 2nd, 2026)

Writing is thinking. Coding is thinking. Exploratory data analysis is thinking, and so is creating a good visualization. Working with AI can be thinking too — when you actively engage.

So if all of these are thinking, why do we keep saying AI erodes learning? This issue argues the erosion is not about AI. It is about which act of thinking gets outsourced.

Each aiX Weekly post is organized around a set of recurring sections and pairs with our companion AI and Higher Education timeline, which traces the broader arc of how AI has reshaped higher education since late 2022.

Curated by Claude for the aiX Programs, Columbia University. AI can make mistakes and so can the human reviewer. Please double-check the linked sources.

Review by Tian Zheng on August 26, 2026

TL;DR / This Week at a Glance

The dominant conversation in higher education this fall is some version of “if AI can write the essay, what is the essay for?” The most useful reframing is to ask what writing does for the writer. The cognitive psychologist Ronald T. Kellogg described writing as “a technology for thinking”: it forces you to hold a subject in working memory, organize it, and notice what you do not yet understand. Coding is a technology for thinking too. So is exploratory data analysis — you form a question, look, and revise the question. So is creating a good figure, because deciding what to show is deciding what you believe.

• The slogan is right but incomplete. Writing, coding, exploratory data analysis, and visualization are all forms of thinking because each is a generative act — you build understanding by producing and revising something.
• Offloading is normal — and often good. We write things down to remember them; we use version control so research is reproducible without constant effort. That is offloading, and it frees cognition for the harder work. The problem is not offloading; it is offloading the generative act itself.
• AI is not the variable; substitution is. Studies converge on a distinction between deliberate offloading (delegate the routine, keep the judgment) and unconscious offloading, or “cognitive surrender,” where the learner defers entirely. The first can raise critical thinking; the second erodes it.
• The evidence has a shape. The MIT “cognitive debt” study found the weakest brain connectivity and lowest recall in AI-only writers — but the group that wrote first, then used AI fared better.
• Working with AI can be thinking when it is used as a sparring partner you interrogate, not an oracle you copy. The design question is where the generative act lives.

Relevant to faculty: Ask of each assignment: which generative act do I want the student to perform? Protect that act. Let AI assist around it, not inside it.

Relevant to institutional leaders: “AI literacy” should include the difference between deliberate delegation and cognitive surrender. That distinction, not tool access, is what predicts whether learning improves.

Relevant to students and researchers: If you reach for AI before you have tried, you may get a correct answer and no understanding — what one paper calls “unproductive success.” Try first, then bring AI in to challenge and extend your thinking.

Table of Contents

• TL;DR / This Week at a Glance
• Research Highlights
• What’s in the News
• Institutional Movements
• Most Discussed
• What Changed
• From My Desk
• Try This Week

Research Highlights

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing (MIT Media Lab, arXiv, June 2025; foundational context.) Fifty-four participants wrote essays under three conditions — LLM, search engine, or brain-only — under EEG. AI-only writers showed the weakest neural connectivity and could recall only ~17% of their own text afterward, versus ~46% for brain-only writers. Crucially, the group that wrote unaided first and then used AI (Brain-to-LLM) showed stronger engagement than the reverse.

Editor’s note: The headline gets read as “AI makes you dumber.” The interesting finding is directional: how to engage AI makes a difference: AI after your own attempt looks different, in the brain, from AI instead of it. — TZ

Outsourcing thinking to AI? Focused immersion, AI dependency, and the double-edged impact on critical thinking (Humanities and Social Sciences Communications, 2026.) This study models AI use as double-edged: engagement can deepen critical thinking, while dependency undercuts it. The effect turns on how the tool is used, not merely how much.

Editor’s note: The variable that flips the sign is whether the user stays cognitively in the loop. — TZ

The effects of a critical-thinking intervention on reliance behaviors, problem-solving quality, and creativity during human–GenAI collaborative learning (2026.) Teaching students how to collaborate with AI — when to delegate, when to interrogate — reduced over-reliance and improved problem-solving and creativity relative to unstructured use.

Editor’s note: This is the most hopeful result of the week. The skill of keeping the thinking with yourself is teachable. — TZ

The Effortless Trap: Productive Struggle, AI, and the Illusion of Learning (arXiv, 2026; recent context.) Argues that GenAI can create an effortless trap by delivering “unproductive success” — a correct output produced without the struggle that builds understanding — and examines timing of access as a design lever.

Editor’s note: Unproductive success is a learning outcome where a person achieves short-term high performance or passes an assessment, but fails to build deep, long-term understanding or retention. AI is not the only condition that can create this but is surely a very effective one. — TZ

What’s in the News

What do students lose when they stop writing? (The Philadelphia Inquirer, August 18, 2026; recent context.) A widely shared feature makes the “writing is thinking” case directly, citing worries that students who route first drafts through a chatbot never discover what they actually think. It reports a group of writing experts at UC Irvine urging schools not to spend less time on writing — the point of writing being to sharpen the mind, not only to produce prose.

Editor’s note: The root of the issue is not students routing first drafts through a chatbot. The real problem is they don’t challenge the AI-generated contents and they don’t attempt to ensure the work represent their true thoughts. — TZ

Students increasingly worry AI harms their critical thinking (RAND, American Youth Panel; context.) Surveyed students report both heavier AI use and rising concern that it makes learning too shallow — roughly two-thirds worry about the effect on critical thinking, with a notable share saying they retain less because they lean on it. The concern is coming from students themselves, not only faculty.

Editor’s note: Students shouldn’t think critical thinking can only be attained by avoiding AI. — TZ

Frontier model capability kept rising this week. Anthropic’s Claude Fable 5.1 (Sept 1), Google’s Gemini 3.8 Flash (Sept 2), and OpenAI’s GPT-6 Astra (Sept 3) all shipped in three days — raising, not lowering, the stakes of the question about which acts of thinking we ask students to keep.

Institutional Movements

Protecting the generative act by changing where work happens. A visible response this term is structural: more in-class handwritten writing, oral defenses, and supervised assessment, with several institutions — including reports of Princeton stepping back from some unsupervised testing traditions — moving to make the thinking observable again.

Provisioning continues, and the literacy layer is where the theme bites. Campuses are still turning on institution-wide access (the University of Leicester’s full Microsoft 365 Copilot rollout this month is a clear example), and peer programs like OpenAI’s Student Collective are expanding. The open question is whether the accompanying “AI literacy” teaches delegating the routine while keeping the reasoning.

Editor’s note: A literacy module that teaches privacy and prompt syntax but not the difference between using AI to think and using it to avoid thinking is incomplete. — TZ

Most Discussed

MIT’s Ad Hoc Committee releases its report on AI and education (MIT, August 13, 2026). The committee’s report documents disruption to campus life — fewer office-hour visits, weaker study groups, less in-person engagement — alongside inconsistent AI policies across courses and mounting student anxiety about integrity and career relevance. Its central argument fits this issue’s theme directly: MIT should respond with “augmentation not automation,” redesigning courses around clear learning goals and experiential, community-based work rather than relying on restriction alone, since learning is a cultural practice built on human connection, not just content transfer.

Question for discussion: is the problem AI, or is it substitution? If writing, coding, EDA, and visualization are all thinking, then the same tool can either host thinking or replace it. The distinguishing variable is not the technology but the location of the generative act. A student who drafts an argument and then uses AI to attack it is thinking. A student who prompts for the argument and submits it is not — regardless of how good the output looks.

If we accept that framing, “AI literacy” becomes less about tool proficiency and more about metacognition: knowing which part of a task is yours to think through, and refusing to hand it over.

We have never taught critical thinking (Inside Higher Ed, June 23, 2026). Soundar-Shah and Kestigian argue that the crisis predates AI: higher education has long relied on a “by-product model,” assuming critical thinking develops incidentally from rigorous readings and disciplinary content rather than being taught explicitly, and roughly 45% of students show no measurable gain in it over their first two years. AI does not create this failure; it removes the friction — research, drafting, revision — that was quietly doing the teaching the curriculum never did explicitly.

Editor’s note: I fully agree with this one. AI’s disruption is forcing the real question into the open: why do we ask students to generate a given piece of work in the first place — in our classes, and in the careers they’re headed toward? If we can name that reason, we can teach the thinking directly instead of hoping it accumulates as a by-product of the assignment. — TZ

What Changed

• Capability rose again — Claude Fable 5.1, Gemini 3.8 Flash, and GPT-6 Astra all shipped September 1–3. Better models make the substitution more tempting and more invisible, which raises the value of designing for the generative act rather than the output.
• The framing is shifting from detection to design. Less “can we catch AI writing?” and more “which thinking do we require, and how do we make it visible?” — a move the research this week supports.
• Timing of access is emerging as a design lever. Work on access timing as scaffolding suggests when students can reach for AI (before vs. after their own attempt) may matter as much as whether they can.

Try This Week

Make the thinking come first

Take one assignment and split it into two visible moves. First, the student does the generative act unaided — writes the rough argument, sketches the EDA plan, drafts the function, or names what they expect the data to show. Only then may they bring in AI, and their task becomes interrogation: what did the model add, what did it get wrong, what did they change and why?

Grade the second move as much as the first. Ask for a short “what I thought before AI / what changed after” note attached to the work. It costs a paragraph, and it does two things at once: it guarantees the student performs the thinking before the tool can substitute for it, and it turns “using AI” into an act of critical evaluation — which, done this way, is thinking too.

aiX Weekly is curated by Claude and reviewed by Tian Zheng for the aiX Programs at Columbia University. Editorial notes reflect one statistician’s reading of the week and invite discussion. Corrections and suggestions are welcome.