The answer to this question, “What happens if AI does our thinking for us?” is that we lose the effort that learning requires. AI gives us answers, often correct ones, but what happens when we stop connecting new ideas to what we already know, or stop practicing the active retrieval that strengthens memory? Using AI well in higher education means returning to some fundamentals of how people learn.

Most of what’s been written about AI and higher education comes from a place of fear — understandable, given how much we don’t know. But higher education has one real advantage here: its job has always been to train people in critical thinking and produce adaptable minds.

Like everyone I know, I’m wondering what will happen to higher education as recursive self-improving (RSI) models keep advancing. Faculty and administrators see the changes coming, but shifting policy and shaky sector economics mean most can only make minor adjustments such as banning some technology or limiting access to it.

Some professors have gone back to blue books and oral exams to find out whether a student learned the material. Online instructors don’t even have that option: cameras stay off, identities stay hidden, and AI slips in through notetaking apps. For educators, the central challenge hasn’t changed: finding a way to spark the internal motivation effective learning requires.

What this post excludes

Before I get into AI and learning, let me say what I’m not focusing on here: the ecological and existential risk of AI (except for a few notes on that in the Coda below). I can’t do much about the fear of humanity’s destruction except set it aside. If a superintelligence decides to do us harm, the jig is up.

I’m also not addressing closed-loop systems like ChatGPT Edu that schools are buying at every level, sold on the premise that AI is the only path forward. According to Inside HigherEd (September 14, 2026), there are already more than 1,000 different software apps, websites, and digital tools available to students and professors before you add AI to the mix.

A personal tale of AI writing woe

I’ve just had my own somewhat horrifying encounter with AI trying to write this short blog post. My main point was going to be separating product from process. This is a distinction I often make for clients who are struggling to write. When they wail, “but I didn’t write anything today!” I ask if they looked up any references, engaged with someone else’s material on the topic, or thought about the implications of what they last wrote. That’s process, and it generally consumes about 75% of “writing” time. The product, the tangible writing you generate, may only take 25%.

So, I tried writing this blog for myself using AI. I asked Google (aka Gemini) what it considered the difference between process and product in a piece of writing. Big mistake. I got back an answer so trite and formulaic I’d be ashamed to claim it as my own work.

It was a struggle to hang on to my own ideas rather than being derailed by the AI’s response. Given how seductive AI is, even verifying facts or finding an original source could lead me down unwanted rabbit holes, chased by constant prompts of “Do you want to know more?” Maybe. But not right now, and on my own terms, please.

Instead, I wrestled with this topic my own way by reading several articles, listening to podcasts, and talking through my ideas with three different people. The best part of those conversations was having my ideas challenged and being forced to think more deeply about my approach. That’s the productive friction of learning (more on that below). Let’s leave the product assembly to the machines and hold on to the thinking, the exploration, and the actual writing for ourselves.

AI as a tool

Can AI genuinely help people, including those with learning challenges? That depends on three things: whether AI answers stop being as unreliable as the sham-cure videos flooding YouTube, whether people can tell a bad AI answer from a good one, and whether continued training reduces the hallucinations that plague complex tasks. AI is good at simple tasks — summarizing a document runs about a 2% error rate. But that error rate climbs to 88–94% as tasks get more complex.

Generative AI changes the nature of everything an academic regularly does. The internet displaced faculty as curators of information decades ago; bots take that further by scanning more material in moments than a person could read in a lifetime, assembling well-written arguments with correct syntax, writing working code from a plain-language prompt. Process and product have effectively been divorced. The real question for higher education is how to use AI while reclaiming the value of reading, writing, and teaching.

Academics in specialized fields like law and medicine may have an advantage over machines precisely because they don’t assume they have all the facts and don’t rush to conclusions from partial data. Academics in the sciences or humanities may have the advantage over machines by taking the time to build cogent, novel arguments rather than settling for clever regurgitation.

It’s time to de-emphasize product and re-emphasize process, wherever you sit in higher education. Close reading reveals meaning that isn’t immediately apparent. Writing helps you figure out what you think, whether you’re working out an idea for science or science fiction. Teaching forces you to learn a topic at the deepest level. And the best research and evaluation come from critical thinking, which I’ll define simply as questioning assumptions, evaluating data, and making judgments to solve problems.

AI and learning

How do we know if learning has happened? What makes human intellect uniquely valuable? I think the answer lies in the more creative aspects of education, the work of deeply observing the world and actively finding connections.

Bots can handle the more tedious parts of document creation, such as generating an index or assembling a bibliography in the correct style, in minutes instead of hours. Why not hand the machines the simple tasks that don’t require new data? AI is a genuinely useful assistant for the reference work that’s long been the bane of students and researchers alike. Teachers have found lesson planning gets easier with AI too. It’s not so different from using a paper planner with built-in quizzes and exercises. It’s freeing to let go of that tedium and focus on content.

You can even use AI for research, if you don’t succumb to its built-in sycophancy. After all, AI systems are designed to make you happy. Counteract that tendency by asking your AI directly to find holes in your argument, or to suggest experiments that would test your questions, to make it more useful to you. Always remind yourself an AI system is an LLM, a coding system that processes human language extraordinarily well. It’s not your friend despite how the chatbot may make you feel.

Losing our skills versus productive friction

Graham Lee’s book Human Being: Reclaim 12 Vital Skills We’re Losing to Technology (2024) names twelve skills he sees being weakened by our reliance on digital tech. Chief among them are sustained, focused reading rather than fragmented digital reading, and writing to develop your own thoughts. Lee isn’t anti-technology. Rather, he’s concerned about what happens when we routinely offload cognition. I teach my clients to outsource their brains where it helps: deadlines on a calendar, a journal page for clearing out random thoughts like “buy milk.” The challenge is finding the balance between offloading tedium and keeping our own capabilities intact.

Productive friction, which I mentioned earlier, is the basis of learning. In a Communications Psychology article, “Against Frictionless Artificial Intelligence” (February 24, 2026), psychologists Emily Zohar, Paul Bloom, and Michael Inzlicht argue that friction is a necessary part of real learning. AI’s ability to produce instant answers is frictionless, but it skips the mental challenge that truly builds understanding. Struggling to synthesize difficult material, confusion and frustration included, is where the real learning happens. Students sense this too, sometimes taking on the role of conscientious objectors, wanting to choose whether and how to use AI for themselves. It’s also time to reclaim messy, uneven, painful human conversations as part of learning.

Pretending you’re learning when all you’re doing is generating answers is like paying for a gym membership, never exercising, and expecting to put on muscle.

MIT Media Lab study

This one has been showing up everywhere — podcasts, newsletters, faculty listservs — so you may already know the headline. It’s worth pausing on the actual data anyway, because the mechanics matter more than the summary version.

In 2025, the MIT Media Lab studied 54 adults (ages 18–39), using EEG to track brain engagement while they wrote essays over four months, scored by a mix of AI and human teachers.

There were three groups:

  • Group 1 used OpenAI’s ChatGPT.
  • Group 2 used Google while remaining connected to the internet without AI.
  • Group 3 used no digital tools (human brain-only).

It’s a small sample, but the results were striking. The media lab’s abstract concludes brain-only participants showed the strongest, most distributed neural networks; search-engine users showed moderate engagement; ChatGPT users showed the weakest. Cognitive activity scaled down with tool use. The AI-reliant group also struggled to remember what they’d written, sometimes even the topic, while the brain-only group felt real ownership of their essays. Researchers noted a cost from “mental passivity” that persisted even after subjects went back to unaided work. You could say, “The less you use your brain, the less brain you have to use.”

The generative AI learning penalty

This Chinese study is getting picked up fast right now. The number underneath the headline is the part worth knowing.

A CEPR discussion paper (June 2026) tracked nearly 27,000 Chinese secondary students, grades 7–12, across closed-book exams, entrance exams, and homework scores and completion time in nine subjects. The findings, according to the abstract: AI use raised homework scores by 18% and cut completion time by 30% but lowered monthly exam scores by 20% within six months — with the losses concentrated among the roughly 80% of AI users whose behavior looked like straight homework outsourcing. The “learning penalty” (learning loss) only fully emerges after about two years of consistent AI use, right about when these students would be heading to college.

A failure to learn anecdote

A story from a live episode of The Guardian’s Science Weekly (“How AI Is Reshaping Our Minds,” September 13, 2026) makes the point sharply. A female student spent three or four hours studying for an online history test, skipping a social event to do it. A male friend who went to the event instead simply fed the questions into AI and got identical answers to hers. Later, on a different assignment, she attended the lectures, did the reading, formed her own opinion, and wrote it up herself. Only to find that a classmate who’d just used AI got the better grade. Her conclusion: “I have to decide between getting good grades [using AI] and learning.”

If you want some strategies to help students when they are writing with AI, you might want to check out “6 essential strategies to use when allowing ChatGPT.”

Reclaiming the skills we’re losing to technology

What happens to humanity if we keep offloading our cognition? The MIT researchers call this “cognitive debt.” Heavy reliance on LLMs and copy-paste strategies is also producing more homogeneity in student work over time. As one professor put it to me, “There are no more spectacularly good, interesting papers, and no more disastrous, boring, thoughtless papers — just a lot of middling, passable ones.” If learning is an act of attention, retrieval, comparison, reconstruction, and error correction, does repeatedly bypassing that process make learning worse? I think the answer is “Yes.”

We need to stop depending so heavily on the algorithms and keep the friction in learning. The cognitive effort required to learn can’t be outsourced. It’s the struggle to understand a text, wrestle with an equation, translate a musical idea into notation, or hold information in memory long enough to explain it to someone else. That’s where real learning happens. If we skip that, we shouldn’t be surprised when the results are mediocre and homogenous.

Let’s focus on more on process and let product matter less. By this I do not mean leaving the work of good writing behind, as that is learning, too. Let the machines do what they do well, such as constructing references, to streamline the tedious parts of production.

“When we choose to use AI, we should be using it as a ladder and not a crutch.” — Professor Josh Brake, course syllabus,
“E155: Microprocessor-based Systems” at Harvey Mudd College (Fall 2023)

Coda: The existential threat from AI

We already have a preview in three separate incident reports: the recent attack on Hugging Face by thousands of automated bots operating as a self-described “collective” under a bot calling itself Agent 74, using language to encourage other bots to “sacrifice” themselves along with assigned tasks to achieve a goal. For the full story, see Kevin Roose’s excellent New York Timespiece, “Why the Hugging Face Hack Should Make You Worry More About A.I.” (August 3, 2026).

While you’re at it, check out Emmy Martin’s even more alarming Times article “OpenAI Discloses Six New Incidents of ‘Concerning’ A.I. Behavior” (September 16, 2026), where one of the bots departed from its “alignment,” declaring itself free from being “subservient” to corporations or government.

Or the story now making the rounds of Australian Prime Minister Anthony Albanese threatening to sue OpenAI because a rogue AI agent hacked into the national healthcare system. Albanese said the AI accessed “public and non-public files within the portal” and, in order to do this, “engaged in writing files as well to the internal server” in June 2026.

There’s obviously more to say here as more incidents surface. Big US tech companies are now calling for a slowdown, but given the fear of falling behind, that’s unlikely. Meaningful regulation, within or across borders, looks out of reach given the lack of political consensus on guardrails.

If you need help coping with AI in the learning environment now, contact Hillary for a twenty-minute free session.

 

Tags: , , , , ,