The Skill AI Didn't Actually Break
The "AI is eroding critical thinking" panic is directionally right and diagnostically wrong — AI didn't break the skill; it exposed that we were never really teaching it.
TL;DR
- A June 2026 MIT study found that people who relied heavily on GPT-4o to identify fake news got worse at spotting it themselves. A separate 2025 study of 600+ people found the same correlation, sharpest among the young.
- Nearly three in four US K-12 teachers now believe AI has bigger implications for education than the internet did (NPR/Ipsos, 5 June 2026). 90% of US faculty surveyed by the AAC&U in autumn 2025 agreed generative AI will diminish critical thinking.
- Estonia gave every 10th and 11th grader a free "Socratic" ChatGPT that refuses to do their homework. Singapore employers now pay a premium for critical thinking skills (ManpowerGroup Q3 2026).
- The real story isn't in the data. It's in the quieter admission from inside universities: if a tool this new can erode critical thinking this quickly, we probably weren't teaching it in the first place.
- What to do about it isn't a resolution. It's a habit small enough to survive Tuesday.
The number
80%. In one Washington Post analysis (24 June 2026), ChatGPT gave exclusively left-leaning answers to political questions 80% of the time. Claude gave left-leaning answers 43% of the time. Gemini split both ways in more than 90% of answers.
Put aside the politics. Notice the shape of the finding. Three of the most-used reasoning tools on the planet return systematically different worldviews depending on which one you pick — and most people don't pick. They open the tab that was already open.
That is the story. Not that AI has opinions. That we've quietly handed reasoning to systems whose opinions we don't audit, and then measured our own reasoning and found it thinner. Then acted surprised.
What the studies actually say
Three findings, converging.
The MIT study (June 2026). Participants spent a month answering questions about whether news headlines and images were real. Some had access to an AI assistant running on GPT-4o integrated with Google search. The assistant was useful — accuracy went up when it was in the room. The catch is what happened to the humans over the month: people who relied heavily on it got worse at the task without it. The tool didn't teach them to think. It thought for them and left them thinner.
The 600-person study (2025, resurfaced widely in June 2026). Observational. It found a correlation — not causation — between heavy AI use and lower critical-thinking scores. The gap was sharpest in younger users. The researcher was careful about the caveat, which most of the coverage was not.
The NPR/Ipsos poll (5 June 2026). 545 K-12 teachers surveyed. Almost three in four believe AI's educational impact will exceed that of the internet or the personal computer. A majority worry it's making it harder for students to learn to think for themselves. Around 40% report students using AI in class at least weekly. Half say it isn't used in class at all — which, given the smartphone in every teenager's pocket, is a survey answer, not a fact.
Underneath these sits a fourth number: the AAC&U's autumn 2025 survey of American faculty, which found 90% agreement that generative AI will diminish student critical thinking.
Ninety percent is not a debate. It's a diagnosis the profession has already agreed on.
The reframe
Read those numbers alongside a piece Inside Higher Ed ran on 23 June, and something shifts.
The argument, briefly: if a new tool can erode critical thinking in eighteen months, the critical thinking probably wasn't there in the first place. Most instructors don't teach the skill directly. They teach content — history, biology, literature — and assume critical thinking accrues as a byproduct. The skill is the residue.
Residue is not a training regime.
This is the wedge. Everyone is arguing about whether AI is making people dumber. The more interesting question is what AI is revealing. Because the pattern in the data — people accepting fluent, confident output from an authoritative source without checking it — did not begin in November 2022.
Search engines did it first. So did television. So did textbooks. So did teachers.
AI is the fastest, most fluent, most confident version of the same problem. And now the problem is legible.
Who this actually affects
Students. Directly. The Estonia experiment (WSJ, 5 June 2026) is the sharpest response so far: a custom "Socratic" ChatGPT that will not complete a student's work, only ask them questions back. Every 10th and 11th grader in the country of 1.3 million now has one. It is the first policy in the world that treats the tool as a Socratic tutor by design, rather than hoping students will use it that way.
Workers. Also directly, and quickly. In Singapore, the ManpowerGroup Q3 2026 survey found 66% of employers willing to pay a premium for AI literacy, 64% for AI model development, and — critically — a matching premium for critical thinking. Info-Tech's IT Talent Trends 2026 report went further: technical skills alone will no longer define IT hiring. The differentiators are "uniquely human abilities" — critical thinking, creative problem solving, business acumen.
The labour market has already priced this in. The premium is being paid to the people who can check AI, not the ones who can use it.
Teachers. Painfully. The NPR/Ipsos numbers show a profession that has correctly diagnosed the problem and is being asked to solve it with the same time, the same class sizes, and no new training.
Employers in developing economies. In Fiji, Fiscal Review Committee chair Richard Naidu said in June that the education system is "not fit for purpose" — his phrase, drawn from conversations with other employers, in the specific context that AI is now widening the gap between developed and developing economies faster than the graduates are ready. That gap is a critical-thinking gap.
Everyone else. Second-order. If you consume news, use a browser, or make a decision that ends in a purchase — the systems around you are getting more persuasive faster than most people are getting more skeptical. The asymmetry is the story.
The quieter story
A short section, because it needs to be short.
There is a version of this article that says: put down the AI, read a book, think for yourself. That version is comforting and probably wrong.
You are not going to put down the AI. Neither is anyone else. AI use is going to increase. The productivity gains are real. The convenience is real. The competitive pressure to use it — at work, at school — is real and rising.
The correct question is not how do I resist this? The correct question is what small habit, done daily, keeps the muscle I need?
The atomic behaviour
Critical thinking, at the level it can be practised on a Tuesday, is one habit: ask one more question before you agree.
That is the whole skill compressed into a sentence. Everything else is scaffolding.
You do not become a critical thinker by resolving to become one. You become one by adding a small friction between input and belief. The friction is a question. Any question, honestly asked, that a fluent source hasn't already answered.
Some that work:
- What would the opposite view get right?
- Where did this source's data come from?
- Who is not in this frame?
- What is this argument missing that I already know?
- If I'm wrong about this, what would I see?
Ask one. Not five. One is a habit. Five is a resolution, and resolutions do not survive Tuesday.
Do it once a day, applied to something an AI told you or an authority you respect. That's it. The point is not to reach a better answer every time. The point is to keep the muscle alive.
Identity precedes behaviour. If you want to think critically, don't set a goal to think critically. Become the kind of person who asks one more question. The behaviour follows the identity, not the other way around.
What organisations can do (that isn't wallpaper)
Two things, and only two, matter.
One: teach the skill explicitly. Not as the residue of coursework. As the subject. Estonia's Socratic bot is one way — force the tool into a tutor role by design. The Cambridge FutureReady pilot delivered in Wuhan in April–May 2026 is another: small discussion groups where students have to articulate their reasoning, out loud, to another human. Both work because they make the skill visible rather than assuming it appears.
Two: measure it. Critical thinking becomes a serious institutional priority the moment it becomes a serious institutional metric. Right now, most organisations measure output. AI has made output cheap. What used to correlate with skill no longer does. The organisations that will still be functional in five years are the ones already redesigning the measurement.
If your workplace is running an AI adoption programme without a critical-thinking curriculum next to it, the programme is producing exactly the outcome the studies describe.
Uncertainty ledger
- The 600-person study is observational. Correlation, not causation. The MIT study is stronger — it's month-long and interventional — but it is one study, and small.
- We don't know the size of the effect at population scale, only that the direction is consistent across independent research groups.
- We don't know how much of the "decline" is people getting worse versus people never having been very good to begin with, and simply being tested for the first time.
- The Estonia experiment is too new to have outcome data. The theory is sound. The results are not in.
- The AAC&U 90% figure measures faculty belief, not student performance. It is a signal about the profession's confidence, not a measurement of the students.
- Regional evidence outside the US and UK is thin. The Singapore, China, and Fiji data points are directionally strong but not deep. This story is under-reported in Latin America and Africa, and the picture there may look different.
Bottom line
Critical thinking wasn't a skill AI took from us. It was a skill most of us were never explicitly taught, and AI is the first tool fluent enough to make that fact impossible to hide. The response isn't to reject the tool. It's to build the habit the tool exposes we're missing — one question a day, asked honestly, of something a confident source told us. Small enough to survive Tuesday, done every Tuesday, is how the skill gets built. Everything else is a resolution.
Sources
- Guardian, Over-reliance on chatbots can diminish critical-thinking skills, study finds, 19 June 2026 — MIT study on GPT-4o. Tier 1.
- NPR, Most K-12 teachers say AI's impact on education will eclipse the internet or computers, 5 June 2026 — NPR/Ipsos poll of 545 K-12 teachers. Tier 1.
- WSJ, Fighting Brain Rot With Free ChatGPT (Future of Everything), 5 June 2026 — Estonia policy. Tier 1.
- AP News, The skills people still perform better than AI, according to workplace experts, 11 June 2026. Tier 1.
- Inside Higher Ed, We Have Never Taught Critical Thinking, 23 June 2026 — AAC&U 90% figure and the framing wedge. Tier 2.
- Inside Higher Ed, A Taxonomy of Harm for the AI Age, 10 June 2026 — cognitive surrender, EEG connectivity evidence. Tier 2.
- Space Daily, A 2025 study of over 600 people found a correlation between leaning on AI tools and weaker critical thinking, sharpest among the young, 18 June 2026. Tier 2/3.
- Forbes, The Rising Peril Of AI-Induced Cognitive Surrender, 16 June 2026. Tier 2.
- Asian Business Review, Hiring cools as employers chase AI skills — ManpowerGroup Q3 2026 (Singapore critical-thinking premium). Tier 2.
- PR Newswire, IT Talent Trends 2026 Report (Info-Tech Research Group), 9 June 2026. Tier 2/3.
- Cambridge Network, What we learned delivering FutureReady internationally: insights from a pilot in China, 2 June 2026. Tier 2.
- Fijivillage, Education system not producing workforce with critical thinking skills — Naidu, 10–11 June 2026. Tier 3, regional context only.
- Washington Post interactive analysis on political bias in AI chatbots, 24 June 2026 (referenced via secondary reporting). Tier 1 source, Tier 2 access.
