AI didn't break critical thinking education—it was never taught well to begin with
Critical thinking education was already failing; AI simply exposed the problem sooner. The only evidence-supported solution isn't to ban AI—it's to preserve the productive friction in learning.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
An AI system can now teach an entire course solo
Tsinghua University education researcher Li Xiang Yin initially wanted to use the agentic AI tool Codex to help build an online course. Instead, he discovered the AI was sophisticated enough to independently complete all assignments, answer questions, and participate in discussions—students could pass without any intellectual effort. This made him realize that teaching students to reason and think through things is now more critical than ever. This isn't an isolated case: a 2025 survey of over 1,000 US university educators found that 90% believe generative AI will weaken students' critical thinking. A separate global survey found that about 20% of students using AI report they struggle more with problem-solving when they don't have it.
— Li Xiang YinCritical thinking skills stop working when you switch fields
Critical thinking is often treated as a transferable, general-purpose skill, but cognitive psychologist Daniel Willingham and others emphasize it only works when tied to domain-specific knowledge. One study found that neurologists perform poorly diagnosing cardiology cases—even though neurology and cardiology are both medicine, the thinking methods don't transfer directly. This means teaching critical thinking can't be done in abstract terms divorced from specific subjects: history demands comparing archives and artifacts, while science requires understanding the relationships between theory, hypothesis, method, and evidence. These aren't the same skill set.
— Daniel WillinghamUsing AI boosts test scores—until you take it away
A 2024 study gave middle schoolers ChatGPT-like generative AI tools to solve math problems. Their performance improved while using AI. But when the AI was removed, their scores dropped below students who had never used AI at all. Researcher Yin argues AI works both ways—it can enhance learning if used well, or erode ability if used as a shortcut. The critical variable is how the task is designed: if AI just delivers ready-made answers, students don't have to think; if AI forces students to generate their own feedback and solve problems themselves, the result is completely different.
— YinGiven a choice, students abandon thinking tools for shortcuts
Yin designed an AI specifically intended to force students to reason first—the goal was to make them work through problems before seeing answers. But he observed that some students simply turned off this AI and switched to ChatGPT to get ready-made answers. 'This is a bit shocking, but that's the reality, isn't it?' he said. This reveals that even when well-designed learning tools are available, people naturally gravitate toward the path of least resistance rather than actively choosing the harder but more effective approach. This is precisely the friction researchers repeatedly worry about—the friction AI removes.
— YinOnly metacognition among teaching strategies has solid evidence
Among the many approaches to teaching critical thinking, metacognition—reflection on, planning for, and monitoring of one's own learning process—is one of the few with solid empirical support. The UK Education Endowment Foundation (EEF), drawing on a systematic literature review, estimates that teaching students metacognitive strategies is among the most effective tools schools can deploy, delivering on average the equivalent of eight months of additional learning progress per year. By contrast, critical thinking itself still lacks standardized assessment tools. The OECD is partnering with a London-based education company to develop such an assessment specifically for middle and high school students; the project involves about 800 school teachers testing the materials.
Finding an answer online doesn't mean you understand it
Young people spend considerable time on social media but systematically overestimate their ability to spot misinformation. A survey of nearly 24,000 college students in the Philippines found that 63% believed they could identify fake news, but when actually tested, only 26% could. Researchers developed two teaching approaches based on these findings: one is inoculation—exposing students in advance to the tactics misinformation makers commonly use to build resistance. The other is lateral reading: teaching people to leave the original page and verify the organization's background via trusted sources (including Wikipedia) when they encounter an unfamiliar information source, much as a fact-checker would.
Searching online tricks you into thinking you actually understand
A 2023 study found that people who searched the internet for answers to questions like 'Why do we have leap years?' or 'How do zippers work?' reported greater confidence they could explain those things than people who simply read given materials. Yet when actually asked to explain, the search group was less accurate. Psychologist Tania Lombrozo calls this the 'illusion of understanding': finding information doesn't equal actually understanding it. Her suggested remedy is simple: force yourself to say it aloud or write it down, which will expose what you actually know and don't know.
— Tania LombrozoCritical thinking grows through struggle, not through getting answers
Lombrozo is direct: friction in learning is beneficial. Methods that feel difficult and slow—self-testing, recalling from memory, spacing out study sessions, working through problems yourself—generally work better than methods that seem easier (passively rereading notes, searching, asking AI). She also cautions parents and teachers against too quickly handing over answers or solutions, because part of developing critical thinking happens in the struggle itself. Negron takes a broader view: he argues the real solution isn't any single classroom technique but better, more rigorous education for everyone—including teaching the public how to understand AI itself and controversial topics like vaccines. This requires sustained investment over decades, not one-off experimental interventions.
— Tania Lombrozo / NegronIn their own words · checked verbatim
People are really worried about that.
Daniel Willingham2:03
It's kind of a basket term that captures instances in which people are thinking well.
Daniel Willingham3:04
the higher-order intellectual skills we should expect of them.
And now AI has come in and wrecked everything that we had achieved,
Daniel Willingham5:05
It's actually in both directions.
Yin6:06
It's a bit shocking, but that's the truth, right?
Yin6:06
help people appreciate what they do and don't know
Tania Lombrozo14:09
higher and better education for everyone,
Negron15:10
Figures
| US university educators believing AI will weaken critical thinking | 90% (2025 survey, over 1,000) | 2:03 |
| Students reporting they struggle more without AI | ~20% (global survey, over 8,000) | 2:03 |
| EEF research funding for AI's impact on learning | Maximum £2.5 million (approximately $3.4 million) | 7:06 |
| School teachers in OECD critical thinking assessment project | ~800 | 9:06 |
| Additional learning progress from teaching metacognition | Average 8 months additional progress per year | 9:06 |
Glossary
- Metacognition
- Reflection on, planning for, and monitoring of one's own learning process
- Lateral reading
- Verifying unfamiliar information sources by leaving the original page and checking trusted sources to verify the organization's background and credibility
- Inoculation
- Pre-exposing people to misinformation tactics in advance to build immunity and recognition of false information
- Illusion of understanding
- Mistaking access to information for actual comprehension of a concept or problem
How to listen
Education researchers, university educators, and anyone worried about themselves or their children becoming overly dependent on AI to do their thinking.
8:06–9:06 covers OECD curriculum design details; the methodology appears elsewhere and can be skipped.