Everyone thought schools would be forced to drastically transform by AI.
But three and a half years after ChatGPT's debut, education has barely changed.
In May of this year, Sam Altman returned to his alma mater, Stanford University, took the stage in the CS153 class, and admitted his mistake:
This was one of my prediction errors.
He also issued a stern warning: if we don't change, human thinking capacity will atrophy.
Not long ago, on the CS153 stage at Stanford, OpenAI CEO Sam Altman was asked: What do you think about education?
He paused: "I'm very worried. I thought it would have changed by now."
Altman appears at Stanford CS153 class discussing education in the AI era, admitting he underestimated the speed of the education system's transformation. (Source: Stanford Online)
Three and a half years ago, right after ChatGPT launched, Altman predicted that students would cheat for a year, then the entire education system would be forced to restructure itself, teaching people to think better than before.
However, three and a half years later, the script didn't unfold as Altman imagined.
On the AI side, it evolved from GPT-3.5, which could only write copy, to models that can disprove mathematical conjectures that stumped mathematicians for decades.
On the school side, they're still testing students with the same old things: rote memorization, standard answers, closed-book exams.
Homework, exams, theses... everything remains the same. Scouring the entire education system, he couldn't find a single significant structural change.
A man who bet correctly on the "Scaling Law" turned out to be mistaken about education.
He said this was one of his biggest prediction mistakes in recent years.
A person who constantly talks about Artificial General Intelligence (AGI) is actually anxious about the classroom.
What exactly is he afraid of?
He Thought Schools Should Have Changed Already
Rewind to November 2022, just after ChatGPT's release.
Back then, Altman's assessment was still optimistic:
The first year, students would use it to cheat and learn little; then the education system would reinvent itself, teaching courses much better than before.
According to his vision, teachers would assign projects that necessitated using AI, forcing students to use their brains more and come up with new ideas.
In 2024, he publicly expressed optimism: superintelligence would bring a personal tutor for everyone, education would shift from rote memorization to problem-solving, to critical thinking.
The result? AI advanced by leaps and bounds year after year, while education remained stagnant.
AI Outsourcing Is Hollowing Out Critical Thinking
This disconnect is what Altman is truly worried about.
He says if we continue to teach and evaluate students using the methods of the "pre-AGI world," it will not only render these methods ineffective but also cause people to "fail to learn how to think," leading to the gradual atrophy of critical thinking.
Outsourcing thinking to AI starts as just a shortcut.
But use it or lose it—the mental muscle responsible for independent thinking, like an unused arm, will subtly shrink and weaken, or as Altman puts it—muscle atrophy.
Is this just Altman's concern, or is it already a reality?
Some research shows that after ChatGPT entered the classroom, monthly test scores dropped by about 20% within six months; for high-stakes entrance exams that truly determine one's future, scores fell by 18% and 24% respectively, and this deficit takes up to two years to gradually become apparent.
More telling is an analysis from the University of California, Berkeley (UC Berkeley).
Among over 500,000 grade samples, grades in subjects like writing and programming noticeably shifted upward after ChatGPT, but the increase was entirely in homework scores, while exam scores remained unchanged.
UC Berkeley analysis of over 500,000 grades: After ChatGPT's release, the proportion of A and A- grades in writing and programming courses significantly increased (blue significantly positive), while B+ and below barely moved. (Source: Chirikov/CSHE)
Why? This is "outsourcing," not "learning."
Another study spanning ten years and covering millions of US math interactions points to the same conclusion: with the arrival of chatbots, problems are solved faster, but less is learned.
Homework submissions are getting more polished, but minds are getting emptier.
Where Is the Promised Renaissance in Education?
Altman isn't the only one puzzled.
OpenAI technical team member Ryan Brewer posted, saying it's shocking that large language models haven't sparked an educational renaissance:
Shouldn't I be able to learn a language in a month? Where did we go wrong?
Similar doubts spread rapidly on X: With the most powerful learning tool in history in hand, why haven't AI tutors entered every household? Why is the education revolution so delayed?
The answer lies not in technology, but in institutional inertia.
The university's evaluation system—exams, theses, homework—has been built over centuries on an implicit premise: these tasks are too time-consuming for anyone to take shortcuts.
AI has changed that premise.
But schools are still using pre-AGI standards to measure a new generation raised with AI, even as the first generation of ChatGPT natives have already graduated.
Tools update with a version number; systems update over a generation: the technology is already here, but the rules remain in the previous era.
A 24/7, tireless, personalized, almost-free AI tutor could theoretically be given to every child today.
But it hasn't arrived, and the real reason is the pace at which the education system restructures itself.
In the same speech, Altman offered another prediction:
It's been three and a half years since ChatGPT appeared. Even if AI merely progresses along the same curve for another three and a half years, what human society can do will be on a completely different scale than today.
As technology races ahead exponentially, the gap between it and education will only widen, ultimately to be filled by the students currently sitting in old exams, old homework, and old evaluation systems.
The skills they learn may be taken over by AI the moment they leave school; the judgment they fail to develop may be difficult to recover in a lifetime.
The debt incurred is a generation's "cognitive deficit."
If Machines Can Write, Why Must Humans Still Learn?
So what should we still teach?
Altman's answer is somewhat counterintuitive: Some things, even though machines can do them better, humans must still do them by hand.
He gave his own example.
He said he's the kind of person who "thinks by writing," producing a lot of text never meant for anyone else, just to clarify a problem, and he's grateful he learned to write.
Programming is the same; AI can generate code in a second, but the process of building logic by hand trains the brain.
In essence, writing and programming are like math proofs in the calculator era: the machine already knows the answer, but we still have students derive it themselves. The goal isn't the answer behind the problem, but the meta-skills of "thinking" and "learning," and writing and programming are the tools to train them.
Following this logic, Altman advocates shifting the goal of education from "remembering more knowledge" to "asking better questions"; from testing memory to testing judgment, creativity, and genuine interdisciplinary skills.
And the root of the problem lies precisely in the evaluation system.
What do today's exams still test?
Memory, standard answers, completing tasks alone and closed-book. These three things are exactly what AI is best at and can most easily do for you.
When schools still evaluate students by "who remembers more, who answers more accurately," AI has turned "remembering more, answering accurately" into a zero-cost commodity.
How much meaning remains in numbers measured by a ruler that AI can easily pass, when used to gauge the abilities of the next generation?
This is what truly worries Altman: It's not most crucial whether students use AI; what's crucial is whether they can *verify* AI.
More concerning than over-reliance on AI is using AI without knowing how to verify it, accepting the machine's output at face value.
If this inertia persists for another three and a half years, a generation will gradually lose the training ground for independent thought, only to realize too late: they've almost forgotten how to think for themselves.
References:
https://www.youtube.com/watch?v=F_7M4Hc-usM
https://x.com/hesamation/status/2073884828861071557
https://x.com/ryanbrewer/status/2073812031988535760
This article is from the WeChat public account "Xin Zhi Yuan", author: ASI Revelations
