AI has scaling laws: more compute, more data, smarter models. We face the opposite: more AI, less thinking, gradual disempowerment.
Figure 1
Two curves, going opposite ways.
1. We were already slipping
The cognitive state of rich countries is, at best, plateauing.
- September's PISA results showed a 22-point drop in math since 2015. That's more than a year of school.1
- 35% of US adults can't use basic math to check whether a claim is true.2
2. AI didn't start this, but it's about to make it worse
More than half of US teens have used chatbots for their homework.3 The problem: today's chatbots are trained to please the person asking, through a method called RLHF. For a kid, pleasing means giving the answer. So kids don't struggle, and they don't learn.
A Wharton study showed it. Kids who practiced math with GPT-4 scored 17% worse on the exam than kids who never had it. A version that gave hints instead of answers did no harm.4 We need dedicated learning products. We wrote more about this in Why Sophie is not a chatbot.
At MIT, people who wrote essays with ChatGPT showed the weakest brain activity of any group, compared with writing alone or with Google. Worse: when they later had to write alone, their brains stayed under-engaged. They didn't bounce back right away.5 It is a small, early study, but it points the same way.
AI safety researchers call this anticipatory disinvestment.6 Once you expect AI to do a task, you stop learning it. That makes you more dependent, which justifies learning it even less. It's a loop.
3. Gradual disempowerment
The same researchers call this scenario gradual disempowerment.6 We don't lose control in one moment. We just hand over more and more, until we become strangers in our own world.
I just learned the word solastalgia: the homesickness you feel while still at home, because home changed around you. It was coined for environmental change, but it describes this feeling beautifully.
The researchers name signals to watch:
- Workers' share of the economy. 52.8% in the US this year, the lowest since 1947. It started falling before ChatGPT.7
- New content made by AI. About half of new articles on the web.8
- Decisions handed to AI. Nobody measures this yet.
- Emotional relationships with AI. 12% of US teens already turn to chatbots for emotional support or advice.3
The paper also says education needs its own signals. PISA, controversial as it is, is a good start.
It's only by rethinking our education systems that future generations will benefit from scaling laws: how we bring AI into the classroom, how we grade work a chatbot can do, which skills kids must still build on their own, and whether the AI they use is built to teach or to answer.
AI will keep scaling. Let's make sure our kids do too.


