Happy Friday, and welcome back to another Human Prowess article.
On October 7, I had the chance to co-moderate a panel on AI literacy at Johns Hopkins in Baltimore. We discussed how these systems actually work, ethical and responsible development, and what waits for undergraduates when they move into the workforce. I wanted to push the conversation toward something I think deserves more attention: are we actually preparing students for greatness?
I mean the capability that belongs to a person and stays with them when the tool is no longer there. Being able to generate an impressive answer is useful. Being able to explain it, challenge it, and take responsibility for what happens next requires something from YOU. If intelligence is becoming abundant and commoditized, what are we doing as a society to make sure human potential does not become increasingly neglected?
That's the question I want Human Prowess to lead on. The evidence is, frankly, messy.
What can someone do when the tool is gone?
Researchers ran a randomized trial with nearly 1,000 high school math students in Turkey. Students using a basic GPT-4 chatbot scored 48% higher on assisted practice than students with textbooks alone. Then the tool was taken away for an exam, and the chatbot group scored 17% lower than the textbook group. Both figures are relative differences in scores (not percentage point changes). [1]
I think that distinction should bother anyone responsible for teaching another person. The work looked better while the help was available, but the students did worse when they had to perform independently. If we only evaluated the practice, we could have celebrated an intervention that undermined the very thing we wanted to develop. Performance is not learning, and our assessments need to account for that.
FROM THE LITERATURE. The design of the help matters. Bastani and colleagues also tested a GPT-4 tutor with safeguards intended to protect learning. Its students scored 127% higher on assisted practice, but their later exam performance without AI was statistically indistinguishable from the textbook group. The safeguards largely removed the harm; they did not produce a demonstrated independent learning advantage in this experiment. [1]
Bastani et al., "Generative AI without guardrails can harm learning," PNAS, 2025.
Other research gives us reasons to pay attention without pretending the question is settled. An MIT Media Lab preprint followed 54 participants through three essay writing sessions, with 18 completing a fourth. The AI group showed the weakest EEG connectivity, reported the least ownership of their essays, and struggled to quote their own writing. That's a small study of a particular writing task, still a preprint; an EEG result is not a diagnosis of a generation losing its intelligence. [2]
A study of 133 undergraduate art students found that AI use was associated with stronger perceived support and skill transfer, without a corresponding improvement in assessed artistic performance. A Microsoft Research survey of 319 knowledge workers found that greater confidence in AI was associated with less reported critical thinking. Neither establishes that AI causes a general decline in ability. They do make me question how much weight we put on feeling productive or supported when we haven't checked what a person can actually do. [3, 4]
With that being said, I have to challenge my own framing. In a randomized Harvard physics study involving 194 students, a purpose-built AI tutor produced more than twice the median learning gain of an active learning classroom, measured through immediate tests around two lessons. The tutor was designed around teaching practices, feedback, and a sequence of problems. That is a very different interaction from handing someone a chatbot and asking them to finish the assignment. [5]
I want AI in education when it helps someone develop. These studies give us no excuse to treat every use as equivalent, or to assume that getting the assignment done means the student learned it.
The test of learning is what you can still do after the help leaves the room.
Have we actually set the next generation up for success?
Scott Galloway's April 2024 TED talk, "How the US Is Destroying Young People's Future," puts the economic position of young Americans at the center of the conversation. His argument challenges the priorities of the generations holding the most wealth and power. I want to direct that same scrutiny toward education: if you love your kids, what opportunities and capabilities are you actually helping them build? Affection alone doesn't answer that question. [6]
The research behind the generational concern is substantial. Raj Chetty and his coauthors found that about 90% of children born in 1940 earned more than their parents, compared with about 50% of children born in the 1980s. They compared inflation adjusted household incomes at age 30. Those figures describe specific birth cohorts, rather than a current forecast for every student, but the change is difficult to dismiss. [7]
If you're twenty or thirty years older than me, I'm genuinely asking: have we set my generation up for success? And for people my age, what are we going to do with the responsibility that is already ours? I don't think blaming the generation above us is a particularly useful life strategy. I do think institutions with authority over our education should have to explain whether their version of success leaves us more capable.
We need to be just as careful about blaming AI for everything happening in the labor market. Stanford researchers report that employment among workers aged 22 to 25 in AI exposed occupations stands 19% below where it would be if it had kept pace with less exposed peers. They describe this as an early, descriptive indicator, rather than a causal estimate. Meanwhile, research linking surveys of 25,000 Danish workers to administrative records through December 2024 found no detectable effects of chatbot adoption on earnings or hours. Different settings and measures leave us with a real question, not a universal verdict. [8, 9]
Schools and employers still have choices to make while that evidence develops. Are students getting opportunities to defend an argument, work through uncertainty, and own a decision? Are junior employees receiving the feedback that used to come from working through a difficult task? Saving time is valuable; I want to know where that time goes, and whether any of it returns to developing the person.
Can we improve ourselves as deliberately as we improve the tools?
The idea of recursive self improvement in AI is useful here, provided we keep the comparison honest. In 2025, researchers introduced the Darwin Gödel Machine, a coding system that modifies its own code and tests changes against benchmarks. Its reported SWE-bench performance improved from 20.0% to 50.0%. That result belongs to a specific research system and coding benchmark; it tells us nothing about a machine acquiring every form of human judgment. [10]
What interests me is the feedback. There is an attempt, a test, and evidence about whether the change helped. We learn, develop, adapt, and iterate too, although human development includes relationships, values, and responsibilities that a coding score cannot capture. The comparison becomes useful when it makes us examine the quality of our own feedback.
If I submit something AI helped produce and receive a good grade, what exactly improved? The result may be excellent, and I may still be unable to explain how it works. A better learning process would expose that gap early enough to do something about it. Otherwise, we keep improving the artifact while leaving the person largely untested.
I would start simply: attempt the problem before asking for an answer, use AI to question the reasoning or give a hint, then close it and explain the decision. Come back later and see whether you can do a similar task independently. The point is to make development observable. A better output should lead us to ask what changed in the person who produced it.
What does betting on people require of me?
I am a student asking my own institution to prepare people for greatness, so I belong inside the question. I don't get to demand independent thinking from a university and then excuse myself from doing it. If I want capability that stays with me, I have to be willing to discover where it is missing. That can be inconvenient when the assignment is due and a complete answer is available immediately.
Betting on people requires us to spend time on their development, including our own. Schools need assessments that reveal what students understand. Employers need to give people opportunities to build judgment, even when a faster route exists. And I need to judge my own work by what I can explain and take responsibility for after the screen closes.
This week, choose one task you usually hand to AI. Do an honest attempt yourself, use the tool for feedback, and then try again without it. What can you do on the second attempt that you couldn't do on the first?
Forward this to someone teaching, learning, or hiring, and have that conversation with them.
See you in two weeks,
Jaron
The minds we pulled from this week:
1. Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS. Read source
2. Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv preprint. Read source
3. Generative AI in creative education: baseline-linked evidence on learning processes, artistic performance, and authorship. (2026). Frontiers in Psychology. Read source
4. Lee, H.-P., et al. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. CHI. Read source
5. Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports. Read source
6. Galloway, S. (2024). How the US is destroying young people's future. TED. Read source
7. Chetty, R., et al. The fading American dream: Trends in absolute income mobility since 1940. Opportunity Insights research page. Read source
8. Brynjolfsson, E., Chandar, B., & Chen, R. Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab, current revision. Read source
9. Humlum, A., & Vestergaard, E. (2026). Still waters, rapid currents: Early labor market transformation under generative AI. Brookings. Read source
10. Zhang, J., et al. (2025). Darwin Gödel Machine: Open-ended evolution of self-improving agents. arXiv. Read source