How Do Normal People Actually Get Good At AI?
Description
A sequel to the AI-power-consumption episode, which argued the real bottleneck is adoption, not energy. This asks what efficacious adoption actually looks like for average people. Adoption is wide (39.6% of working-age Americans by late 2024, 76% of under-30s) but access is not value. The evidence shows real, proven bright spots — customer support (+14%, +34% for novices), a World Bank tutoring RCT in Nigeria worth 1.5-2 years of schooling, and accessibility (Be My AI) — concentrated among people climbing toward a skill. It also shows a jagged frontier: inside AI's range, big quality and speed gains; outside it, a ~23% performance drop. METR even found experienced developers 19% SLOWER with AI while believing the opposite — though a larger 2026 cohort cut that to ~4%, the curve bending in real time. The harms are real (cognitive debt; $186/month of workslop) but mitigable. The frame is Brynjolfsson's Productivity J-Curve: GPTs depress measured output until humans build the complementary know-how. The turn: that know-how is a public good — messy early adoption is collective R&D, waiting for best practices free-rides on a map only explorers draw, and the right response to the harms is better scaffolding and sharing, not abstention.
Sources & further reading (10)
- How are Americans using AI? Evidence from a nationwide survey — Brookingshttps://www.brookings.edu/articles/how-are-americans-using-ai-evidence-from-a-nationwide-survey/
- Generative AI at Work — Brynjolfsson, Li & Raymond (NBER w31161)https://www.nber.org/papers/w31161
- Navigating the Jagged Technological Frontier — Dell'Acqua et al. (SSRN)https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
- From Chalkboards to Chatbots: GenAI and Learning in Nigeria — World Bankhttps://openknowledge.worldbank.org/entities/publication/15e1ff08-15ae-4f7a-b2a8-d146e6c113ee/full
- Introducing Be My AI for people who are blind or have low vision — Be My Eyeshttps://www.bemyeyes.com/news/introducing-be-my-ai-formerly-virtual-volunteer-for-people-who-are-blind-or-have-low-vision-powered-by-openais-gpt-4/
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METRhttps://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- We are Changing our Developer Productivity Experiment Design — METRhttps://metr.org/blog/2026-02-24-uplift-update/
- Your Brain on ChatGPT: Accumulation of Cognitive Debt — MIT Media Lab (Kosmyna et al.)https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- AI-Generated 'Workslop' Is Destroying Productivity — Harvard Business Reviewhttps://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
- The Productivity J-Curve — Brynjolfsson, Rock & Syverson (NBER w25148)https://www.nber.org/papers/w25148
Script
Cold open
Most people now carry a tool that, used one way, hands them an expert tutor — and used another way, quietly makes them worse at their own job. Same tool. So here is the question almost nobody asks cleanly: what does it actually mean to use AI WELL — not more, not less, but well?
Frame
This is a sequel to our last question — whether worrying about AI's power bill is quietly anti-climate. We argued the real bottleneck is not energy, it is adoption: actually using the thing to push the frontier. Fine. But 'use it' is not advice. More than half of American adults have now tried these tools, and plenty came away with slop, distraction, or a vague sense they were cheating. So this episode is about the gap between owning AI and getting value from it — where the value is already real, where it backfires, and a stranger idea: that even the fumbling may be worth more than it looks.
How far has this spread — and does using it equal benefiting from it?
Start with how far this has spread. By late twenty twenty-four, about 39.6 percent of working-age Americans had used generative AI; among the under-thirties, 76 percent, half of them every week. At work, use climbed from 33 percent to roughly 41 percent in a single year. But adoption is not benefit. Having the tool open in a tab tells you nothing about whether it made your work better, worse, or just faster at producing more of nothing. The real question was never how many people use it. It is what separates the people it helps from the people it quietly hurts.
Where is the value already real, not promised?
So where is the value already real, not promised? Three places the evidence is hard to argue with. Customer support: in a study of 5,179 agents, an AI assistant lifted problems solved per hour by 14 percent — and 34 percent for the newest, least-skilled workers, by handing them the habits of the best ones. Education: a World Bank trial in Nigeria gave students six weeks of after-school AI tutoring and saw learning gains worth one-and-a-half to two years of school — among the most cost-effective programs ever measured, with the biggest jumps for girls. And accessibility: Be My AI now describes the visual world, instantly and free, to hundreds of thousands of blind and low-vision users. Notice the pattern: the clearest wins go to people climbing toward a skill, not experts already at the top — which is to say the model looks less like a genie and more like a teacher. One who knows more than you about most things, is confidently wrong in unpredictable places, and is worth the most when it makes you need it less.
Why does the same tool help one person and harm another?
Which is exactly why the same tool helps one person and harms another. Researchers studied 758 consultants and named what they found a jagged frontier. On tasks inside AI's range, the ones using it were 25 percent faster and produced work rated 40 percent higher in quality. But step outside that range — to tasks the model is quietly bad at — and it did not merely fail to help. Performance fell about 23 percent, because a confident wrong answer is worse than no answer. The skill, it turns out, is not 'using AI.' It is knowing where the edge of the frontier is — and that edge stays invisible until you have felt it.
So why do careful studies sometimes show AI making people SLOWER?
And here is the finding that should keep us honest. When METR ran a careful trial with 16 experienced open-source developers across 246 tasks, letting them use AI made them 19 percent SLOWER. The twist: they had expected to be 24 percent faster, and even afterward swore it had sped them up by 20 percent. The tool felt like a gain while it was a loss — the most dangerous kind of error. But watch what came next. When METR repeated it with a bigger group, 57 developers, the slowdown shrank to about 4 percent — well inside the noise. The curve was already bending, as the tools and the people using them got better.
What are the real harms of careless use — and are they disqualifying?
Before we romanticize any of this, the harms are real. An MIT Media Lab team wired up 54 people writing essays and found the ones leaning hardest on a chatbot showed the weakest brain activity, the least sense of ownership, and often could not quote a sentence they had just produced — what they called cognitive debt. And in offices, a survey of 1,150 workers found 40 percent had been handed AI workslop in a single month — confident, empty output dressed up as work — costing nearly two hours each to clean up, about 186 dollars per worker a month. These are the dead-ends: the dependency, the distraction, the slop dumped on a colleague. They are real. The question is whether they are disqualifying — or just the price of something.
So what does good use look like for a normal person, concretely?
So what does good use look like for a normal person, concretely? The studies all rhyme: the wins come when you are climbing toward something and can check the result. Learning a subject from scratch, like those students in Nigeria. Getting un-stuck — a first draft, a summary, a rough plan you will revise. Translation, and access for people the old interfaces shut out. Code for someone who could never write it. The losses come when you invert that — outsourcing the judgment you are supposed to be exercising, on a task you cannot check, in a domain past the frontier. Same tool, opposite outcome. The practical rule is almost embarrassingly simple: use it where you could catch its mistakes, and where catching them teaches you something.
If individuals get mixed results, why is mass adoption still rational?
Because step back, and there is a pattern economists have seen before. Erik Brynjolfsson calls it the productivity J-curve: every general-purpose technology — electricity, computers — first makes measured productivity DIP, because the gains are not in the machine. They are in everything humans have to reinvent around it: new skills, new workflows, new habits, none of which exist yet. The technology arrives in an instant; the know-how takes years. So a tool that slows an expert today and speeds a novice tomorrow is not a contradiction. It is a society part-way up its own learning curve — and the curve only bends because people are on it.
Turn
And that is what the whole 'is it worth it for me' framing misses. The map of what AI is actually good for does not exist yet. It is being drawn, right now, by millions of people trying things — most leading nowhere, some somewhere, a few becoming the practice everyone copies. Every dead-end someone hits fills in a square of that map. Which means the early, clumsy, often-wasteful adopters are doing unpaid research for everyone who comes after. The person who waits until the best practices are clear is free-riding on a map that exists only because other people refused to wait. There is no settled best practice waiting offstage — best practice is the residue of a million experiments, and someone has to run them. So the harms — the cognitive debt, the workslop — are not reasons to sit out. They are reasons to explore BETTER: to share what works, to scaffold the dead-ends so they stay cheap, to catch the distractions early. The failure mode is not using it badly. It is everyone waiting for someone else to figure it out.
Closer
So — back to using it well. For a normal person it is not a secret prompt. It is a posture. Use it where you are climbing toward a skill, not coasting on one. Stay alert to the edge of the frontier, and treat a confident answer as a claim, not a fact. Keep doing the thinking you are actually trying to keep — and treat it like a good teacher, the kind whose job is to work themselves out of a job. And when you find something that works, say so, out loud, to someone. Because the real question was never whether AI makes YOU more productive this afternoon. It is whether you are contributing to the map — or waiting on the people brave enough to get lost first.