Free Thinking
Where AGI fits in global development
It seems likely that within a year, AI will be able to do much of the everyday expert work of development, like diagnosing, tutoring, counselling and auditing, as well as a good expert, at almost no cost. I became interested in the question of what this means for the world's poorest countries: if we take seriously the idea that this "free thinking" will be here soon, what should we be building for development?
Much of the work on AI in development I've read starts from a service that already exists and asks how AI can make it better or reach further. This is important work. But I wanted to start from scratch and ask: which of development's biggest problems are really a shortage of expert work? And if we add free thinking, how much impact can we have?
AI already beats experts at some expert work, and it's improving fast
Boston: diagnosis at triage
% exact or very close diagnosis, reading the records of 76 real emergency cases
Rwanda: community health workers’ questions
Mean score out of 5, rated blind by senior Rwandan clinicians
In teaching, we're starting to see similar results.
What students learned in an hour
Mean gain in test score, percentage points
What each point of learning cost
US dollars per percentage point gained
And the technology just keeps getting better.
METR says measurements above 16 hours are unreliable with its current tasks.
So I think we should plan for a near future in which expert-level work is close to free, rather than for what today's AI can do. There is obviously a lot that AI can't do yet. But I think the future worth building for is one where AI does the expert's work for everyone who needs it.
Development's biggest problems
So, say you agree that free thinking will be here soon. The next question is, "what problems in development can this really help?"
I was interested in taking a first-principles look at this question, and not just turning up the dial on the (important!) "tech for good" work that already exists. So I had Claude read 172 global development reviews and country diagnostics. It extracted 218 barriers to development and wellbeing, grouped them into kinds of work, then asked a series of questions to see how much impact intelligence on tap can have on the problem.
172 reviews and country diagnostics name 218 barriers to development and wellbeing.
They cover health, learning, growth and public goods. Hover over a dot to read one.
Grouped by the work it would take to remove them, they make 46 kinds of work, in six areas.
Teaching children, collecting tax and diagnosing illness are examples. The areas are education, health, farms, firms, markets and the state.
In 15 of the 46, the missing work holds back development or wellbeing a lot.
For each kind of work, I had Claude weigh the evidence on what the missing work costs, in lives, learning or growth, and score the gap big, medium or small against one rubric. Bigger dots are bigger gaps. Click one to see the reasons and sources.
Free thinking supplies expert work, but expert work is just one factor of production.
So I asked three questions about every kind of work, to understand how much impact unlimited expert work might unlock.
1. How much of the work is thinking?
More expert work helps most where the work is judgement, analysis, or conversation, like therapy and tax auditing. It helps little where the work is physical, like building infrastructure.
2. Are the other inputs there?
Expert work only helps if the money, roads, supplies and legal powers it needs exist. Where something else is missing, that limits how much more expert work can do.
3. Will it get adopted?
It won’t be where someone profits from the gap, or where the people who’d have to act on it don’t want to. On these three questions alone, more expert work wouldn’t close any of the 46 gaps by a lot.
But a well-designed service can get around some of these limits.
We can create new channels to reach users, build in missing supervision, or work for different user groups. I judged that it could in 7 cases.
With a service designed that way, free thinking could close some of the big gaps.
They are teaching, mental health, drug discovery and tax. It would partly close the gaps in clinical decisions, agricultural R&D and finding buyers, and most of the smaller gap in courts. But most of the big gaps that hold back development and wellbeing, such as building infrastructure, land records and migration, would barely move.
Most of these services won't get built by the market.
The kinds of work still in colour have no payer with the means and a reason to build them for poor countries.
It's important to acknowledge that for most of the biggest things holding poor countries back, powerful AI is unlikely to make a big difference. Things like building infrastructure, land records, or migration are much more constrained by money and politics than they are by a lack of expert work. But there are places, across society and the economy, where I think we can have a lot of impact if we increase our ambition and build for a future with unlimited expert work.
of the deaths health care could prevent in low- and middle-income countries are among people who reached care but got poor care: 5.0 million a year, against 3.6 million who got none.5
ten-year-olds in low-income countries can’t read a simple text.6
of people with depression in low- and lower-middle-income countries don’t get even minimally adequate treatment.7
of the VAT owed in low-income developing countries isn’t collected.8
What it would make possible
In each of the seven kinds of work below, free thinking could supply the missing expert, for everyone, rather than a better tool for the few experts there are.
These are sketches of what's possible, not plans.
Across all seven areas, much of the technology work is already happening, and building a great product is only going to get easier with time. Which raises the question... where should an organisation which wants to build for this future spend its energy today?
Next in the series: II. Build What Stays Hard →
Sources
- Brodeur et al. (2026), Performance of a large language model on the reasoning tasks of a physician, Science 392:524-527
- Rutunda, Williams, Kabanda et al. (2026), Large language models for frontline healthcare support in low-resource settings, Nature Health
- Northcutt, Hasmani, Feng et al. (23 September 2026), StudentBench: AI and human tutoring yield equivalent GRE learning gains, arXiv preprint
- METR (updated 8 May 2026), Task-completion time horizons of frontier AI models, Time Horizon 1.1
- Kruk et al. (2018), Mortality due to low-quality health systems in the universal health coverage era, Lancet
- World Bank, UNESCO et al. (2022), The State of Global Learning Poverty: 2022 Update, Annex 4
- Moitra et al. (2022), The global gap in treatment coverage for major depressive disorder in 84 countries from 2000-2019, PLOS Medicine
- Barra and Prokof’yeva (2025), The Reverse Method: A New Indirect Technique for Global VAT Compliance Gap Monitoring, IMF Technical Notes and Manuals TNM/2025/15