AI and development · Decision-making under uncertainty · Government
Developing countries can’t wait for certainty about AI
Last week, I joined the “Governing with AI in Low- and Middle-Income Countries” conference at the LSE, convened by International Growth Centre, the LSE Data Science Institute, and the The World Bank Group. Bravo to all involved (in particular Shahrukh Wani and Sharif Kazemi!), and it was great to finally meet Joanne Ke Edelman, Christian Meyer, Oliver Hanney, Alexander De Ville and more in person + see many familiar faces.
To no one's surprise, I believe this is the defining issue of our day. It’s great that major institutions in development are now (finally!) actively engaging with this issue, including with lots of data and evidence. However, in periods of rapid change and high uncertainty, evidence should not be an excuse for averting our eyes from what’s right outside our windows. By definition, we do not have data about the future. During Covid-19, rigidly insisting only on “gold standard evidence” and under-relying on forecasting early on actively harmed us. It was like asking for clinical trials for why we should not jump out of flying planes without parachutes.
I remain really concerned that LMIC governments and development actors are poorly prepared for the next 5-10 years. In climate change, we listen to experts on climate science and forecasting, and we plan accordingly within our respective fields and expertise. Why not in AI? Today’s publicly available AI capabilities are (i) vastly under-recognised and under-utilised by corporations, governments and development actors, (ii) behind what is available to frontier labs behind closed doors, and (iii) increasing incredibly rapidly, with no end in sight. The high uncertainty around the next 5-10 years in AI therefore means we must be clear-eyed about where we are already and think ahead, planning across multiple plausible scenarios and course-correcting based on an accurate picture of the world as we go along.
For that reason, I asked two of the panels essentially the same question: What is different about this conversation from the one we would have had in 2021? And, given the uncertainty and pace of change, how should we think about speed and agility in governing and leveraging AI to benefit the world's poorest people? I ask this question repeatedly because I see a vast gulf in global development and LMICs more broadly: a few people live in the age of frontier AI and use today’s tools curiously across every part of their work, and plan for AI capabilities to keep racing ahead. But most folks in government and in development seem to be quite sure that nothing much will change, and see AI as simple chatbots that will change the world only very slowly and in specific vertical “use cases” in health and agriculture.
In the spirit of constructive disagreement emphasized by the LSE, I think this assessment is wrong. I see a very real probability that major change will come to LMICs and will come between 2030-2035 at the latest, and we should actively plan for this. More below ⬇️
As Parth Ahya noted, if you had seen today's AI five years ago, you would have called it magic. We get used to a lot very quickly, yet rewiring workflows to harness the magic takes years (due to slow diffusion, poor incentives, lacking training, group dynamics, etc.). But governments and development agencies are slow-moving beasts facing long lead times, so often it takes us 1-3 years from planning to delivery - and yet it’s still only been ~1300 days, or 3½ years, since GenAI was released to the general public. You see the problem! How can we do better here?
Friederike Schüür asked the right question: which lessons are the same from 10-20 years ago, when we first talked about AI and machine learning, and what has changed? I would add: are there lessons from back then that we carry with us, but which no longer serve us? I will volunteer one outdated lesson to drop: “AI is very expensive, complicated, and technically demanding.” Actually, and thankfully, general-purpose AI, LLMs and agents today are much simpler to use and implement out of the box than previous waves of AI and technology. The narrow machine learning AI we have used for 20 years does not run off-the-shelf as a commodity and is much harder to deploy, with much greater prerequisites.
So, how DO we move forward?
💡 Invest time to learn and be up to speed with where we are today and where we are going. Don’t let yourself be surprised at present and future developments that are relatively easy to forecast today simply by looking at existing trends and extrapolating them forward within reason (this goes for AI and for lots of things outside AI!). Mainstream media coverage of AI is improving, but still not great, and we need much more of a translation layer into where global development and LMIC decision-makers get their news and advice.
💡 Don't overcomplicate it. Just get started. No one would try to run a government or organization without computers, emails/messaging, or spreadsheets. Frontier LLMs are a foundational organizational commodity and access and use are bare minimums at this stage (not least to monitor how fast capabilities change…). Securing AI is not radically different from our email or word processors, where we all already input lots of sensitive information. Overcomplicating this actively harms resource- and time-constrained governments - and customised solutions using internal data can always come later.
For those of us advising governments, foundations, development agencies, and nonprofits:
💡 Offer an immediate painkiller, not another multivitamin to remember to take. AI and LLMs work out of the box with the right training and advice, with very few pre-requisites. At the same time, many bottlenecks and missing “foundations” for AI only become visible when doing the thing that will actually deliver public value. Starting with long diagnostics and foundation-building exercises burns political capital, often wastes time and resources, and in any case begs the question of which AI foundations to start with. False sequencing is a huge risk here: governments must “learn by doing” while delivering a portfolio of AI initiatives to build the muscle and continuous pipeline for their next deployments, with low-risk first wins and public value delivered in parallel with building AI foundations.
This is not to sing AI's praises. The AI labs are commercial actors and partners that governments and development actors need to negotiate hard with, like we do with pharma companies to get access for the poorest. But AI is coming, whether we like it or not, and we need to be ready. Before we can agree on what to do about AI in LMICs and in development, we have to have an accurate view of where it is today, and where it's going.
I have much more to say about this, and I’m actively working on projects on what to do for LMIC governments if we get transformative AI in the next few years. But I will leave that for another post! (Thanks for reading, if you got this far!)
The post continued in my comments under it on LinkedIn, which I’ve added here in order. Original post and comments ↗