Olu Online

RE: “The robots work at night”: How AI helps the world’s poorest, and where it falls short

Last updated 10 minutes ago

This is a reply to one of GiveDirectly's blog posts, and also the accompanying video. I'd advise you read and/or watch before you engage with this1.

The tl;dr of this piece is pretty simple: putting AI tools in the hands of the world's poorest, to ask questions such as “How to make a profit in three months?”, “What is the number of neighbors I should be cautious about?” and “My child is in trouble, what should I do?” (real questions from the blog post!) instead of just directly giving them more money (the raison d'etre of a group called GiveDirectly you'd assume!) is misguided at best and harrowingly awful at worst.

GiveDirectly says it gave users education on the tools they were using. I can also be charitable and say that these are mostly idle questions, questions to see what the robot thinks, not to be taken wholesale as The Answer. However, rampant misinformation/"hallucinations", gaps in the data covering their villages and lives and context, and the way LLMs work in the first place means that the most kind thing I can think to say is that this is a strange experiment.

I'm unsure of the utility of this kind of experiment. If one of GiveDirectly's core beliefs is that people in poverty know what people in poverty need, have people in poverty requested this? They reach people in countries like the US where they would have more exposure (if not necessarily more understanding) of the technology due to more people speaking in English; why not run a pilot there, if it had to be done at all?

To pull from the article again, if the right information at the right time is so important, doesn't GiveDirectly have experts at its fingertips (or sure, at the other end of its chequebook) who could create online resources that are definitely not making up information to compile useful information or pointers? If it has to be online or a chatbot, old style "a bunch of if statements in a trenchcoat" AI would make a lot more sense. If "traditional" training programs are needed but inflexible, as the piece suggests, offer something flexible and untraditional? Even offering these solutions seems outside of the remit of a wealth redistribution group, unless they are being directly asked for it.

I'd be remiss not to mention GiveDirectly's effective altruism links, given all the AI, but you can decide whether you care about that on your own.

if you liked this post, please:

|

  1. As ever, I'm not your mum or your boss so do what you like, it'll just make more sense.↩

#ethics