Who can get AI to follow them?

AI is already shaping the future of work, and I’m curious of how that will affect developing countries. Last time, I talked about how it will affect the returns to education. Today, I’m reviewing some research relevant to who will be able to get AI to help them at work.

In my previous post, I said the following about education and working with AI:

As AI advances, the winners will be those who understand how to use AI to achieve increasingly complex tasks. Indeed, this divide has already begun. We're already using AI to do simple tasks. Analyze this memo. Write a Python script that does a specific function. But complex tasks require the human and AI to work together. And by work together, I mean the human needs to find out how to heckle the AI until it does exactly what he wants it to do, and he needs to discern when the AI has not yet achieved it. I'm hypothesizing that users with more education will be better equipped to do this. And when they do this, they will be significantly more productive, which will bring them higher wages.

Playing cooperative games

The NBER paper by Ben Weidmann, Yixian Xu, and David Deming is titled Measuring Human Leadership Skills with AI Agents. They look at how well people did at leading teams in a cooperative game. A cooperative game is one where all players work together to beat the game itself, rather than competing against each other. Everyone wins or loses together, usually by achieving a common goal or surviving against challenges controlled by the game. The key to these games is that players have different clues, and no individual has sufficient clues to solve the game on their own. Thus, it takes a leader to coordinate players to contribute their clues.

This game is a standard way to measure leadership, but the authors added an interesting twist. Sometimes they assigned leaders to teams of human players, and other times they assigned them to teams of AI players. The authors’ goal was to look at whether the factors that correlate with strong leadership among humans are similar to those for leadership among AI.

But I am only interested in the factors for leading AI. So here are a few points that stood out to me.

1. Education is more important for leading AI

Below is the figure where they compare the correlates of leading AI to the correlates of leading humans. The 45-degree line indicates that the two factors the same across team type, but if the factor is below the linethe factor was a stronger predictor for good leadership with AI teams.

One thing you’ll notice is that education is below the line. In fact, it’s negatively correlated with leading humans but positively correlated with leading AI. I don’t want to oversell this results for two reasons. (1) The magnitudes of the correlations are pretty small, among the smallest in the factors; (2) the sample did not have a lot of variation in education (about 75% had attended university) so I don’t think it’s actually designed to pick up differences in education. But education being more helpful in leading AIs is consistent with my hypothesis, and I would love to see future research diving into this to see if it’s a true difference.

2. Education correlates are important for leading AI

If you look at the biggest correlates of leading AI, a lot look like the kind of things that improve with more education: decision-making and task-skill being the two most obvious. Like I said in the previous post, getting AI to work for you is a complex task, and the benefit of education is developing the ability to handle complex tasks.

Implications for developing countries

I don’t want to put too much weight on this one study, but it should move the needle a little on how you see the relationship between education and AI productivity. If AI is a complement to higher levels of education, then it should also increase the returns to those high levels of education. We need to prioritize getting students in developing countries up to that level.