Salesforce's economic simulations have created some very unexpected tax models
In environments that look a lot like a video game, Salesforce is running simulations of miniature economies to see whether AI can develop novel approaches to taxation.
It calls its automated policymaker the AI Economist. Here’s how they describe it:
The simulation uses a two-dimensional world. There are two types of resources: wood and stone. Resources are scarce: they appear in the world at a limited rate. Workers move around, gather and trade resources, and earn income by building houses (this costs stone and wood). Houses block access: workers cannot move through the houses built by others. The simulation runs this economy over the course of an episode, which is analogous to a “working career.”
“In our simulations, AI agents pay taxes, while the AI economist recommends how to set taxes,” a developer says in the video above. “It also advises how to subsidize and redistribute wealth.”
The simulations themselves involve AI agents trained to look out for their own interests. In each, the laws themselves are made to replicate various economic models like free markets or those promoted by well known economists like Emmanuel Saez.
In Salesforce’s simulations, their AI Economist outperforms other models in terms of a balance between economic equality and productivity. The development team claim that it did even better when compared with a simulation based on free market economics using the US tax system.
What makes Salesforce's simulations interesting is the introduction of AI as policymaker. In past economic simulations, only citizenry were studied. In these, the policies themselves are also the result of machine learning.
The results are odd, as one might expect. Agents earn coins, and are taxed based on their income.
From MIT's analysis:
Unlike most existing policies, which are either progressive (that is, higher earners are taxed more) or regressive (higher earners are taxed less), the AI’s policy cobbled together aspects of both, applying the highest tax rates to rich and poor and the lowest to middle-income workers. Like many solutions that AIs come up with—such as some of AlphaZero’s game-winning moves—the result appears counterintuitive and not something that a human might have devised.
Finally, what happened with the taxes themselves? Salesforce eplains:
The collected taxes are redistributed evenly among the agents. In effect, the lower-income agents receive a net subsidy, even though their tax rates are higher (before subsidies). In other words, under the AI Economist, the lowest incomes have a lower tax burden compared to baselines.
The team admits that their simulations are very simplistic, and say they’ll continue to increase their complexity.
AI and Personhood
From brain-machine interfaces to digital assistants and automation, AI's growing presence in everyday life may alter what it means to be human.