Salesforce’s AI unit published
a fascinating piece on training a reinforcement learning model - the “AI Economist” (sorry) - to design optimal tax policies. They run a simulation of a mini-economy where the agents learn to maximise income and the AI Economist learns to optimise the tax rate. This is both an interesting technical problem (see
the full paper) and has potentially important real world implications.
One common trade off in tax policy is between output and equality. Higher tax rates (
sometimes) reduce work, but also reduce inequality. Salesforce claims that the AI Economist’s tax policy is better at managing this trade off than any real-world policies it tested against - and indeed is
strictly better than the US tax code, which it beats for
both output
and equality.
It’s interesting to look at how it achieves this, though, as the AI Economist’s optimal tax code is
odd, at least compared with what we’re used to. For example, compared to the US tax code, almost all high earners would pay a lower tax rate and most low earners would pay a higher one (before redistribution, importantly). Of course, Salesforce is at pains to emphasise that this is a toy model, but it’s nevertheless an interesting special case of the problem of
explainable AI: even if AI delivers better
results, are we willing to trust it if its
methods conflict with our political intuitions?