Income inequality is an overarching problem in economics and, for policymakers, taxation has long been the answer. The classic model is that governments collect money according to their earnings and redistribute it via welfare schemes and public projects. But while taxation can help equalise the economy, taxing too much can also undermine the economy by discouraging people from working and encouraging tax avoidance.

Could AI have the answer? 

Salesforce, the cloud-based American relationship management company, has devised a way that AI can help. Richard Socher and his science team are building an AI-driven system called the AI Economist that uses reinforcement learning to identify optimal tax ratios for a simulated economy.

The tool is an exciting step forward as it enables us to evaluate policies and economies in a new way. “It would be amazing to make tax policy less political and more data driven,” says Alex Trott,  member of the Salesforce science team.

In an early result, the tool found a policy that could maximise both productivity and income equality – 16% fairer than an advanced progressive tax framework developed by expert economists.

The simulation involves four AI workers, each controlled by bespoke reinforcement-learning models, who are interacting with a two-dimensional world. They gather wood and stones, trade their resources with others, and allocate resources to buying houses that earns them money. The workers differ in skill, which leads them to specialise. AI workers with fewer skills learn better if they gather resources whereas those with more skills learn better if they buy resources to build houses.

An AI-controlled policymaker at the same time is working towards its goal of boosting both productivity and income of all workers. As each simulated year comes to an end, all workers are taxed fitting with a rate devised by an AI-controlled policymaker running its own reinforcement-learning algorithm.

The AIs meet with an optimal model formulated by millions of repetitions of the simulation. The workers and policymaker act through trial-and-error, with their simple instructions leading to very complex behaviours.

Soon, the project team has said, the number of AI workers will need to increase for the tool to model realistic scenarios. LeBaron, economic researcher at Brandeis University in Massachusetts, said in reviewing the tool, “There are people who argue you can get deep intellectual insights with just a few agents,” he says. “I’m not one of them.”

Playing the system

Even though the double use of AI workers and AI policy makers is key to the model, with its dynamic adaption, the AIs have found ways to game the system. In one example, AI workers began avoiding tax by reducing their productivity. This allowed them to qualify for a lower tax bracket, only to increase it again once qualified. This type of response from the AIs might show the high-level potential of the tool to gauge real-world responses, the Salesforce team says.

The optimal tax policy so says AI

It’s a bit different from our current. Existing policies usually fall into either progressive – where higher earners are taxed more – or regressive – where higher earners are taxed less. The AI Economist instead has combined the two: it applies the highest tax rates to the rich and the poor, and the lowest to middle-income workers. While seeming counterintuitive, its economic impact was a smaller gap between rich and poor.

LeBaron believes the tool is ready to start sanity-checking existing economic models, though. “If I were a policymaker, I would fire this thing up to see what it says.” If the tool disagreed with other models, then it could alert us if those other models were missing something, he says.

‘The world is too complicated’

In its current state, the tool is arguably too simple to factor in all of the complexities of real-world, human behaviour. An economist at the University of Oxford, Doyne Farmer, is not entirely convinced. While welcoming the use of reinforcement learning in economics, he believes it will be some time before the tool is actually useful. “The real world is way too complicated,” he says.

But he points out that, “If you are using an AI to recommend that certain people get lower or higher taxes, you’d better be able to say why.”

The team acknowledges that classic economists will need some persuading. To help lessen such concerns, Salesforce are also making their code open-source and allowing others to run their own scenarios through it. It’s this kind of openness, Socher believes, that, in the long run, will be crucial to keeping AI tools trustworthy.

Discover more:

The Digital Council’s advice to the New Zealand Government about building an empowered digital economy, post pandemic.

Helping government make the right decisions: We spoke to the Government Director behind Dot Loves Data’s COVID-19 Response Dashboard

Rising to meet the AI opportunity