Imagine a world where policymakers or politicians could not only propose economic policies that could lift millions out of poverty and boost economic growth, but also have millions of years' worth of simulations and data to back up their lofty claims.
While there may not be enough room for trial and error with different policies in the real world without severe economic consequences, AI can be used to simulate millions of economies, with numerous economic policies, to find the optimum economic choice.
That is exactly what the researchers in a paper titled, 'The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies' from Harvard University and Salesforce Research tried to showcase through their AI framework that designed economic policies.
Income inequality is one of our world's biggest problems, which has only worsened over time, and exacerbated by the ongoing COVID-19 pandemic. In Bangladesh, the COVID-19 pandemic forced countless households into poverty and many people have lost their jobs or source of earnings due to the mandated lockdowns which worsened the country's income inequality.
To redistribute wealth and reduce income inequality, the government regularly adjusts its tax and subsidy policies, for example, heavily taxing the rich and/or offering financial aid to the poor. However, there is a trade-off, higher taxation can serve as a disincentive to work and invest, which lowers the country's productive potential and economic growth rate.
How can future governments design policies that optimise both income equality and productivity? Well, as the researchers in the aforementioned paper noted, through simulated interactions and reinforced learning (machine learning), the 'AI economist' was able to devise a tax policy that was able to improve income inequality.
Moreover, the AI-designed policy was able to attain similar results of productivity and income equality from an economy populated by human participants. Notably, the proposed tax policy by the AI offered a bigger piece of the economic pie to the middle class and subsidised the lower-skilled 'AI workers'.
The figure on the left shows the trade-off between equality and productivity of different models, the policy proposed by the AI Economist optimes for both equality and productivity outclassing every other model (except the free-market model in terms of productivity which results in the greatest inequality. | The figure on the right shows the wealth distribution of the agents with different skill levels (colour coded), unlike other models, the AI economist's tax model offers a more equal share of the economic pie.
Source: (Zhang and Trott et. al, 2020)
How the AI economist works
The simulation is called the 'Gather and Build Game' that is populated by randomly generated AI agents possessing four levels of skill that can move around, collect resources like wood and stone and build houses, which signify trading activities.
Each agent is programmed to maximise its utility, observe its environment, learn about the current tax rate and income, pay taxes and work accordingly. The difference in skill levels of the AI agents (replacing workers and consumers) allows lower skilled agents to specialise in collecting, gathering and selling building materials to the high skilled agents who bought the materials and built houses.
More importantly, like the AI agents, each simulation consisted of a policymaker AI, which monitored the level of wealth and the market and determined the level of tax to impose on agents with different skills, while simultaneously subsidising lower-skilled agents to redistribute wealth and improve social welfare.
As a result, there were two reinforced learning loops in the model, where the inner loop (the AI agents) optimised their behaviour depending on their surroundings (tax, price and income) and the outer loop (policymaker AI) which adjusted taxes and subsidies, to redistribute wealth and grow the size of the economic pie.
After these simulations were run millions of times with the agents and policymakers actively adjusted to each other's behaviour, ideal tax and subsidy rates were obtained by the AI economist, which maximizes both productivity and income equality. Interestingly, the tax schedule yielded by the AI economist was a mixture of both progressive and regressive taxation, which gives it an unusual shape compared to the tax schedules of other models (there is no taxation in the free market system).
Source: (Zhang and Trott et. al 2020)
What are the future implications?
The application of AI in different fields has become more and more widespread over time and will open new doors along the way. Economics is no exception. The AI framework discussed in the paper is still relatively simple and makes lots of assumptions (no tax evasion, rational decision-making, closed economy, etc.) to make the simulations intuitive and coherent.
However, as the researchers themselves argue, in the future, with the availability of greater computing power, the scale of modelling and more realistic parameters, the AI-driven models may no longer be as simplistic and instead be used for designing tax/ economic policies that maximise social welfare. Similarly, other economic goals like sustainability (environmentally sustainable policies) can also be optimised in the future using AI-driven policies to help make the world a better place.
If you want to read the paper for yourself or want to learn more about this fascinating area of research, I have included some additional resources below:
Zheng, Stephan, et al. "The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies." Salesforce Research, Salesforce Research, 30 June 2020, blog.einstein.ai/the-ai-economist/.
The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies, Stephan Zheng, Alexander Trott, Sunil Srinivasa, Nikhil Naik, Melvin Gruesbeck, David C. Parkes, Richard Socher.