Could the Minister of Finance be replaced with an AI?

How an algorithm might be the future of economic policy

I wrote this piece after being intrigued by a new AI designed to create economic policy. It’s interesting to think about what kinds of things AI is and isn’t well-suited to - arguably, economics should fall within AI’s wheelhouse. Whether we want to give an algorithm that much power is another question.

The idea that the work of our elected representatives could be done better by a computer program would strike most sensible people as absurd. Managing the economy and all of its moving parts, for instance, is a herculean task for any government. Balancing the implementation of new policies with keeping a steady hand on the tiller is no easy feat, even with a legion of civil servants in Wellington standing by to shower decision-makers with memos and white papers.

On the other hand, just a few years ago, the idea that a computer program (an AI) could write poetry was absurd. Now, AI can write poetry good enough to fool reputable judges into thinking it must have been written by a human. You can add cancer diagnosis, theoretical mechanics and professional poker to the list of surprising areas where AI performance can surpass the humble homo sapiens. A new initiative from Salesforce, an American software company, aims to add ‘economic policy’ to that list.

Stephen Zhang, the head of the ‘AI Economist’ research program, says that typical economic models relied on by economists and politicians suffer from serious challenges. To enable a controllable mathematical analysis, the models rely on assumptions. They might only focus on one policy lever in isolation (say, just GST rather than all taxes), and struggle to cope with large-scale shocks like Covid-19. Most of all, economic models rely on the debunked notion of humans as rational actors.

Zhang and Salesforce think that they can solve these issues with AI. Through a ‘reinforcement learning framework’, the Salesforce AI develops economic policy without any recourse to economic models or existing theory. The AI does this by simulating the impact of different economic policies on an economy and observing the results. The AI is tasked with achieving different social outcomes and coming up with the optimal economic package that balances those objectives.

The ‘reinforcement learning’ bit refers to the way the AI runs millions of simulations in the blink of an eye, each instance learning from what ‘worked’ in the last simulation. The AI is not given any prior knowledge about economic theory or policy whatsoever — it figures out what kind of policies to implement on its own, after observing simulation after simulation. You can think of it as though the AI starts from a blank slate where it can focus solely on results, unencumbered by decades of economic theory that usually dominate policy discussions.

An early iteration of the AI Economist was built for a scenario called ‘Gather and Build’. This scenario reflected a simple 2D economy, with different ‘agents’ collecting resources, earning money, and building houses. The AI could tax the workers however they liked, as well as redistribute money through welfare. The AI was tasked with maximising the productivity of the agents while minimising inequality (the agents had different skillsets, changing behaviours and tended to earn different amounts of money).

Over countless simulations, the AI Economist trialled different combinations of taxation and welfare to work out the optimal balance. Just like in the real world, taxing agents too much dropped their productivity, but not taxing them meant that there wasn’t money to redistribute.

The results of the reinforcement learning phase were impressive. At its conclusion, Salesforces’ AI had calculated the exact system of tax and welfare that would result in the highest combined score of productivity and equality. Those results were replicated in experiments with human participants.

As a comparison to the AI Economist, the research scientists ran the scenario under more orthodox systems of economic policy. These included a free market system with no taxation or redistribution, the American tax system, and a popular tax formula proposed by economist Emmanuel Saez. The AI Economist, despite beginning the scenarios with no pre-existing knowledge about tax theory or economics, outperformed all others in terms of maximising productivity and equality.

While this is an encouraging start to AI economic policy, the simplified 2D scenario is clearly a pale imitation of the complicated mess that is the real economy. And of course, governments have more objectives to balance than just productivity and equality. However, the same reinforcement learning techniques that power the AI Economist could well lay the groundwork for a more complicated AI to manage all of the complications of the real world. Salesforce is optimistic that it can scale up their promising work.

Given the radical potential of an AI better than humans at designing economic policy, it is worth thinking ahead of time about what the consequences might be for politics and government. While it is understandable to worry that delegating economic decisions to AI might erode our democracy, this doesn’t necessarily have to be the case. Remember that the AI isn’t deciding what outcomes (productivity, equality etc) are desirable, but which economic levers to pull to best achieve those goals. Arguably, voters might be freed up to focus on what they care about (outcomes) instead of needing to understand the minutiae of economic policy.

For an example of this post-AI politics, parties could run campaigns on which priorities they would feed into the AI Minister of Finance. Labour might say it will task the AI with balancing productivity and redistribution. ACT may campaign on a pledge to give the AI the freedom to do whatever it takes to maximise economic growth. The Greens would prioritise environmental protection over GDP. And so on. Whichever parties are successful on election day get to provide those priorities to the AI, which would accordingly set economic policy for the next three years.

The Mother of All Algorithms

This is a drastically different political landscape to imagine, but at the very least we might get more transparent politics on economic issues. Currently, all political parties campaign on delivering an economy that is “strong”, “fair”, and other bland epithets. Parties don’t even necessarily translate their campaign positions into policy if they form a government. Having an AI Minister of Finance that converts the societal objectives it is set into economic policy would force parties to be a little clearer on what trade-offs they are willing to make in terms of their priorities.

Whether or not we actually end up automating the country’s second-most important political position, AI will play an increasingly important role in helping governments craft policy. With the power and scale of well-designed AI, this trend should be welcomed as a step towards better government. However, a greater reliance on AI does raise a number of unique issues.

First and foremost, the integrity of the AI frameworks and their data must be sacrosanct. Particularly if the country’s economic outcomes are at stake, the AI would have to be protected from hacking from the outside world and interference from insiders. Relatedly, the data that the AI processes must be free from the biases that plague far too many algorithms at present. AI engineers tend to be white and wealthy, and great care should be taken to ensure the AI has no unconscious tilt in favour of those same privileged groups. Existing economic policy does that quite enough as it is.

Finally, the AI would have to be ‘explainable’ — that is, the particular AI would have to be able to show the working behind its policy recommendations, so they can be cross-checked by experts. This is not an easy technical feat, as the reasoning of advanced AI is sometimes too complicated even for the AI’s creators to understand. However, handing over the economy to a mysterious and opaque AI is a dystopian nightmare.

Partly due to these technical hurdles, Grant Robertson is at no risk yet of needing to use his own unemployment scheme. But Salesforce’s AI Economist has trailblazed a path for AI policy design. An AI whose simulations can today be summarised with a 2D GIF might in a decade be able to simulate the entire New Zealand economy, after absorbing all of the considerable data held by the government. Long before then, the very lodestar of ‘objective’ and ‘optimal’ governing should raise some welcome conversations about what it is exactly we seek from our representatives and the policies they pass.