1. General

Category

Category

Other

Please describe Other

Public policy: Economics, Public Health, etc

2. Project Details

Company or Institution

Salesforce

Project

The AI Economist

General description of the AI solution

The AI Economist designs and evaluates optimal economic policies using reinforcement learning and data-driven economic simulations.
The simulations contain AI agents (emulating real people) and model (parts of) real economies. In particular, reinforcement learning is a key AI technique that compares millions of economic policies and evaluates how the economy responds to (changes in) policy.
In this way, the AI Economist iteratively finds policies that optimize social welfare, taking the response of the economy into account.
This approach has discovered policies that outperform classic economic tax models, and real-world US COVID policies in simulation.

Website

https://www.einstein.ai/the-ai-economist

Organisation

Salesforce

3. Aspects

Excellence and Scientific Quality: Please detail the improvements made by the nominee or the nominees’ team or yourself if your applying for the award, and why they have been a success.

1. Type of AI:

Reinforcement learning and data-driven economic simulations. For example, we've designed optimal tax policies in 2d spatial worlds, and optimal responses to COVID-19 in data-driven simulations of the COVID-19 pandemic and its public health/economic impact in the US.

2. Quality of AI solution and algorithm: To what extent is the research work clear and detailed?

The AI Economist has found tax policies that achieve 16% higher equality and productivity than the classic Saez optimal tax model. Compared to US policies in 2020, executed in simulation, the AI Economist also finds COVID-19 response and subsidy policies that reduce deaths by 3x, while maintaining similar levels of unemployment over 2 years.

For details on the COVID-19 project:
https://einstein.ai/the-ai-economist/ai-policy-foundation-and-covid-case-study
https://arxiv.org/abs/2108.02904

Income taxes, equality, and productivity: https://arxiv.org/abs/2004.13332

3. Describe status of technology:

The AI Economist is under active development and is increasing the scale and realism of the economic simulations. This means simulating more AI agents, multiple levels of government, firms, consumers, etc. AI policies have outperformed classic economic policy models and real-world policies in simulation so far. A high-level goal is to work with economists and policy makers to see how this technology can best inform future policy.

Scaling of impact to SDGs: Please detail how many citizens/communities and/or researchers/businesses this has had or can have a positive impact on, including particular groups where applicable and to what extent.

Overall view:
Equality, productivity, public health, unemployment, and other metrics are key economic indicators and issues. The AI Economist has flexible, versatile, and can be used to analyze any economic policy problem, given the right economic simulation. Given future advances in bottoms-up modeling (starting with the individuals in the economy), the AI Economist has the potential to guide and improve policy globally, towards improving equality, productivity, public health, etc everywhere. In simulation, AI policies can improve social welfare by at least 16% (equality x productivity), or reduce deaths by 3x (COVID-19 simulation). In particular, the latter simulation fitted historical outcomes well, and used all publicly available data. As such, it holds significant promise in improving social welfare in the real world as well.

Measurable progress
The realism of economic simulations is rapidly increasing. The COVID-19 simulation used a large amount of publicly available data on the pandemic in the US. Building more realistic simulations requires more responsible and timely data collection. Further progress will predominantly depend on the quantity and quality of economic data, and the scale and realism of economic simulations. As such, the quality of the AI Economist system can be measured by potentially deploying + evaluating the impact of AI policies in the real world, and comparing outcomes with the simulation projections.

Clarity of SD components: See above.

Projection of impact and uptake of the AI solution to Sustainable Development:
The impact is global, as equality, productivity, public health, etc, are issues that affect every citizen in every country.

A key bottleneck is educating economists and computer scientists about each other's domain, and ensuring that this technology is developed in an interdisciplinary fashion. There is increasing interest and uptake from prominent academic labs in the US and elsewhere on this technology. For instance, the Github respo with economic simulation + RL code has 652 "stars" — a significant number for a up-and-coming research direction.

Global impact: See above.

Global added value: See above.

Scaling of AI solution: Please detail what proof of concept or implementations can you show now in terms of its efficacy and how the solution can be scaled to provide a global impact ad how realistic that scaling is.

Evidence for impact: See above.

Scalability: As mentioned above, a key bottleneck is educating economists and computer scientists about each other's domain, and ensuring that this technology is developed in an interdisciplinary fashion. There is increasing interest and uptake from prominent academic labs in the US and elsewhere on this technology. For instance, the Github respo with economic simulation + RL code has 652 "stars" — a significant number for a up-and-coming research direction.

Such education takes time and requires economists to acquire experience with the latest AI techniques, and implementing these. For computer scientists, this requires learning about economics, policy making, and translating statistical and AI concepts to real-world economic problems. This will take time and adaptation.

Network effort: Our simulation and RL code is publicly available. A community of ~300 people is on Slack where they can discuss this project. Over time, as more and more people get familiar with our work, they can build on our code and insights to apply the AI Economist to their own problems. So far, 34 other scientific papers have already cited our preprint paper. Our preprint remains in review at a top-tier academic journal. As such, momentum is growing.

Ethical aspect: Please detail the way the solution addresses any of the main ethical aspects, including trustworthiness, bias, gender issues, etc.

In both the income tax and COVID-19 projects, we have spent a significant amount of time doing ethical reviews with internal and external ethics teams and experts. We have consulted tax experts, economists, AI ethics, and other domain experts. These findings have been summarized in our public papers and demos. We have also created simulation cards for our open-source software, which detail the intended use and known risks of our software. As such, our projects have been subjected to significantly higher-than-average ethical analysis and assessment.

Our economic simulations have the potential to describe all segments and parts of society, specifically including underrepresented groups. Our research highlights the need for more representative datasets, and the potential for AI policies to target equality for all as an explicit objective for AI policy design. This is a specific and significant technological breakthrough: classic economic methods *cannot* optimize for equality directly, and therefore often cannot optimize policies directly for equality. As such, AI techniques have significantly more potential for real-world applicability and relevance than classic economic models that rely on simplified analyses and models of the economy and its complexities.