Predistribution over Redistribution: Beyond the Windfall Clause
By Saffron Huang and Sam Manning
“Rawls's predistributive state achieves fairness in economic competition and equality of opportunity by [...] invest[ing] equitably in the human capital of its citizens. [...] It responds to the ideal of non-domination, and helps to realise a republican brand of freedom, by ensuring that all citizens can be active and independent participants in economic life [...]. And it is designed to realise the value of social equality by [...] giv[ing] everyone [...] the means to secure their status and self-respect as participating members of the economic life of that society.”
- Martin O’Neill and Thad Williamson, The Promise of Predistribution
If AI could bring enormous benefits to humanity, but also threatens to put us all out of work and concentrate economic production in the hands of a few, what should our policy approach be? Perhaps more importantly, what world does that policy approach presume, and what alternative world does it seek to bring about? While economic redistribution might be essential to ensure widespread prosperity in a world with advanced AI systems, we argue that focusing on predistribution — proactively ensuring widespread opportunity to benefit from AI — can reduce the likelihood that AI exacerbates inequality in the first place.
To make this argument, we start by examining the Windfall Clause — a prominent policy proposal for distributing AI company profits in a world where extremely powerful AI systems are developed and economic power is heavily concentrated in developers’ hands. We’ll then discuss why this flavor of after-the-fact redistributive economic policy may be insufficient to foster economic and political inclusion in a world where transformative AI gets developed. Next, we’ll walk through the rationale for why efforts to predistribute the benefits of AI could help promote widely shared prosperity, drive broader participation in shaping technological development, and bolster democracy en route to more advanced AI systems. We’ll end with a list of starter ideas for how predistribution could work in practice and put out a call for others to expand on them.
What is the Windfall Clause?
The Windfall Clause is a suggested mechanism for redistributing the economic benefits of AI, wherein AI firms commit to donating a significant amount of any enormous profits they earn. The Clause’s authors suggest that this obligation would kick in if the AI company came into a “windfall”, defined as “a level of income greater than a substantial fraction (e.g. , at least 1%) of the world’s total economic output.” The scenario’s driving assumption is that this firm would have invented and be in ongoing control of transformative AI, the deployment of which could deliver an enormous concentration of profit and power into its hands. The authors leave the issue of where and how to allocate donated profits as an open question, suggesting that the objective should be to “maximiz(e) expected good produced per dollar received.”
While the Windfall Clause is one of the first mechanisms specifically and seriously proposed for addressing the potential for extreme levels of global inequality driven by AI automation, the authors acknowledge that it is only one of many possible solutions. We share the hope that their report sparks more discussion and ideas, and aim to present an alternative perspective on distributing AI’s economic benefits. Our central critique of the Windfall Clause is that it only kicks in once AI has caused quite extreme wealth concentration.
What if, when it comes to AI, we could foster inclusive growth and avoid unprecedented inequality in the first place?
The world in which a Windfall Clause kicks in is one in which a company has accrued extreme global monopoly power. In the Windfall Clause report, the authors outline a potential tax schedule which triggers obligatory donations of 1% of their marginal profits once they reach 0.1% of world GDP and half of marginal profits once they reach 10-100% of world GDP. This sort of concentration of value might never come to fruition, but it's worth envisioning such a scenario for a moment. According to the example schedule, a company whose profits were equivalent to a staggering 10% of global GDP would donate ~1.81% of those profits under the Clause. In this scenario, a handful of employees, perhaps numbering in the tens of thousands, would control an unprecedented share of the world’s industrial, service, and other economic outputs. Their task would be to allocate less than 2% of the company’s profits among billions of people, many of whom may have been displaced from their jobs by AI automation.
While redistribution in the face of such extreme wealth concentration may be necessary to foster human flourishing, it is unlikely to be sufficient. Envisioning the reality of this highly unequal world compels us to consider whether there are actions we can take sooner to avoid such inequalities unfolding in the first place.
A chief issue is that the Windfall Clause—along with other ex-post (after-the-fact) financial benefits sharing regimes such as many post-AGI universal basic income proposals—does not address the forces that would continue to generate inequality. In Donella Meadows’s terms, this is fiddling with downstream parameters and constants (last on the list of effective leverage points for changing a system.) This is why, below, we suggest a comparatively more predistributive rather than redistributive approach. The idea of predistribution relative to some economic outcome specifies that these outcomes should be prevented from occurring in the first place, rather than rectified after the fact in a redistributive manner (i.e. as a Windfall Clause would).
A Windfall Clause would only be triggered after power has already become highly concentrated in the hands of a few, as economic power often confers political power. History has many examples of wealthy and powerful people controlling economic means, who used their influence in myriad ways to safeguard their economic interests. None of them had anywhere near as much power as some anticipators of transformative AI (TAI) believe its creators might. If such transformative AI exists, whoever owns a means of production that accounts for greater than 1% of world GDP (let alone greater than 10%) will have forms of soft and hard power scarcely imaginable. This may effectively nullify or render marginal any impact of a Windfall Clause to address material inequality. Further, this concentration of economic wealth might make any legal commitment mechanism to the Clause susceptible to being overturned, circumvented, or simply ignored by those wielding such considerable power and influence.
To be clear, a Windfall Clause scenario would be radical. The scenario where a private company operates independently at 1% of world GDP is staggering. Compare this to historical monopolies: in the 70s, the Bell System’s revenue (not profit) peaked at under 2% of U.S. GDP, or 0.4% of world GDP at the time. Ma Bell was broken up. Coupled with the fact that U.S. GDP is decreasing over time as a percentage of world GDP, if the 1% world GDP company is American, it would have had to first establish greater domestic dominance than Bell. Although there are no accurate GDP statistics from then, the British East India Company was arguably the largest corporation ever at its peak, ruling over large parts of the most productive nation in the world, larger than several nations, with its own private armies. It was eventually nationalized by the British Crown. The Windfall Clause anticipates a company that would truly be the king of all monopolies. It likely would be split up, regulated, and possibly even nationalized before the Clause kicks in—fates that have befallen many monopolies in the past—because the world’s countries, with armies at their back and national interests to protect, would likely not allow such power to accrue to a private company.
In some sense, such historical examples of governments curbing corporate power suggest that we may not need entirely new policy mechanisms to address the monopolistic potential of AI. Still, we do need to think about what those policy mechanisms might be, especially if we are to anticipate and ameliorate the forces of extreme wealth concentration before the fact.
We believe that we need to prioritize creating and implementing economic policy proposals for transformative AI (TAI) that predistribute rather than redistribute.
The world requires more decentralized and broadly-distributed abilities to contribute to AI value generation along the way. We favor enabling equality of opportunity ex-ante over enforcing equality of outcomes ex-post.
The rationale for prioritizing predistribution and equality of opportunity is part economic, part sociopolitical. For one, if AI progress is to be truly transformational, leading to unprecedented material abundance, then it could be one of our best opportunities to correct – rather than exacerbate – existing economic inequalities in the world. A world where AI’s incredible economic potential is directed toward the benefit of those with the greatest need could level the playing field between the rich and the poor, enabling a broader swath of the world to participate in the economy alongside AI systems that make new production possibilities and human aspirations a reality.
Furthermore, fostering inclusive growth can encourage more robust democratic systems relative to a world with highly concentrated economic influence. Economic growth that triggers a Windfall Clause would likely be driven by significant labor automation, which could increase inequality, particularly between workers and owners of capital. This rising inequality could undermine democracy through several mechanisms. Highly unequal societies tend to see increased elite influence over politics, as the wealthy leverage economic power to shape agendas and policies in their favor. At the same time, if AI drives massive productivity gains through labor automation, elites may find that they no longer need to depend on a large workforce to generate profits. This could make them less inclined to politically compromise with the working class. If AI advancement continues to accelerate, and we don’t set up institutions to proactively nudge progress in a more inclusive direction, we could kick off a vicious cycle in which AI-fueled inequality steadily erodes both the economic and political power of average citizens.
On the other hand, inclusive growth could significantly bolster democracy. Broad-based prosperity reduces economic disparities, leading to a more balanced distribution of political power. This environment discourages elite dominance, as there is less concentrated wealth to manipulate the political landscape. Furthermore, when people feel economically secure, they may be more likely to vote and participate in civic activities. By ensuring that all citizens have a stake in the economy as AI becomes a larger part of it, we can create a virtuous cycle where economic and political inclusion mutually reinforce each other, leading to a more just and democratic society. At a global level, as calls for international governance of AI get louder, it will be critical to ensure that communities around the world have access to use and benefit from AI so that they can play a role in shaping its future.
How could predistribution work in practice?
We don’t have all the answers about how exactly a predistributive benefits-sharing option should operate in practice. Any such proposal should be the product of a participatory design process, with intended beneficiaries playing a leading role. Also, certain methods of predistribution come with economic tradeoffs. For example, tax incentives might advantage the deployment of accessible personal assistant interfaces relative to enterprise-level API access for large firms, which could temporarily stunt growth and innovation in favor of more broadly distributed AI-supported productivity gains. Our goal in drafting this post is to spark more thinking and creative ideas around policy mechanisms to encourage an inclusive path towards shared prosperity and greater abundance.
As a start, we’ve listed a few potential directions below. This is not meant to be a comprehensive or exhaustive set of options. Please reach out to us if you have ideas you’d like to collaborate on to develop further.
Illustrative predistributive proposals could look something like:
Policies that enable more people to have an economic share in AI companies. E.g. policies that ensure the “human infrastructure” and labor that goes into AI is well rewarded and their bargaining power is increased, encouraging public listings of AI companies.
Using push and pull incentive mechanisms to direct the development and deployment of AI systems towards mobility-enhancing use-cases, such as education and skills-training, addressing global health challenges, health systems strengthening in low-resource settings, therapeutic and diagnostic development, agricultural productivity and innovation, job search and placement tools, housing, transportation, and others.
Predistributing key inputs to AI development to foster broader participation in AI development. This could include international pooling and sharing of compute resources, fostering a vibrant open-source AI development ecosystem, and investing in global talent development and capacity-building programs.
Ensuring language representation and accessible user interfaces for frontier capabilities prior to the release of a new model or system.
Pairing the deployment of increasingly capable AI systems with complementary investments in public digital infrastructure that localities need to fully reap economic benefits from such systems. For example, if the most economically advantageous deployments are through an API, but over a billion people worldwide don’t have internet access or aren’t positioned to benefit from an application built on the API, then digital infrastructure investments can foster more inclusive access to benefits across diverse global contexts.
Supporting the development of applications aimed at expanding worker capabilities rather than outcompeting workers at tasks they already perform.
We are not saying redistributive policies are unimportant. They may be necessary. Our claim is that predistributive policies can complement redistribution and lead to a more pluralistic and resilient society in a world where TAI is developed. They may also ultimately be more effective for safeguarding people’s interests and rights, especially in extreme cases, like if AI accounts for most or all of the world’s economic output. It is difficult to imagine a system that awards the vast majority of people little share, little power, and little value will still be able to enable those people to meaningfully advocate for their interests. Alignment requires stake.
Saffron Huang is the Research Director and Co-Founder of Collective Intelligence Project.
Sam Manning is a Senior Research Fellow at the Centre for the Governance of AI (GovAI). His work focuses on measuring the labor market impacts of AI and designing policy options to help ensure that advanced AI can foster broadly shared economic prosperity. Sam can be reached at sam.manning@governance.ai.
Claire Dennis provided helpful feedback on this piece.