Optimism Foundation

A productive outcome would be the development of a legitimate, multistakeholder governance framework that could be adopted by leading AI institutions. A practical AI governance framework must include actionable recommendations as to how to create institutional accountability. We have no shortage of well-intentioned "voluntary commitments" and "constitutional principles" authored by leading AI companies. What we lack are mechanisms that make those commitments binding, verifiable, and enforceable. Such a framework must also define the full set of stakeholders impacted by AI institutions and create avenues for their participation in institutional governance. Right now, the people most affected by AI systems often have the least say in how decisions about those systems are made. A usable framework would define scalable input processes, rapid feedback loops, and create avenues for objection. This framework would not just give these stakeholders visibility, it would also give them a voice. I believe a taxonomy of AI governance questions would be a useful tool for people, policymakers, and corporate leaders to better reason about specific issues and evaluate different solutions sets. I often hear people conflate technical, societal, and institutional matters when speaking about AI governance and it prevents more robust dialogues. When policymakers respond to concerns about AI labor displacement with red-teaming requirements, they've applied a technical solution to a societal problem — and neither the societal nor technical governance challenges get properly addressed.

Transparency, accountability, and human oversight;

As Head of Governance at the Optimism Foundation, a private sector steward of an open source technology platform, I've spent the past four years running novel governance experiments. The goal of these experiments was to address some of the shortcomings of the traditional corporate governance models, especially as it pertains to governing large technology platforms. My work most closely corresponds to SDG 16.6 and 16.7, specifically "the building of effective, accountable and inclusive institutions," and "ensuring responsive, inclusive, participatory and representative decision-making at all levels." I believe the thematic area of transparency, accountability, and oversight most closely maps to this SDG and my professional experience. It is of critical importance in institutional AI governance, which is my main focus and which is often overlooked in AI governance dialogues.

It's important to clearly distinguish between technical, societal, and institutional challenges in AI governance. Technical solutions largely focus on governance of the models themselves including data governance, evaluation frameworks, red-teaming, compute thresholds, and catastrophic risk policies. Societal solutions focus on bias, labor impacts, upskilling, the potential redistribution of profits, and various AI taxes. The third, often overlooked, category is institutional governance, or the corporate governance of AI Labs. In the private sector, we tend to overestimate technical solutions and underestimate institutional ones. I want to ensure the Dialogue includes conversations about creating accountability for decision-makers, defining the full set of stakeholders that should have a vote in governance, and creating legitimate procedural avenues for the affected to weigh-in on important decisions. The UN is one of the few institutions positioned to define a global, multi-stakeholder governance framework - one that produces institutions capable of being effective, accountable, and inclusive as AI reshapes the world. But only if the agenda reserves sufficient time and attention to the very important institutional aspect of AI governance.

As a private sector governance practitioner, I see important governance gaps in the governance of AI institutions. AI companies are self-regulating themselves via voluntary commitments. These commitments are well-intentioned but insufficient. While some AI companies are transparent in their efforts, this does not translate to accountability if these commitments are non-binding and changeable at will. Decisions that have wide-reaching externalities are often made by a small subset of private sector employees. Broad portions of the affected population are not represented in decisions made on their behalf. There is no defined process by which impacted stakeholders may object if these decisions negatively impact them. Many important stakeholders simply have no voice. This matters more for AI than for previous technology platforms. With traditional tech platforms, we reference what Albert Hirschman calls the "right to exit," which is the ability for a user to stop using a product and/or use an alternative. This is presented as a powerful alternative to having a "voice" in how platforms are run. However, the actual ability to exit decreases as a platform reaches market dominance. It does not translate well to AI models, which have externalities that can impact non-users as much, or more, than users. This makes "voice," which is currently absent in institutional AI governance, all the more important. There's a structural reason this dynamic won't fix itself. Even if an AI company values their governance commitments, these companies are for-profit corporations and their decisions must, at the very least, consider profit. This creates incentives to walk back the commitments that create accountability, in the name of competition. The traditional corporate governance model, which favors financial stakeholders, is an inappropriate means of addressing the breadth of governance challenges presented by AI. This presents an opportunity to update the institutional governance model.

At the Optimism Foundation, we've run novel governance experiments aimed at addressing problems in platform governance that are similar to those presented by AI .I am a practitioner, and my practical experience with novel governance structures is what makes my perspective distinct. Here are three approaches the UN might consider: *Stakeholder Voting* defines the full set of stakeholders impacted by institutional decisions and gives them a voice in the decisions that impact them. This could expand upon the work done by the Optimism Foundation to include platform users as a key stakeholder group by also including non-users impacted by the wide-reaching externalities of AI. This gives all stakeholders a voice in the governance of powerful AI institutions, not just financial shareholders. *Inclusive Input Processes* allow for representative participation among stakeholder groups. Importantly, these processes should occur continuously and at global scale. Some of the scalable input processes Optimism experimented with include sortition, deliberative processes, prediction markets, and optimistic approval processes. The AI institutions themselves have built powerful tools to collect input such as Anthropic Interviewer. Other organizations have used pol.is. *Credible Commitments* prevent institutions from going back on their commitments or changing them at will. At the Optimism Foundation, we've used smart contracts, novel corporate structures (such as Public Benefit Corporations, DUNAs, or Anthropic's Long Term Benefit Trust), and public elections to ensure leadership remains accountable.