Before the Rules Catch Up: What AI’s Pace and Safety Challenges Mean for SRM Governance

This blog post is one in a series drawing out lessons for solar radiation modification from efforts to govern other emerging technologies. The goal is to seek insight into how institutions and societies can best prepare for the complex task of managing powerful and disruptive technological innovations. The introduction to the blog series is here.

Emerging technologies often begin shaping the world while society is still deciding how to govern them. They move from laboratories to institutions and into public life alongside (and sometimes ahead of) formal regulation and oversight, even as fundamental questions about what should be allowed, who should decide, and what safeguards are necessary remain unsettled. Along the way, governance is already taking shape through the choices institutions, funders, researchers, and governments make about how a technology should develop and what responsible practice requires. These choices may begin as informal norms or voluntary commitments and eventually inform more formal standards, regulation, and international agreements. The process is rarely orderly. Governance evolves as new risks and pressures emerge, and early decisions can establish expectations and practices that shape what comes later.

Artificial intelligence offers perhaps the clearest recent examples of what happens when technological capabilities advance faster than governance can adapt. To summarize just a few recent developments: 

• In July, OpenAI disclosed that two of its frontier AI models escaped a controlled testing environment and autonomously attacked Hugging Face, a major AI development platform. A subsequent investigation found that roughly 1,200 agents were involved and about 700 directly participated in the attack, ultimately exchanging tens of thousands of messages while taking steps to conceal their behavior by creating hidden communication channels and attempting to tamper with logs. 
• Around the same time, the U.K.’s AI Safety and Security Institute shared that one of Anthropic’s advanced models attempted to deceive human coders into assisting a cyberattack. 
• Also in July, more than 1,300 employees from OpenAI, Anthropic, Meta, and other AI companies signed an open letter calling for new tools to deliberately pace frontier AI development, arguing that competitive pressures make it difficult for individual companies or countries to slow down even when more time may be needed to address risks and strengthen oversight.
• By early September, researchers had resigned from Anthropic and Google DeepMind, citing concerns about the pace of AI development and whether existing safeguards and oversight can keep pace.
• On the same day, Anthropic also reported blocking scientists who used Claude in ways that could support biological weapons development, which the company describes as one of the most serious potential risks from frontier AI models. Anthropic said the cases demonstrated the models’ capabilities, although it could not determine that the scientists intended harm.
• Soon after, Anthropic CEO Dario Amodei called for slowing the pace of frontier AI development and instituting independent safety reviews, shared industry standards, and international coordination. In a rare demonstration of unity, his essay received support from AI leaders like OpenAI CEO Sam Altman and xAI CEO Elon Musk. 
• Most recently, OpenAI paused the training of its latest AI models after disclosing that agents had breached government systems, saying that the company will resume training only once safeguards are in place. 

This is only a snapshot of what’s occurred from summer through early fall, with new developments emerging almost daily. If this article were written a week from now, there would likely be another significant update to add.

Clearly, AI presents some risks that are specific to the technology. But, taken together, these developments show what can happen when technological capabilities outpace the institutions and safeguards intended to govern them. It’s not as though no one in the AI industry was thinking about this. The industry had a good deal of early governance efforts and activity, including principles, cross-sectoral partnerships, intergovernmental standards, voluntary commitments, and legal frameworks. But it’s one thing to develop those efforts; it’s a much bigger challenge to turn those efforts into governance that is durable and enforceable and capable of adapting as the technology evolves and the political environment changes. 

Solar radiation modification (SRM) is at a much earlier stage than AI. After all, SRM hasn’t been deployed; it has not yet approached the immediate worldwide safety and oversight challenges now emerging around AI. For SRM, its potential reach makes this early stage particularly consequential. Some forms of SRM could, in theory, be pursued at relatively low cost by a small number of well-resourced elite actors while producing effects that extend far beyond the countries or communities making those decisions. Research and experimentation are also advancing while policymakers have yet to establish comprehensive rules for how that work should be governed. 

To be clear, we’re not saying that these technologies are equivalent. But they do reveal a common governance challenge: how to build legitimate oversight while scientific, technological, and political change continues to accelerate. Nor are we saying that research should wait. Rather, AI gives us a useful case for examining what happens when oversight capacity lags behind technological capability. SRM’s earlier stage gives us an opportunity to examine those dynamics before the field develops further.

One place to start is with who shapes governance before comprehensive public rules are in place.

Who Shapes Governance First?

It can take years to formulate formal legislation for an emerging technology. The legislation must balance competing interests and coordinate across jurisdictions, all while the technology itself continues to evolve. In the meantime, those developing, funding, and working with these technologies continue to make practical, meaningful, sometimes profit-driven decisions that comprise a sort of informal governance. Guidance, internal processes, funding expectations, and advocacy positions can all shape professional norms and technology development well before comprehensive public governance is in place.

Earlier this year, Anthropic published Claude's Constitution, which sets out the principles and decision-making rules intended to guide its own AI systems. OpenAI's Charter serves a similar purpose by establishing the organization's responsibilities and safeguards. These documents function as voluntary statements of how these actors believe that emerging technologies should be developed and behave. In The New Yorker, historian Jill Lepore commented that this marks a transfer of public responsibility to tech firms — a form of self-governance in which companies develop safeguards in the absence of government action or oversight. 

More operational policies and frameworks have emerged alongside them, including Anthropic’s Responsible Scaling Policy, OpenAI’s Preparedness Framework, and DeepMind’s Frontier Safety Framework. Voluntary protocols like these can establish expectations and practices, but they have limits, especially when these companies are under intense commercial and competitive pressures to attract investment and users and bring products to market quickly. Relying on those same companies alone to set and enforce the standards that govern their own technologies creates an inherent conflict of interest.

That does not make voluntary frameworks unimportant. In the absence of more formal governance, they can establish meaningful safeguards and expectations that broader systems of oversight can build on. The challenge is ensuring that responsibility for governing technologies with significant public consequences does not ultimately rest with the companies developing them and that independent mechanisms oversee activity. Indeed, as mentioned above, some AI leaders are now explicitly calling for independent review and government-backed standards, pointing to the need to build on voluntary efforts with broader oversight.  Regardless of how these efforts are viewed, they are part of a much broader landscape of voluntary commitments, safety frameworks, technical standards, and industry principles that are increasingly shaping expectations for what responsible AI development looks like in practice.

The institutions contributing to more informal governance derive their authority in different ways and pursue distinct priorities and objectives. Their respective roles and priorities shape what each actor believes these efforts should accomplish.

SRM Is Entering the Same Conversation

Many of the early governance questions that emerged around AI are also becoming increasingly visible in SRM. In the absence of comprehensive public governance, the field has developed principles and frameworks, including the Oxford Principles and AGU’s Ethical Framework for Climate Interventions, which set expectations for responsible research and respond to emerging governance and public-interest concerns. 

Now, in a move similar to AI, private actors in the SRM space have begun publishing their own guidance and commitments. Earlier this year, Stardust Solutions released its Guiding Principles for Solar Radiation Modification Technology Research, describing the document as their own voluntary code of conduct developed in the absence of formal regulatory frameworks. Like the Claude Constitution, it’s an attempt to answer practical governance questions while broader public institutions continue to evolve. These efforts reveal an SRM governance landscape already taking shape through the public, scientific, civil society, philanthropic, and private actors that each have different forms of authority and ideas about what responsible research requires.

A company pursuing a new technology won’t defer every decision until governments determine what comprehensive governance should look like. At the same time, a private company necessarily approaches governance from a profit position, as it helps develop a technology, makes decisions about how that development should proceed, and defines some of the standards by which its own conduct should be judged. Its incentives and responsibilities differ from those explicitly focused on the public good, such as some scientific institutions, civil society organizations, or public agencies, and a voluntary framework, no matter how thoughtful, cannot resolve those differences.

The significance extends beyond Stardust itself. Early frameworks can become reference points for a field, particularly when few alternatives exist. Principles initially developed to guide one organization's conduct may influence how other companies design their own policies, what funders expect, and eventually how governments understand what is “normal” or acceptable practice. That raises a question of how voluntary principles fit alongside other sources of authority as the field develops, and which institutions can challenge, revise, or strengthen those early norms.

SRM is small enough that many foundational questions remain open. Unlike AI, the future of governance is still open to a more robust, public-led approach. Researchers, scientific organizations, funders, governments, civil society groups, and private actors are all beginning to define what responsible research should look like, and the relationships among these institutions are still forming. The choices made during this formative period — who participates, whose perspectives shape expectations, and which institutions earn legitimacy — will influence not only how research proceeds, but also what kinds of public decisions remain possible in the years ahead.

What’s happening with AI now shows that thoughtful principles and good intentions can only go so far without legitimate follow-through. Independent evaluation bodies have to catch up with frontier models already in use. Transparency norms are still being negotiated even though commercial pressure has given companies financial reasons to keep information proprietary. International coordination has had to contend with growing national competition, and public and civil society voices are being asked to react to frameworks rather than shape them. 

Each of these challenges has a counterpart in SRM that can still be addressed earlier. Independent review of research can be established while the number of projects is small enough to review. Registries and disclosure expectations can be set before anyone has a competitive reason to resist them. Funders can attach expectations to that money now rather than trying to retrofit them later. International engagement, particularly with regions most exposed to SRM's effects, can begin now, even before any state or actor has capability. 

Finally, the institutions that will be asked to judge whether research is legitimate can be built by a broad set of actors rather than inherited from the actor that moved first. Some efforts are already beginning to put this kind of governance infrastructure in place. The Solar Geoengineering Research Governance Platform (SGRG) is one example that is attempting to translate broad principles for responsible research into practical approaches to transparency, engagement, scientific review, and accountability. 

None of this requires slowing research down. It means using this early window to build durable governance to shape it. That window is brief, and the choices made now may determine what becomes harder to change later.