The debate about stratospheric aerosol injection (SAI) — releasing reflective particles into the upper atmosphere to reduce how much solar heat reaches the surface — tends to get stuck on the wrong question: whether it should be deployed. No one can answer that responsibly right now. SAI has never been tested in the real atmosphere at any meaningful scale, and what we have instead are climate models, data from imperfect volcanic analogues, and a long list of open questions about how a deliberate intervention would actually behave.
The more urgent question is how we get the evidence those open questions require — under what conditions, governed by whom, with what safeguards, and with what shared understanding of what the results would mean. The reason we need answers to such questions is not necessarily because deployment is inevitable, or even advisable, but because the alternative to a rigorous research framework isn’t no research. It’s uncoordinated, contested, unaccountable research that produces findings no one can agree are a sufficient basis for action under remaining conditions of uncertainty.
Medicine, however, has built a framework for a parallel set of problems. It took the better part of a century to build this framework, and it was built largely in response to unintended harmful outcomes that forced the issue.
We have solved this kind of problem before
In 1937, a pharmaceutical company in the United States began selling a liquid form of sulfanilamide — one of the first effective antibiotics, genuinely capable of treating infections that had previously been fatal. But to make it drinkable, the company dissolved it in diethylene glycol, a sweet-tasting solvent similar to antifreeze. No animal testing was conducted. Under the law at the time, none was required. More than a hundred people died, many of them children.
The public outrage that followed produced the 1938 Federal Food, Drug, and Cosmetic Act, legislation that for the first time required drug manufacturers to demonstrate safety before a product could be sold. The Food and Drug Administration, which had existed since 1930 but primarily as a reactive enforcement body responding to violations after they occurred, acquired genuine pre-market authority. Companies now had to submit evidence of safety and wait for approval before distribution.
That authority is part of why the United States largely avoided the thalidomide catastrophe a generation later. Thalidomide, prescribed to pregnant women across Europe as a treatment for morning sickness, caused thousands of birth defects — malformed or absent limbs, among other devastating effects — before it was withdrawn. In the United States, an FDA official named Frances Kelsey repeatedly refused to approve it; the 1938 law gave her the standing to do so. Without it, the American outcome would have likely looked like Europe’s.
By 1962, the thalidomide disaster — and the narrow American escape from it — made clear that safety testing alone was insufficient. Drugs also needed to be proven effective. The Kefauver-Harris Amendment that year added this requirement and, in doing so, codified the phased clinical trial structure that drug development relies on today: Phase I to assess safety and dosing in a small population, Phase II to evaluate efficacy and side effects, Phase III to confirm efficacy against a control in a larger population, and Phase IV post-licensure monitoring once a drug is in use.
Every part of that structure was built reactively. The 1938 Act followed a hundred deaths; the 1962 amendment followed thousands of birth defects. The logic they arrived at is right — build the evidence base before an intervention reaches the public, and decide in advance what would count as a reason to stop. Diethylene glycol in 1937 was a surprise.
Medicine developed the clinical trials system slowly, largely in response to crises that forced the issue. SAI doesn’t require the same sequence. Various events could force the issue here, such as a single government or private actor attempting a very large experiment before anyone has agreed how to evaluate what they did, or on what basis to object. Agreeing in advance what evidence would be sufficient, and what result would be a reason to stop, is possible before any experiment runs. And with the hindsight afforded by our experiences with a well-functioning, trusted system in medicine, not doing it is a choice, not an accident.
Research and deployment are different activities
One source of persistent confusion in the SAI debate is the tendency to treat research and deployment as points on a single continuum, such that more of one inevitably leads to more of the other. They are not. They are categorically different activities, with different justifications, different oversight requirements, and different decision-making processes — and conflating them is part of why the debate often ends up circular and heated.
Research is the process of generating evidence. Deployment is a decision about how to act on it. These require different justifications. A research program is, in part, warranted when key uncertainties are large enough that the expected value of resolving them is high; a deployment decision requires a separate weighing of risks, benefits, and affected interests that no research program can perform on its own. The research is designed to inform that decision. It cannot predetermine one.
During the COVID-19 pandemic, the FDA issued Emergency Use Authorizations (EUA) for vaccines while the clinical trials were still running. An EUA is not approval. It’s a separate and temporary legal authority that permits use of an unapproved product during a declared public health emergency, granted on a judgment that the known and potential benefits outweigh the known and potential risks, and revocable when the emergency ends or the evidence changes. The authorizations didn’t replace the trials or suspend the protection they provide. They operated alongside them, under a parallel legal and ethical framework, with their own justification, namely, that under emergency conditions, the expected cost of waiting for full validation is outweighed by the risk of acting on incomplete evidence. The trials ran to completion, and the vaccines went through the ordinary approval process afterward.
These authorizations drew heavy criticism, and that criticism was possible because there was a published standard to measure the departure against. A framework doesn’t stop anyone from acting early — it makes the shortcut visible as a shortcut, and keeps it reversible. A deployment decision for SAI, even a cautious one made in response to a climate emergency, is the potentially relevant analogue of the authorization, not the trial. It would need its own justification and possibly its own governance pathway, separate from and not derivable from whatever research program preceded it.
To be clear, clinical trials are a useful directional analogy for how SAI research should proceed, not a perfect one. Clinical trials in medicine are built around consent. A trial participant is told what the risks are, agrees to take them, can withdraw, and stands to benefit personally if the drug works. None of that holds here. There is no consent form for the stratosphere, no control planet, and no individual who benefits directly from an early experiment. That difference is part of why early outdoor experiments are generally designed to stay small enough to produce no detectable climate effect.
It’s also why the harder version of the consent question — who has standing to accept a risk on behalf of people who didn’t choose it — sits with the deployment decision, where it has to be engaged with rather than assumed, and where what counts as an adequate answer is itself unsettled.
Similarly, the specific phases of medical clinical trials – safety and dosing, then efficacy and side effects, finally efficacy against a control – do not transfer. But the discipline does: a carefully sequenced manner that minimizes undue risk, predefined endpoints, stopping rules, independent oversight, and a shared understanding of what each phase of research can and cannot tell us. The phases for SAI research will need to be defined on their own terms, by people who understand both the science and what decision-makers need to know. They will also need to change as the research exposes where the design was wrong, just as the phases in medical trials were revised repeatedly as experimenters and regulators learned where the old structure failed.
The clinical trials framework is an answer to the slippery slope, not its enabler
Treating research on and deployment of SAI as points on a single continuum is part of what has been called the “slippery-slope” objection. Its more fleshed out form goes something like this: any research program creates momentum toward deployment. Institutions get built. Scientists become advocates. Data gets interpreted by people with a stake in a particular outcome. One experiment leads to another. The research program acquires a life of its own, and before we know it we’re at deployment.
While not an inevitability, the history of science provides examples of such dynamics which make these concerns reasonable. But the features of a genuine clinical trials framework — predefined endpoints, binding stopping rules, pre-registered protocols, independent oversight, criteria established before any experiment is designed — can constrain those potential dynamics. A research program that defines in advance what a disqualifying result looks like, and stops when it finds one, resists the institutional capture and motivated reasoning that turn cautious inquiry into unstoppable momentum. The slippery slope is not an argument against structured research. It is an argument for why the structure has to be real — specific, stable, and set before anyone knows what the results will be.
Take an early outdoor experiment: releases on the order of tens of tonnes of sulfur dioxide, designed to characterize how aerosols form and disperse. A release this small would not produce any detectable climate effect, nor tell us what deployment would do. But they can constrain the microphysics that climate models currently can’t resolve, which is a narrow and specific thing to learn. So the experiment succeeds or fails on narrow questions. Did the aerosols behave the way the models predicted? Did the plume spread and dilute the way they said it would? Those questions have to be well considered before the release, because if you decide afterward what would have counted as a match, almost any result can be called one.
Questions like these however only settle whether models match observational data. They don’t tell us what to do with the answers: whether the uncertainty this experiment was designed to reduce has actually narrowed, whether the projections of what deployment would do are better constrained as a result, and whether that’s enough to justify a larger experiment — or whether the result means repeating this one differently, or stopping. The judgment to proceed – or not – is the one most exposed to the momentum the slippery slope describes. Name the uncertainty and the reduction you’re aiming for in advance, and the decision to proceed becomes a scientifically-informed, by and large, objective decision. Leave it open and it’s decided by the people with the most invested in continuing.
The slippery slope is not the only serious objection to SAI research, and it may not even be the most common one policymakers will raise, such as the argument that research itself weakens the motivation for cutting emissions. But the slippery slope is the objection the clinical trials framework answers most directly, and answering it exposes what’s actually missing: something that can play the role of the FDA.
In medicine, a stopping rule has force because the FDA (or local equivalent agency) can refuse to let a trial continue; Frances Kelsey could say no to thalidomide because a statute gave her the authority to do so. There is no equivalent for the stratosphere — there is currently no body that can review a protocol, withhold permission, or hold anyone to a threshold they wrote down in advance.
What exists instead is self-governance, and the research community has taken it seriously. Researchers in this field have drafted ethical frameworks, established independent oversight committees, and in at least one case, halted an experiment in response to community objection. That should continue. But voluntary standards only bind the people who volunteer. They do nothing about an actor who declines to participate, and they leave the field’s credibility resting on the good behavior of people who were going to behave anyway. Self-governance is what researchers can offer. It isn’t enough, and they shouldn’t be the ones deciding who gets to say no.
What this asks of governments
The clinical trials framework only works if governments are in the room when the research agenda is designed — not brought in after results are in hand.
This is a narrower ask than international climate negotiations usually present. It doesn’t require every country in the world to agree on whether SAI deployment is a good idea, and it doesn’t require anyone to commit to a particular outcome. It requires that governments willing to engage in SAI research help to generate initial answers to a specific list of questions: what uncertainties to prioritize, what counts as sufficient evidence, what results would be disqualifying, what oversight looks like at each stage, and who has standing to make those calls. These are political and ethical questions as much as scientific ones. Researchers can’t answer them on their own, and no single government can answer them for everyone. Answers reached without the civil society organizations who will either vouch for the results or challenge them won’t survive the first contested finding. What gives the answers weight is who was at the table when the questions were framed — which is where confidence in the results comes from, especially among the publics and communities most affected.
Building a clinical trial-like governance model before the fact would be new. Yet, it’s also more available now than the medicine story implies, because the thing that made 1938 and 1962 necessary hasn’t happened: nobody has deployed at scale, and there is no product already in the public’s hands. The evidence base from outdoor research is thin enough that no one’s position is locked in by results they’ve already published, and the number of governments and research groups involved is still small enough to get in a room.
For a government that’s curious about SAI but wary of committing, this creates a choice that matters more than it may initially appear. An imperfect framework is more or less being built now, incrementally, through the decisions of whoever is willing to engage in research. Governments that wait to evaluate the evolution of research norms once results are in will find that the consequential choices — what questions to ask, what counts as sufficient evidence, what outcome would be disqualifying — have already been made without them.
Thank you to Ignacio Mastroleo, Jessica Seddon, and Daniele Visioni for discussions and feedback on this idea. All shortcomings are our own.