Methodology

Several parameters the model relies on are not well established, and reasonable changes to them can meaningfully shift the final allocation. We've consulted experts in our network to arrive at best guesses, but readers should treat these as informed estimates rather than settled facts. We don't think the effective altruism community currently has good answers to many of these questions, and we plan to involve more experts in future iterations. Several examples are worth flagging:

• Attributing extinction risk across causes. How to split total extinction risk among AI, biological, and nuclear risks is genuinely uncertain, and small changes in these shares can shift significant portions of the allocation. Similarly, how much these extinction risks can be reduced by specific interventions is an uncertain assumption in our modeling work.
• Far future and extinction-risk impact. It is very difficult to confidently assess the long-run impact of a project. We mitigate this by assigning probability distributions to all inputs, by considering different scenarios with different levels of conservatism (for example, capping the time horizon at roughly 100 years in some worldviews), and applying different risk-weighting functions (some are risk-neutral, others penalize outcomes that are very unlikely to occur). In the future, we may expand upon the methods we're using to treat future uncertainty.
• How much credence to give to each moral theory. How much weight to assign to different moral theories also affects the result, and there is no clear-cut answer to how to set these credences.
• Cost-effectiveness estimates at the fund level. We estimate each fund's cost-effectiveness by combining estimates for the interventions we expect it to support, using a mix of historical and forward-looking data. The underlying data varies in quality: not all funds have detailed cost-effectiveness estimates available, and those that do vary in how transparent or self-critical their assumptions are. We have tried to level the playing field through adjustments so that conservative, self-critical funds are not penalized for their conservatism, but this requires subjective judgment. The model also does not capture our subjective confidence in particular grantmakers, and funds may reasonably change direction over time, so what each fund actually supports going forward may differ from the mix we have assumed. We hope to keep improving this as capacity allows.

In addition to involving more experts, we are working on a more advanced sensitivity analysis to assess how changes in the distributions of our inputs affect our ultimate recommendations.

We've included a small number of funds, and several cause areas are not represented at all. Most notably, climate and democracy preservation are not currently covered, and neither are "meta" causes aimed at improving giving decisions or expanding the number of people working on priority causes. We hope to include democracy and potentially climate work later this year. Rethink Priorities is a 501(c)(3) organization and is legally prohibited from supporting or opposing political candidates or parties, so giving to political candidates is also outside the model. Beyond these specific exclusions, we did not include many strong funds due to capacity constraints. Absence from this list should not be read as a judgment that a fund is less cost-effective than those we've covered.

Interaction effects between grantmakers

In certain cases, what one funder does will impact what another funder does. Sometimes they may pull a new funder into the space or crowd out a different actor (this is sometimes known as "funging"). If what we care about is overall impact, then we should consider these effects. However, these effects are difficult to model due to a lack of data on what someone would do in a counterfactual scenario, and they can differ substantially within and between areas.

For our purposes, we have selected funds that are in touch with other actors in their respective spaces, which should limit the impact of this concern. However, actors not in those spaces, particularly outside effective giving circles, may be causally affected by their actions in ways that are difficult to account for.

This may be a smaller issue in smaller fields, where it's less likely that other funders will step in. And a bigger issue in fields with more money is that a number of actors can make any given grant and may be interested in doing so.

Interaction effects between grant recipients

In some areas, grant recipients largely operate independently of one another, but in others, particularly in the policy space, a single bad intervention or action could harm the prospects or outcomes of related interventions. In our model, this type of concern may be most salient for AI grantmakers (where multiple actors are trying to influence policy in the same countries and locations) and less relevant for GHD (where interventions operate under different theories of change in very different countries). For animal interventions, certain actions may be more tightly correlated, like corporate campaigns where success may depend on keeping the public on your side, and a single rogue actor could harm the chance of success for all.

A number of the methodological decisions behind the model do not have a clear best answer, and choosing a different one would change the recommendation. For example, one of the risk profiles we use is skeptical of benefits that are very unlikely to occur. We currently implement this by fully counting all outcomes except the top 2.5% and discounting those in the top 2.5%. The exact threshold is a judgment call, not a settled methodological standard, and a different threshold would have produced different results. Similarly, all interventions are currently modeled as having diminishing overall effectiveness as funding scales, with no distinction between an intervention whose probability of success falls and one that becomes more likely to backfire or have no effect. Different moral theories and risk attitudes would treat those situations differently. We are running sensitivity analyses to better understand how much these choices matter, and we hope to refine our treatment in future iterations.

Aggregating across moral views is a young and unsettled area, and the choice of method does influence results. We have extensively reviewed the available literature on this topic and commissioned a report from Harry Lloyd, a philosopher specializing in moral uncertainty. Our model implements several aggregation methods and combines them using the weighted factor model described in this document.

Our model uses simplified versions of non-utilitarian theories, and does not yet include virtue ethics or the capabilities approach. We hope to add both in future iterations. One simplification is worth flagging.

To produce a recommendation, the model has to combine impacts across many people or animals (for example, weighing small harms to many against a large harm to one) and assign each option a single number so options can be compared. This naturally fits the reasoning of consequentialist theories. But non-consequentialist theories, such as deontology, do not fit this structure. The non-consequentialist theories represented in our model are better thought of as proxies for the recommendations these theories would make, rather than as reflecting the structure of their reasoning. We plan to keep iterating on our treatment of these theories in future versions.

More broadly, in its current form, our methodology rests on the assumption that the goal of giving is doing as much good as possible, where 'good' is defined by the donor's own worldview weights across welfare, risk reduction, and moral views. It is aimed at readers who broadly accept this quantitative lens.