I used an LLM to help review this post and it likely contains some AI-generated re-formulations. The ideas are not fundametally new and inspired by Nassim Taleb and trading lore.
In effective giving, it makes sense to be close to risk-neutral and focus on high expected utility per dollar.
However, in practice, organisations and people are risk-averse for good reasons.
It can add value to take more risk as an individual donor to generate higher expected utility at the margin. But when?
One important angle that I increasingly think matters a lot more than commonly appreciated is the following: If the downside is small and bounded, but the potential upside is very high, variance is good (cf. venture investing, convex macro trades). If you add diversification, you just need a few winners to have a good chance of realising a great outcome.
If the distribution of possible outcomes is close to symmetrical, variance per se does not matter (only the mean matters). If there is non-linearity of utility on the downside, variance is bad (selling convexity).
This line of reasoning makes it much easier to donate to uncertain global health interventions that might not work, but are unlikely to cause harm. However, it makes it harder to donate to longtermist cause areas where the theory of change is less direct, such as AI safety research which might lead to more dangerous AI, or to political lobbying that could really backfire for the organisations involved.
The variance framework clarifies a lot. The harder part may sit inside the "bounded downside" category itself.
Global health interventions look convex partly because the downside appears measurable and contained. But "unlikely to cause harm" and "bounded downside" often get conflated in ways that deserve more scrutiny. An intervention with weak evidence, scaled substantially, can produce harm that only surfaces in aggregate — substitution effects crowding out stronger alternatives, institutional credibility consumed on interventions that underperform, donor attention patterns that persist beyond the evidence. None of those show up as direct harm in a single-program evaluation.
The distribution may look bounded from inside one intervention and fat-tailed from the portfolio level. That asymmetry seems worth naming before treating global health as the safe side of the variance line.
Yes, I agree, it is non-trivial to assess (approximately, probably) bounded downside and this needs to be done on a case-by-case basis, including in global health.
However, if you find that there are fat tails on a portfolio level but none on the individual intervention level, then the individual intervention was probably not assessed comprehensively? Interaction effects and second order effects matter, of course, but at least a portion of them should arguably be credited back to individual iterventions as you assess them.