Maximal cluelessness is a Global Priorities Institute Working Paper by Andreas Mogensen. This post is part of my sequence of GPI Working Paper summaries.

If you’d like a very brief summary, skip to “Brief summary.”

Introduction

Suppose you may choose to donate to the Against Malaria Foundation (AMF), which is estimated to save a child’s life for every ~$5,000 received, or donate to Make-A-Wish, which grants an ill child’s wish for ~$7,500 on average. If you care about maximizing the impartial good brought by your donation, the decision might seem obvious. But Mogensen argues it isn’t, and, more broadly, many EA priority rankings seem to be in tension with reasonable assumptions you might make about rational decision-making under uncertainty.

Cluelessness

If we choose actions based on their consequences, we have to consider all possible consequences—some of which stretch into the far future. As long as we don’t discount future value, future consequences dominate the differences in expected value between choices, meaning they determine our choices. The problem: we are clueless about our actions’ long-term consequences.

Mogensen anticipates a Naive Response to this: Cluelessness doesn’t impede our rational decision-making because we can still maximize expected value—we simply account for uncertainty by assigning probabilities.

He calls this “the Naive Response” because it doesn’t take the depth of our uncertainty seriously—our evidence often doesn’t even allow us to assign precise probabilities.

For example, consider the AMF. Our evidence suggests saving children’s lives may indirectly affect population size, but we are uncertain about how. Worse, we are uncertain about whether, at the margin, we want the population to increase or decrease. We also don’t know the political effects of saving lives through a charity instead of local health institutions. Hence, we are so deeply uncertain about the effects of donating to the AMF that we can’t create a precise probability distribution.^[1]

The maximality rule

One way to adapt to such uncertainty is to assess decisions with a set of plausible probability functions instead of just one. Unfortunately, doing so is incompatible with simple expected value maximization, leaving us to find a new decision criterion.

Mogensen proposes a decision criterion, the maximality rule, that he finds plausible enough to prevent us from ruling it out. 

The maximality rule (definition)

The maximality rule: We’re allowed to choose an act if no other act has greater expected value according to every probability function in our set of plausible probability functions.

When we have two acts, and one has greater expected value in some of our set’s probability functions but not all of them, the maximality rule doesn’t prefer that act over the other, but it also doesn’t think they are exactly equal. This is a key attraction of the maximality rule: It “does not contrive a preference between incommensurable options where there is none.”^[2]

Unfortunately, it also has its faults. First, because it considers every probability function, it can prefer acts that don’t have the highest expected value in any of our set’s individual probability functions.

Mogensen discusses a similar alternative to the maximality rule that doesn’t have this problem, called the liberal rule, but its problems lead him to prefer the maximality rule.^[3]

Mogensen also notes a fault in the maximality rule in the following case:^[4]

• p can be true or false. Your set of probability functions gives all probabilities from 10% to 80% to p being true.
• You're first offered bet A, which you can choose to take or leave. It pays $15 if p is false, but costs $10 if p is true.
• Then you’re offered bet B, which you can also choose to take or leave. It pays $15 if p is true, but costs $10 if p is false.
• Accepting both bets guarantees a $5 gain, regardless of whether p is true or false.

The maximality rule permits us to decline both bets, even though we are foregoing $5 for sure.

Another similar alternative to the maximality rule, the Γ-maximin rule, doesn’t have this problem, but, as before, its drawbacks lead Mogensen to prefer the maximality rule.^[5]

So, although some decision criteria are better than the maximality rule in some cases, Mogensen doesn’t think any of them are better overall; we can’t rule out the maximality rule and ought to avoid conclusions that aren’t consistent with it.

Implications

Donating to the AMF has numerous potential indirect effects, such as the population size. Some of these affect the long-term future, including slightly changing the probability of existential risks.^[1] Because minute changes to the far future have great significance, they dominate the differences in expected value between our options. The problem: we’re so uncertain about these long-term consequences that we can’t create a precise probability function to define expected values, giving us a reason to consult the maximality rule. Under such uncertainty, Mogensen argues that, in a decision between only the AMF and Make-A-Wish, the maximality rule permits us to donate to either.^[6]

While this conclusion doesn’t necessarily hold in reality, as we can donate to many organizations (not just these two), Mogensen thinks it applies to many highly-ranked EA global well-being charities, as we don’t know the sign or magnitude of their long-term effects (including the probability of extinction).

As he puts it:

If our evidence cannot rule out that the chance of extinction is ever so slightly higher given a choice to donate to one of GiveWell’s top charities as opposed to an organization like Make-A-Wish Foundation, [people abiding by the maximality rule] arguably need not prefer the former.

Unfortunately, interventions aiming to improve the long-term future or reduce existential risks also lack sufficient evidence of their effects, forcing us to rely on intuitive conjectures with a demonstrably poor track record.^[7]

Mogensen feels we need more research before we’ll know what the maximality rule would suggest about long-term focused interventions, but he thinks we should be skeptical that it would say we must choose them.

Conclusion

He concludes:

I do not insist that the maximality rule is correct. I merely claim that it is sufficiently plausible that we cannot rule it out. For all we know, orthodox effective altruist conclusions about cause prioritization are all true. In fact, I am inclined to believe they are. The problem is that I do not know how to set out and argue for a decision theory that is consistent with a long-termist perspective and supports these conclusions without downplaying the depth of our uncertainty. Then again, as a philosopher, I know that I am inclined to believe a great many things for which I lack an adequate response to certain apparently compelling sceptical challenges. Some may share my conviction that this is just one of those cases. But those who are already sceptical of effective altruist conclusions undoubtedly will not.

Brief summary

Mogensen argues:

1. When it comes to our actions’ broad effects, we are clueless—we’re so uncertain we can’t assign precise probabilities.
2. When we’re clueless, we can’t rule out using the maximality rule and shouldn’t draw conclusions contrary to it.
3. Using the maximality rule, we don’t have to donate to the AMF over Make-A-Wish, which probably applies to most other global well-being interventions.
4. We need more research to know what the maximality rule would suggest about long-term-focused interventions, but we should be skeptical that it would say we must choose them.

1. ^{^}

   See pages 13 to 16 for a detailed and concrete discussion (especially if you are skeptical of this notion).

2. ^{^}

   Quote from Bradley and Steele (2015).

3. ^{^}

   See pages 10 and 11 for the definition and drawbacks of the liberal rule.

4. ^{^}

   Elga (2010) identified this problem.

5. ^{^}

   The Γ-maximin rule requires us to act as if the worst possible expected utility of each act is correct (see page 12 for the proper definition of the Γ-maximin rule). Mogensen thinks this is extremely restrictive.
   The Γ-maximin rule also violates a plausible principle called Restricted Conglomerability (see pages 12 and 13 for discussion).

6. ^{^}

   Mogensen states he cannot prove this conclusion, though, because of numerous barriers. See the bottom of page 16 and all of page 17 for his argument.

7. ^{^}

   See Hurford (2013).