While longtermism is an interesting ethical principle, I believe the consequence of the extent of uncertainty involved with the impact of current decisions on future outcomes has not been fully explored. Specifically, while the expected value may seem reasonable, the magnitude of uncertainty is likely to dwarf it. I wrote a post on it and as far as I can tell, I have not seen a good argument addressing these issues.
To be clear, I understand the argument of risk-reward tradeoff and how one is often irrationally risk-averse but I am not talking about that here.
One way to think of this is the following: if the impact of an intervention at present to influence long term future is characterized as a random variable X(t) , then, while the expectation value could be positive:E[X(t)]>0
the standard deviation as a measure of uncertainty , σ(t)=√E[X(t)2]−E[X(t)]2
could be so large that the coefficient of variation is very small: E[X(t)]σ(t)<<1.
Further if the probability of a large downside, P(X(t)<−a) is not negligible, where a>>E[X(t)] , then I don’t think that the intervention is very effective.
Perhaps I have missed something here or there have been some good arguments against this perspective that I am not aware. I’d happy to hear about these.
Magnitude of uncertainty with longtermism
While longtermism is an interesting ethical principle, I believe the consequence of the extent of uncertainty involved with the impact of current decisions on future outcomes has not been fully explored. Specifically, while the expected value may seem reasonable, the magnitude of uncertainty is likely to dwarf it. I wrote a post on it and as far as I can tell, I have not seen a good argument addressing these issues.
https://medium.com/@venky.physics/the-fundamental-problem-with-longtermism-33c9cfbbe7a5
To be clear, I understand the argument of risk-reward tradeoff and how one is often irrationally risk-averse but I am not talking about that here.
One way to think of this is the following: if the impact of an intervention at present to influence long term future is characterized as a random variable X(t) , then, while the expectation value could be positive:E[X(t)]>0
the standard deviation as a measure of uncertainty , σ(t)=√E[X(t)2]−E[X(t)]2
could be so large that the coefficient of variation is very small: E[X(t)]σ(t)<<1.
Further if the probability of a large downside, P(X(t)<−a) is not negligible, where a>>E[X(t)] , then I don’t think that the intervention is very effective.
Perhaps I have missed something here or there have been some good arguments against this perspective that I am not aware. I’d happy to hear about these.