Summary
Introduction and Motivation
I recently completed a master’s thesis in economics on the topic of value drift and want to share a compact overview of my approach and findings.
In effective altruism, value drift mainly refers to the tendency of individual EA’s and EA-aligned organisations to change their values over time. Value drift can take on good or bad forms. It might mean growth in wisdom and development of moral character, which seems like a robustly good thing. It might also mean an erosion and abandonment of values which we hold dear in the present. This latter type of value drift seems problematic, as it implies that our future selves and institutions might not act in our present altruistic interest. For this reason, value drift is usually incorporated into the philanthropic discount rate, pushing us towards acting earlier. Yet acting early might forego benefits from knowledge accumulation. Waiting might enable a philanthropist to learn more about the state of the world, thereby enabling him to make better decisions in the future. Given these two considerations, when should a philanthropist act in order to achieve maximum impact?
In my thesis I focus on the tradeoff between delegating a decision to some better-informed future agent, and deciding in the present in order to avoid exposure to future value drift. For example, picture a government which knows that it has to implement drastic measures in order to combat climate change. For this purpose, it wants to make a large, irreversible investment in a promising technology. It is not clear, however, whether this investment really is the best possible, and additional research might unearth a more promising path. So the government might want to wait for future discoveries in order to make a more informed decision. Unfortunately, there is a powerful, climate-skeptic political movement breathing down the government’s neck, which is likely to take power during the next election. If the movement succeeds, then they will de-prioritise climate protection. Should the government make the investment or wait?
It is possible to reframe this problem as a tradeoff between option value and option disvalue. Option value emerges through knowledge accumulation, whereas option disvalue arises as a consequence of value drift. Several EA writers have discussed option value on this forum. For example, Maresca provides a good introduction into different varieties of option value. Schubert has pointed out that “hard-to-reverse decisions destroy option value”, and Brauner and Grosse-Holz argue that the option value of avoiding extinction is likely positive. Moreover, MacAskill argues that option value is a strong argument against human extinction, even if the expected value of the continued existence of humanity is negative. This model aims to add to these discussions, and while its conclusions could point in different directions, it hopefully provides a useful formal approach.
The Model
There is a finite number of time periods. In each time period, there is an agent who can decide whether to make a decision or whether to delegate decision authority to the next generation. The decision will have to be made eventually and cannot be pushed into the future indefinitely. Moreover, the decision is irreversible, and cannot be amended. Agents are endowed with a quadratic utility function that contains three scalar terms: the agent’s belief about the state of the world, her preference, and a variable that denotes the agent’s decision.
Beliefs about the state of the world drift according to a Gaussian random walk. Each period, more information about the state of the world becomes known, and so future agents are in a better position to decide. However, at the same time preferences drift according to a deterministic (later stochastic) law of motion. This implies that present agents will incur some utility loss when delegating their decision, as future agents will decide in their own interest. This basic tradeoff drives the dynamics of the model.
Social welfare is defined through a utilitarian social welfare function. The function takes into account both deviations from the state of the world and deviations from individual preferences. One could say that it measures the aggregate amount of dissatisfaction with a decision across all generations, with dissatisfaction of values measured relative to the eventual decision maker. This is taking an outside view to values - all perspectives are equally valid.
Findings and implications
The socially optimal decision maximises both gains from learning and gains from avoiding value drift. If knowledge accumulates monotonically, then all else equal, a social planner will prefer to see a decision at a later point in time. This can be different depending on the shape of value drift. For example, if there are large variations in preferences in the near term, and relative convergence in the long term, then it might be smarter to lock in an earlier decision, in order to accommodate the massive value differences of earlier agents.
When no social planner is around to intervene, then decisions might be taken inefficiently early. Given sufficiently low-variance value drift, all agents will want to delegate at least one period further. However, as all agents want to do that, the actual decision would be delegated into the future multiple periods, which is not desirable for early agents, as the decision would be taken too far into the future from their perspective. As a result, some of these early agents preempt excessive future delegation by implementing a decision themselves, foregoing the collective gains from knowledge accumulation.
If a government could instate taxes and subsidies across time, then this could restore the social optimum. For example, early agents could be subsidised for being patient, and later agents could be taxed for reaping the benefits of a more value aligned decision. The problem is that in many settings in which this model applies, this would require intergenerational transfers, i.e. earlier generations would enjoy more consumption whereas later generations would have to pay off the debt.
The previous conclusions apply under a deterministic law of motion for value drift. How do they change when drift is stochastic? An interesting case is when preference drift follows a Gaussian White Noise process. This corresponds to a situation in which all agents have the same preferences on average, but in which each generation experiences an idiosyncratic shock in their preferences. Examples for this could be political upheaval or moral progress. It turns out that agents will delegate if and only if this shock falls below a certain threshold. This setting generates a probability distribution over possible outcomes, whose shape is determined by the autocorrelation parameter and variance of value drift, as well as and the variance of knowledge accumulation. In a simulation, this distribution resembles an exponential distribution whose probability mass is decreasing the further a decision is delegated into the future.
Caveats and Further Directions For Research
These results rely on relatively restrictive assumptions. To name a few examples:
Addressing these limitations could yield inspiration for further research. I would also be interested in research that looks at intragenerational dynamics - e.g. the distribution of power within each generation - and how this affects delegation decisions. Moreover, it could be worthwhile to look more closely at possibilities for cooperation between generations with similar values. On the technical side, it seems that many of the complications of my approach could be avoided by a continuous-time model, although I have not investigated which additional challenges this would introduce.
If you are planning to build on my work, or if you are pushing in a similar direction, I am happy to discuss ideas and give feedback. Likewise, I would appreciate any feedback you might have on my thesis.
I want to thank my supervisor Prof. Daniel Quigley, Dillon Bowen, Philip Trammell, Sebastian Krantz, Sören Mindermann, and a seminar audience at the Forethought Foundation for helpful comments and feedback on my thesis manuscript.
For helpful comments on this blog post, I want to thank Holly Elmore, Ronny Fernandez Ishi Crew, and Alejandrina Cristia.
I think this is a valuable contribution—thanks for writing it! Among other things, it demonstrates that conclusions about when to give are highly sensitive to how we model value drift.
In my own work on the timing of giving, I’ve been thinking about value drift as a simple increase to the discount rate: each year philanthropists (or their heirs) face some x% chance of running off with the money and spending it on worthless things. So if the discount rate would have been d% without any value drift risk, it just rises to (d+x)% given the value drift risk. If the learning that will take place over the next year (and other reasons to wait, e.g. a positive interest rate) outweigh this (d+x)% (plus the other reasons why resources will be less valuable next year), it’s better to wait. But here we see that, if values definitely change a little each year, it might be best to spend much more quickly than if (as I’ve been assuming) they probably don’t change at all but might change a lot, since in the former case, holding onto resources allows for a kind of slippery slope in which each year you change your judgments about whether or not to defer to the next year. So I’m really glad this was written and I look forward to thinking about it more.
One comment on the thesis itself: I think it’s a bit confusing at the beginning, where it says that decision-makers face a tradeoff between “what is objectively known about the world and what they personally believe is true.” The tradeoff they face is between acquiring information and maintaining fidelity to their current preferences, not to their current beliefs. The rest of the thesis is consistent with framing the problem as a information-vs.-preference-fidelity tradeoff, so I think this wording is just a holdover from a previous version of the thesis which framed things differently. But (Max) let me know if I’m missing something.