Impact Prizes as an alternative to Certificates of Impact
Epistemic state: quite uncertain
An EA donor puts up a $50k prize for distribution in 2022. In 2022, several projects that have started since 2019 apply. Their net EA impacts are estimated, and these estimates (vs. the total value estimate of all submissions) are eventually used to give them corresponding proportional amounts of the $50k.
Back in 2019, several projects sell “rights” to their prize, and these get sold around. It’s expected that $1M in estimated total value will apply, so the market value of the claim of every $10 of estimated impact is $0.50. One project sets up an estimation service where they publicly estimate the eventual evaluation of every project, to help make the market more efficient, with the goal of themselves getting part of the prize.
I really like the goal of Certificates of Impact, but personally find them suboptimal in practice. I think Impact Prizes present an interesting alternative. It’s also possible Certificates of Impact could be used with Impact Prizes to gain the advantages of both down the road.
The most basic definition of Impact Prizes is something like,
“Declarations and fulfillment of prizes aimed at public benefit.”
Such a definition would apply to many existing charity prizes. They’ve recently been used with success on LessWrong in the interactions of the AI Alignment Prize.
I think these get more interesting with some extra less-explored features.
If one group has an expectation of making $2,000 of prize money from a future Impact Prize, they should be able to sell that claim to a 3rd party. This should be really simple, and that 3rd party should be able to easily resell it.
We can call “parts” of this claim “tokens.”
Suppose there’s a single $10,000 prize, to be awarded in 2020, and a specific group has a 20% chance of winning that prize. Then that group has an expected value of $2,000 of prize money. That group creates 100 tokens representing 100% of the claim to that prize. They sell 50 tokens for $1,000.
Later, they do win the $10,000 prize. However, because they only own 50% of the tokens, they only get $5,000. The other $5,000 goes to the previous share purchaser.
If only one prize was given, then share purchasers would only be interested in projects that have chances of being the top submitted project. This seems suboptimal.
Imagine instead that once the prize evaluation session begins, every single project is numerically evaluated for impact. Then each one gets a reward in proportion to the impact that the project was rated as having.
Say the $10,000 prize attracted 20 project entries, each of which was evaluated to have saved 1 life (these were all efficient anti-malaria projects). Each individual project would be awarded
This means that even small projects would receive rewards, and thus, they could effectively issue and sell tokens.
If a random sample of projects was selected to be evaluated, this may not change the expected value for stock purchases that much. A smaller proportion would get awards, but they would get proportionally more.
If the distribution of project impacts was considered very long-tailed and the sample very small, then this would disincentivize investments in better projects. Perhaps one partial solution would be to do a first quick round of review, ensure that the highest potential projects make the cut to be in the prize round, and then randomly select lower potential projects for the rest of the round.
Instead of using probabilistic evaluations, it could make sense to use decent priors. Imagine that all projects start off with wide distributions based on empirical priors. Then evaluators would gradually narrow these down in multiple passes, spending evaluation time roughly in proportion to the impact on the final result.
Counterfactual Prize Adjustment
I assume the main goal for many Impact Prizes would be to encourage valuable activity. This may not actually correlate that well with total project impact. There could be many submitted projects that would have been done equally well if it wasn’t for the Impact Prizes.
If this was a concern, it may be reasonable to estimate counterfactual prize value on some scale along with project value. Projects that would have been helped more by marginal prize money could be granted proportionally larger prizes. Say that each project is rated on a linear scale of 0-10 in terms of “counterfactual effect of prize amount”, and this was multiplied by its project impact estimate.
Say project A is a large-scale United Nations effort that created $2 Million of value, and project B is a smaller project by an independent organization. It comes out that the United Nations effort would have done the project without any expectation of a reward, while for the independent organization, the reward was a decisive factor. In this case, it seems possibly useful to be able to favor the independent organization in the award outcome.
Tooling and Earmarking
Instead of presenting $10,000 for “all projects”, it may make more sense to divide this pool to encourage a few areas. For instance, it may be common practice to earmark 20% for support and evaluation. The idea of this would be to encourage some people to “do good” by doing things that would help the prize. Some things to help could include setting up a Prediction Tournament to establish common knowledge of prize expectations, or web tools to make purchasing and selling more accessible.
Because some prizes would go towards efforts to help the prize system, this could lead to a minor prize-value-promotion economy. As stated above, some people could set up prediction systems, and other people could make predictions of prize outcomes on them. Others may act as police, detecting and reporting on bad actors.
Users could investigate not only bad behavior for prizes but also good behavior. In many tournament systems groups become quite competitive and indirect services like education or collaboration can be undervalued. If there could be a lot of value in some of these areas, then it should be evidently valuable when people point that out. The presence of some motivated actors actively investigating and promoting overlooked activity would hopefully lead to more of that activity.
Dealing With Multiple Prizes
One disadvantage of Impact Prizes, compared to Certificates of Impact, is that they could get complicated when there are several different prizes by different donors. A naive implementation of Impact Prizes could demand a unique token minting per project per prize, which would make things very messy. Any given project may have dozens of tokens to worry about and trade, and many exchanges may be between clusters of tokens at a time.
A simpler setup would look something more like Certificates of Impact. Only one token is made per project, but that token can be used for all Impact Prizes. Perhaps there would be a few common standards of tokens if Impact Prizes with different parameters.
Say in 50% of tokens of a project are sold for $2,000, and later that project wins $5,000 from an Impact Prize. With the case of shared tokens, this token-holder could expect to possibly win even more money later on from other Impact Prizes as well.
A related technique could just be that future donors often donate to existing Impact Prizes instead of creating new ones. This would mean that Impact Prizes would be lower-bound (the existing cash pool) but not simply upper-bound (it’s not clear how much more money will be added).
It’s possible that Certificates of Impact could work effectively as one of these token standards.
Risks and Insurance
One bias that these systems may create is that actors may be motivated to maximize upside risk, but may not care about minimizing downside risks. As long as Impact Prizes can only give out money (rather than demand money), than the lowest one should expect from a highly risky outcome is zero.
One way to get around this would be with formal insurance systems. All projects that create tokens could be required to purchase insurance upon project formation. When it comes to evaluation time, the Impact Prize could request that the insurer payout for any projects that are evaluated to be net-negative. It’s not obvious how they should strike a balance between charging for the entire cost or for the proportional cost.
In the case of multiple prizes, perhaps damages should be handled outside the prize system.
I think that tokenized Impact Prize systems in particular may be quite legally complicated. Corporate stock systems come with lots of rules, in part because there’s been an established record of people manipulating them in shady ways for personal gain.
If sophisticated financial instruments like shorting became possible, challenges could arise that would normally be addressed by corporate law. For instance, insider trading is regulated, in part, to prevent corporate employees from taking relatively simple actions to short their own stocks and then purposely cause bad things to happen.
If an Impact Prize system was established, it would have to either work within the current legal infrastructure, like stock, or outside of it. Both come with disadvantages.
It’s possible that only accredited investors would be able to purchase Impact Prize tokens, though this may be a fine first step.
These considerations would really need to be evaluated by an actual attorney. I suggest anyone considering doing this at scale hire an attorney first.
That said, similar problems would come up with Certificates of Impact if they was done to a similar scale. They also may be adequately addressed by existing cryptocurrency Token projects.
The final prize evaluations could be quite costly to produce. A few methods above could help, but significant costs would remain. I feel like there’s probably clever ways of thinking about this to incentivize everyone to maximize total value. For example, perhaps the evaluation cost comes out of each project’s value, incentivizing projects not to apply if that total would be below zero, and incentivizing them to make the evaluation easy.
My current model is that there a lot of incentives to not make most kinds of evaluations public. Perhaps the best comparison is prizes that are given out based on rubrics, though here most of the results of most of those rubrics are not made public.
The Impact Prize evaluations may be controversial and are likely to be at least somewhat misunderstood. Public evaluations may really require a community that is quite epistemically mature.
Controversy could create liability. If a Twitter war or similar gets started, it’s possible there could be enough anger for any prize to be canceled, or at least to stop future prizes.
For such a system to work well, a decent of work may be needed both on technical tooling and in implementation creativity.
If Impact Prizes took off, I could imagine some actors drawing into the ecosystem who only motivated by making profits. The token system may be looked at as a form of gambling (somewhat similar to the stock market) and may lead to some gambling tendencies. I think there may be some significant downsides here, but estimate that the upsides will be higher. (but this should be tested!) It could obviously be partly combatted using some of the techniques mentioned above.
Comparison to Certificates of Impact
A Philosophical Comparison
Perhaps the main philosophical difference between Impact Prize Tokens and Certificates of Impact is that Certificates of Impact, according to Paul Christiano, are supposed to represent causal responsibility. As he writes,
Allocating certificates requires explicit and transparent allocation of causal responsibility, both within teams and between teams and donors.
I personally find the causal responsibility bit unintuitive, and don’t expect a much larger community (especially outside the EA sphere) to accept it.
Impact Prize tokens would be decoupled from this idea.
A Ratio Comparison
I believe Certificates of Impact are supposed to be priced at their expected rates of impact, so $1 worth of certificates means $1 of counterfactual impact.
I think this will prove somewhat inflexible. I question the market viability of a $1 to $1 peg. If the demand is much less than what is necessary to create a $1 to $1 peg, then I would expect this to result in an illiquid market.
That said, of course a $1 to $1 would make things very simple if it works. A variable ratio could be fairly confusing and could require sophisticated purchasers (well, ones that could do two multiplications.)
Perhaps the main challenge to Impact Prizes as I discuss them is their additional complexity, compared to Certificates of Impact. It may require setting up a prize in advance, and then when that happens either doing a bunch of evaluations or figuring out clever ways of decreasing that burden.
The feature space is quite large. I’d like it to be larger. I’d be curious to hear other ideas for features or to modify the above features.
One area I’m particularly interested in is how best to structure the openness of evaluations. I think that the “Openness Cost” is very significant, and it would be nice to be able to reduce it while still maintaining much of the benefits of the evaluations.
 There’s much about the Blockchain world I don’t like, but they have used tokens extensively for this specific purpose. I don’t want to use the word “shares” because these parts will not have any voting rights, and legally there are other important distinctions.
Special thanks to Ryan Carey for discussing the concept, providing writing feedback, and suggesting I use the name “Impact Prizes” instead of something more obtuse.