From 2013 to 2021, Open Philanthropy donated $200M to criminal justice reform. My best guess is that, from a utilitarian perspective, this was likely suboptimal. In particular, I am fairly sure that it was possible to realize sooner that the area was unpromising and act on that earlier on.
In this post, I first present the background for Open Philanthropy’s grants on criminal justice reform, and the abstract case for considering it a priority. I then estimate that criminal justice grants were distinctly worse than other grants in the global health and development portfolio, such as those to GiveDirectly or AMF.
I speculate about why Open Philanthropy donated to criminal justice in the first place, and why it continued donating. I end up uncertain about to what extent this was a sincere play based on considerations around the value of information and learning, and to what extent it was determined by other factors, such as the idiosyncratic preferences of Open Philanthropy’s funders, human fallibility and slowness, paying too much to avoid social awkwardness, “worldview diversification” being an imperfect framework imperfectly applied, or it being tricky to maintain a balance between conventional morality and expected utility maximization. In short, I started out being skeptical that a utilitarian, left alone, spontaneously starts exploring criminal justice reform in the US as a cause area, and to some degree I still think that upon further investigation, though I still have significant uncertainty.
I then outline my updates about Open Philanthropy. Personally, I updated downwards on Open Philanthropy’s decision speed, rationality and degree of openness, from an initially very high starting point. I also provide a shallow analysis of Open Philanthropy’s worldview diversification strategy and suggest that they move to a model where regular rebalancing roughly equalizes the marginal expected values for the grants in each cause area. Open Philanthropy is doing that for its global health and development portfolio anyways.
Lastly, I brainstorm some mechanisms which could have accelerated and improved Open Philanthropy’s decision-making and suggest red teams and monetary bets or prediction markets as potential avenues of investigation.
Throughout this piece, my focus is aimed at thinking clearly and expressing myself clearly. I understand that this might come across as impolite or unduly harsh. However, I think that providing uncertain and perhaps flawed criticism is still worth it, in expectation. I would like to note that I still respect Open Philanthropy and think that it’s one of the best philanthropic organizations around.
Open Philanthropy staff reviewed this post prior to publication.
Index
Background information
What is the case for Criminal Justice Reform?
What is the cost-effectiveness of criminal justice grants?
Why did Open Philanthropy donate to criminal justice in the first place?
Why did Philanthropy keep donating to criminal justice?
What conclusions can we reach from this?
Systems that could have optimized Open Philanthropy’s impact
In that blog post, they described their expansion into the criminal justice reform space as substantially a “bet on Chloe”. Overall, the post was very positive about Chloe (more on red teams below). But the post expressed some reservations because “Chloe has a generally different profile from the sorts of people GiveWell has hired in the past. In particular, she is probably less quantitatively inclined than most employees at GiveWell. This isn’t surprising or concerning—most GiveWell employees are Research Analysts, and we see the Program Officer role as calling for a different set of abilities. That said, it’s possible that different reasoning styles will lead to disagreement at times. We think of this as only a minor concern.” In hindsight, it seems plausible to me that this relative lack of quantitative inclination played a role in Open Philanthropy making comparatively suboptimal grants in the criminal justice space [1].
Initially, I thought that might be because of an expectation of winding down. However, other Open Philanthropy cause areas also show a similar pattern of going up in 2019, perhaps at the expense of spending on Global Health and Development for that year:
In 2021, Open Philanthropy spun out its Criminal Justice Reform department as a new organization: Just Impact. Open Philanthropy seeded Just Impact with $50M. Their parting blog post explains their thinking: that Global Health and Development interventions have significantly better cost-effectiveness.
What is the case for Criminal Justice Reform?
Note: This section briefly reviews my own understanding of this area. For a more canonical source, see Open Philanthropy’s strategy document on criminal justice reform.
There are around 2M people in US prisons and jails. Some are highly dangerous, but a glance at a map of prison population rates per 100k people suggests that the US incarcerates a significantly larger share of its population than most other countries:
Outlining a positive vision for reform is still an area of active work. Still, a first approximation might be as follows:
Criminals should be punished in proportion to an estimate of the harm they have caused, times a factor to account for a less than 100% chance of getting caught, to ensure that crimes are not worth it in expectation. This is in opposition to otherwise disproportionate jail sentences caused by pressures on politicians to appear tough on crime. In addition, criminals then work to provide restitution to the victim, if the victim so desires, per some restorative justice framework [2].
In a best-case scenario, criminal justice reform could achieve somewhere between a 25% reduction in incarceration in the short-term and a 75% reduction in the longer term, bringing the incarceration rate down to only twice that of Spain [4], while maintaining the crime rate constant. Say that $2B to $20B, or 10x to 100x the amount that Open Philanthropy has already spent, would have a 1 to 10% chance of succeeding at that goal [5].
What is the cost-effectiveness of criminal justice grants?
Estimation strategy
In this section, I come up with some estimates of the impact of criminal justice reform, and compare them with some estimates of the impact of GiveWell-style global health and development interventions.
Throughout, I am making the following modelling choices:
I am primarily looking at the impact of systemic change
I am looking at the first-order impacts
I am using subjective estimates
I am primarily looking at the impact of systemic change because many of the largest Open Philanthropy donations were aiming for systemic change, and their individual cost-effectiveness was extremely hard to estimate. For completeness, I do estimate the impacts of a standout intervention as well.
I am looking at the first-order impacts on prisoners and GiveWell recipients, rather than at the effects on their communities. My strong guess is that the story the second-order impacts would tell—e.g., harms to the community from death or reduced earnings in the case of malaria, harms from absence and reduced earnings in the case of imprisonment)—wouldn’t change the relative values of the two cause areas.
After presenting my estimates, I discuss their limitations.
Simple model for systemic change
Using those what I consider to be optimistic assumptions over first-order effects, I come up with the following Squiggle model:
initialPrisonPopulation = 1.5M to 2.5M # Data for 2022 prison population has not yet been published, though this estimate is perhaps too wide.
reductionInPrisonPopulation = 0.25 to 0.75
badnessOfPrisonInQALYs = 0.2 to 6 # 80% as good as being alive to 5 times worse than living is good
counterfactualAccelerationInYears = 5 to 50
probabilityOfSuccess = 0.01 to 0.1 # 1% to 10%.
counterfactualImpactOfGrant = 0.5 to 1 ## other funders, labor cost of activism
estimateQALYs = initialPrisonPopulation * reductionInPrisonPopulation * badnessOfPrisonInQALYs * counterfactualAccelerationInYears * probabilityOfSuccess * counterfactualImpactOfGrant
cost = 2B to 20B
costPerQALY = cost / estimateQALYs
costPerQALY
That model produces the following distribution:
Note: mean(cost)/mean(estimateQALYs) is equal to $8160/QALY
This model estimates that criminal justice reform buys one QALY [6](quality-adjusted life year) for $76k, on average. But the model is very uncertain, and its 90% confidence interval is $1.3k to ~$290k per QALY. It assigns a 50% chance to it costing less than ~$19k. For a calculation that instead looks at more marginal impact, see here.
EDIT 22/06/2022: Commenters pointed out that the mean of cost / estimateQALYs in the chart above isn’t the right quantity to look at in the chart above.mean(cost)/mean(estimateQALYs) is probably a better representation of “expected cost per QALY. That quantity is $8160/QALY for the above model. If one looks at 1/mean(estimateQALYs/cost), this is $5k per QALY. Overall I would instead recommend looking at the 90% confidence intervals, rather at the means. See this comment thread for discussion. I’ve added notes below each model.
Simple model for a standout criminal justice reform intervention
Some grants in criminal justice reform might beat systemic reform. I think this might be the case for closing Rikers, bail reform, and prosecutorial accountability:
Rikers is a large and particularly bad prison.
Bail reform seems like a well-defined objective that could affect many people at once.
Prosecutorial accountability could get a large multiplier over systemic reform by focusing on the prosecutors in districts that hold very large prison populations.
For instance, for the case of Rikers, I can estimate:
initialPrisonPopulation = 5000 to 10000
reductionInPrisonPopulation = 0.25 to 0.75
badnessOfPrisonInQALYs = 0.2 to 6 # 80% as good as being alive to 5 times worse than living is good
counterfactualAccelerationInYears = 5 to 20
probabilityOfSuccess = 0.07 to 0.5
counterfactualImpactOfGrant = 0.5 to 1 ## other funders, labor cost of activism
estimatedImpactInQALYs = initialPrisonPopulation * reductionInPrisonPopulation * badnessOfPrisonInQALYs * counterfactualAccelerationInYears * probabilityOfSuccess * counterfactualImpactOfGrant
cost = 5000000 to 15000000
costPerQALY = cost / estimatedImpactInQALYs
costPerQALY
Note: mean(cost)/mean(estimateQALYs) is $837/QALY
Simple model for GiveWell charities
Against Malaria Foundation
Using a similar estimation for the Against Malaria Foundation:
costPerLife = 3k to 10k
lifeDuration = 30 to 70
qalysPerYear = 0.2 to 1 ## feeling unsure about this.
valueOfSavedLife = lifeDuration * qalysPerYear
costEffectiveness = costPerLife/valueOfSavedLife
costEffectiveness
Note: mean(costPerLife)/mean(valueOfSavedLife) is $245/QALY
Its 90% confidence interval is $90 to ~$800 per QALY, and I likewise validated this with Simple Squiggle. Notice that this interval is disjoint with the estimate for criminal justice reform of $1.3k to $290k.
GiveDirectly
One might argue that AMF is too strict a comparison and that one should instead compare criminal justice reform to the marginal global health and development grant. Recently, my colleague Sam Nolan quantified the uncertainty in GiveDirectly’s estimate of impact. He arrived at a final estimate of ~$120 to ~$960 per doubling of consumption for one year.
The conversion between a doubling of consumption and a QALY is open to some uncertainty. For instance:
GiveWell estimates it about equal based on the different weights given to saving people of different ages—a factor of ~0.8 to 1.3, based on some eye-balling from this spreadsheet.
GiveWell recently updated their weighings to give a DALY (similar to a QALY) a value of around ~2 doublings of income.
Commenters pointed out that few people would trade half their life to double their income, and that for them a conversion factor around 0.2 might be more appropriate. But they are much wealthier than the average GiveDirectly recipient.
Using a final adjustment of 0.2 to 1.3 QALYs per doubling of consumption (which has a mean of 0.6 QALYs/doubling), I come up with the following model an estimate:
costPerDoublingOfConsumption = 118.4 to 963.15
qalysPerDoublingOfConsumption = 0.2 to 1.3
costEffectivenesss=costPerDoublingOfConsumption/qalysPerDoublingOfConsumption
costEffectivenesss
Note: mean(costPerDoublingOfConsumption)/mean(qalysPerDoublingOfConsumption) is $690/QALY
This has a 90% confidence interval between $160 and $2700 per QALY.
Discussion
My estimate for the impact of AMF ($90 to $800 per QALY) does not overlap with my estimate for systemic criminal justice reform ($1.3k to $290k per QALY). I think this is informative, and good news for uncertainty quantification: even though both estimates are very uncertain—they range 2 and 3 orders of magnitude, respectively—we can still tell which one is better.
When comparing GiveDirectly ($160 and $2700 per QALY; mean of $900/QALY) against one standout intervention in the space ($200 to $19K per QALY, with a mean of $5k/QALY), the estimates do overlap, but GiveDirectly is still much better in expectation.
EDIT 22/06/2022. Using the better mean, the above paragraph would be: When comparing GiveDirectly ($160 and $2700 per QALY; mean of $690/QALY) against one standout intervention in the space ($200 to $19K per QALY, with a mean of $837/QALY), the estimates do overlap, but GiveDirectly is still better in expectation.
One limitation of these estimates is that they only model first-order effects. GiveWell does have some estimates of second-order effects (avoiding malaria cases that don’t lead to death, longer-term income increases, etc.) However, for the case of criminal justice interventions, these are harder to estimate. Nonetheless, my strong sense is that the second-order effects of death from malaria or cash transfers are similar to or greater than the second-order effects of temporary imprisonment, and don’t change the relative value of the two cause areas all that much.
Some other sources of model error might be:
QALYs being an inadequate modelling choice: QALYs intuitively have a bound of 1 QALY/year, and might not be the right way to think about certain interventions.
I ignored the cost to the US of keeping someone in prison, as opposed to how that money could have been spent otherwise
I didn’t model the increased productivity of someone outside prison
I didn’t estimate recidivism or increased crime from lower incarceration
I didn’t estimate the cost of pushback, such as lobbying for opposite policies
My estimates of the cost of reform were pretty optimistic.
Of these, I think that not modelling the cost to the US of keeping someone in prison, and not modelling recidivism are one of the weakest aspects of my current model. For a model which tries to incorporate these, see the appendix. So overall, there is likely a degree of model error. But I still think that the small models point to something meaningful.
We can also compare the estimates in this post with other estimates. A lengthy report commissioned by Open Philanthropy on the impacts of incarceration on crime mostly concludes that marginal reduction in crime through more incarceration is non-existent—because the effects of reduced crime while prisoners are in prison are compensated by increased crime when they get out, proportional to the length of their sentence. But the report reasons about short-term effects and marginal changes, e.g., based on RCTs or natural experiments, rather than considering longer-term incentive landscape changes following systemic reform. So for the purposes of judging systemic reform rather than marginal changes, I am inclined to almost completely discount it [7]. That said, my unfamiliarity with the literature is likely one of the main weaknesses of this post.
Open Philanthropy’s own initial casual cause estimations are much more optimistic. In a 2020 interview with Chloe Cockburn, she mentions that Open Philanthropy estimates criminal justice reform to be around 1/4th as valuable as donations to top GiveWell charities, but that she is personally higher based on subjective factors [8].
For illustration, here are a few grants that I don’t think meet the funding bar of being comparable to AMF or GiveDirectly, based on casual browsing of their websites:
The last one struck me as being both particularly bad and relatively easy to evaluate: A letter costs $2.5, about the same as deworming several kids at $0.35 to $0.97 per deworming treatment. But sending a letter intuitively seems significantly less impactful.
Conversely, larger grants, such as, for instance, a $2.5M grant to Color Of Change, are harder to casually evaluate. For example, that particular grant was given to support prosecutorial accountability campaigns and to support Color Of Change’s work with the film Just Mercy. And because the grant was 50% of Color of Change’s budget for one year, I imagine it also subsidized its subsequent activities, such as the campaigns currently featured on its website [10], or the $415k salary of its president [11]. So to the extent that the grant’s funds were used for prosecutorial accountability, they may have been more cost-effective, and to the extent that they were used for other purposes, less so. Overall, I don’t think that estimating the cost-effectiveness of larger grants as the cost-effectiveness of systemic change would be grossly unfair.
Why did Open Philanthropy donate to criminal justice in the first place?
Epistemic status: Speculation.
I will first outline a few different hypotheses about why Open Philanthropy donated to criminal justice, without regard to plausibility:
The Back of the Envelope Calculation Hypothesis
The Value of Information Hypothesis
The Leverage Hypothesis
The Strategic Funder Hypothesis
The Progressive Funders Hypothesis
The “Politics is The Mind Killer” Hypothesis
The Non-Updating Funders Hypothesis
The Moral Tension Hypothesis
I obtained this list by talking to people about my preliminary thoughts when writing this draft. After outlining them, I will discuss which of these I think are most plausible.
The Back of the Envelope Calculation Hypothesis
As highlighted in Open Philanthropy blog posts, early on, it wasn’t clear that GiveWell was going to find as many opportunities as it later did. It was plausible that the bar could have gone down with time. If so, and if one has a rosier outlook on the tractability and value of criminal justice reform, it could plausibly have been competitive with other areas.
Each grant is subject to a cost-effectiveness calculation based on the following formula:
Number of years averted x $50,000 for prison or $100,000 for jail [our valuation of a year of incarceration averted] / 100 [we aim to achieve at least 100x return on investment, and ideally much more] - discounts for causation and implementation uncertainty and multiple attribution of credit > $ grant amount. Not all grants are susceptible to this type of calculation, but we apply it when feasible.
That is, Open Philanthropy’s lower bound for funding criminal justice reform was $500 to $1,000 per year of prison/jail avoided. Per this lower bound, criminal justice reform would be roughly as cost-effective as GiveDirectly. But this bound is much more optimistic than my estimates of the cost-effectiveness of criminal justice reform grants above.
The Value of Information Hypothesis
In 2015, when Open Philanthropy hadn’t invested as much into criminal justice reform, it might have been plausible that relatively little investment might have led to systematic reform. It might have also been plausible that, if found promising, an order of magnitude more funding could have been directed to the cause area.
Commenters in a draft pointed out a second type of information gain: Open Philanthropy might gain experience in grantmaking, learn information, and acquire expertise that would be valuable for other types of giving. In the case of criminal justice reform, I would guess that the specific cause officers—rather than Open Philanthropy as an institution—would gain most of the information. I would also guess that the lessons learnt haven’t generalized to, for instance, pandemic prevention funding advocacy. So my best guess is that the information gained would not make this cause worth it if it otherwise would not have been. But I am uncertain about this.
The Leverage Hypothesis
Even if systemic change itself is not cost-effective, criminal justice reform and adjacent issues attract a large amount of attention anyway. By working in this area, one could gain leverage, for instance:
Leverage over other people’s attention and political will, by investing early in leaders who will be in a position to channel somewhat ephemeral political wills.
Leverage over the grantmaking in the area, by seeding Just Impact
The Strategic Funder Hypothesis
My colleagues raised the hypothesis that Open Philanthropy might have funded criminal justice reform in part because they wanted to look less weird. E.g., “Open Philanthropy/the EA movement has donated to global health, criminal justice reform, preventing pandemics and averting the risks of artificial intelligence” sounds less weird than “...donated to global health, preventing pandemics and averting the risks of artificial intelligence”.
The Progressive Funders Hypothesis
Dustin Moskovitz and Cari Tuna likely have other goals beyond expected utility maximization. Some of these goals might align with the mores of the current left-wing of American society. Or, alternatively, their progressive beliefs might influence and bias their beliefs about what maximizes utility.
On the one hand, I think this would be a mistake. Cause impartiality is one of EA’s major principles, and I think it catalyzes an important part of what we’ve found out about doing good better. But on the other hand, these are not my billions. On the third hand, it seems suboptimal if politically-motivated giving were post-hoc argued to be utility-optimal. If this was the case, I would really have appreciated if their research would have been upfront about this.
The “Politics is The Mind Killer” Hypothesis
In the domain of politics, reasoning degrades, and principal-agent problems arise. And so another way to look at the grants under discussion is that Open Philanthropy flew too close to politics, and was sucked in.
To start, there is a selection effect of people who think an area is the most promising going into it. In addition, there is a principal-agent problem where people working inside a cause area are not really incentivized to look for arguments and evidence that they should be replaced by something better. My sense is that people will tend to give very, very optimistic estimates of impact for their own cause area.
These considerations are general arguments, and they could apply to, for instance, community building or forecasting, with similar force. Though perhaps the warping effects would be stronger for cause areas adjacent to politics.
The Moral Tension Hypothesis
My sense is that Open Philanthropy funders lean a bit more towards conventional morality, whereas philosophical reflection leans more towards expected utility maximization. Managing the tension between these two approaches seems pretty hard, and it shouldn’t be particularly surprising that a few mistakes were made from a utilitarian perspective.
Discussion
In conversation with Open Philanthropy staff, they mentioned that the first three hypotheses —Back of the Envelope, Value of Information and Leverage—sounded most true to them. In conversation with a few other people, mostly longtermists, some thought that the Strategic Funders and the Progressive Funders hypothesis were more likely.
I would make a distinction between what the people who made the decision were thinking at the time, and the selection effects that chose those people. And so, I would think that early on, Open Philanthropy leadership mainly was thinking about back-of-the-envelope calculations, value of information, and leverage. But I would also expect them to have done so somewhat constrainedly. And I expect some of the other hypotheses—particularly the “progressive funders hypothesis”, and the “moral tension hypothesis”—to explain those constraints at least a little.
I am left uncertain about whether and to what extent Open Philanthropy was acting sincerely. It could be that criminal justice reform was just a bet that didn’t pay off. But it could also be the case that some factor put the thumb on the scale and greased the choice to invest in criminal justice reform. In the end, Open Philanthropy is probably heterogenous; it seems likely that some people were acting sincerely, and others with a bit of motivated reasoning.
Why did Open Philanthropy keep donating to criminal justice?
Epistemic status: More speculation
The Inertia Hypothesis
Open Philanthropy wrote about GiveWell’s Top Charities Are (Increasingly) Hard to Beat in 2019. They stopped investing in criminal justice reform in 2021, after giving an additional $100M to the cause area. I’m not sure what happened in the meantime.
In a 2016 blog post explaining worldview diversification, Holden Karnofsky writes:
Currently, we tend to invest resources in each cause up to the point where it seems like there are strongly diminishing returns, or the point where it seems the returns are clearly worse than what we could achieve by reallocating the resources—whichever comes first
Under some assumptions explained in that post, namely that the amounts given to each cause area are balanced to ensure that the values of the marginal grants to each area are similar, worldview diversification would be approximately optimal even from an expected value perspective [12]. My impression is that this monitoring and rebalancing did not happen fast enough in the case of criminal justice reform.
Incongruous as it might ring to my ears, it is also possible that optimizing the allocation of an additional $100M might not have been the most valuable thing for Open Philanthropy’s leadership to have been doing. For instance, exploring new areas, convincing or coordinating with additional billionaires or optimizing other parts of Open Philanthropy’s portfolio might have been more valuable.
The Social Harmony Hypothesis
Firing people is hard. When you structured your bet on a cause area as a bet on a specific person, I imagine that resolving that bet as a negative would be awkward [14].
The Soft Landing Hypothesis
Abruptly stopping funding can really be detrimental for a charity. So Open Philanthropy felt the need to give a soft roll-off that lasts a few years. On the one hand, this is understandable. But on the other hand, it seems that Open Philanthropy might have given two soft landings, one of $50M in 2019, and another $50M in 2021 to spin-off Just Impact.
The Chessmaster Hypothesis
There is probably some calculation or some factor that I am missing. There is nothing disallowing Open Philanthropy from making moves based on private information. In particular, see the discussion on information gains above. Information gains are particularly hard for me to estimate from the outside.
Open Philanthropy took several years and spent an additional $100M on a cause that they could have known was suboptimal. That feels like too much time.
They also arguably gave two different “golden parachutes” when leaving criminal justice reform. The first, in 2019, gave a number of NGOs in the area generous parting donations. The second, in 2021, gave the outgoing program officers $50 million to continue their work.
This might make similar experimentation—e.g., hiring a program officer for a new cause area, and committing to it only if it goes well—much more expensive. It’s not clear to me that Open Philanthropy would have agreed beforehand to give $100M in “exit grants”.
On Moral Diversification
Open Philanthropy’s donations to criminal justice were part of its global health and development portfolio, and, thus, in theory, not subject to Open Philanthropy’s worldview diversification framework. But in practice, I get the impression that one of the bottlenecks for not noticing sooner that criminal justice reform was likely suboptimal, might have had to do with worldview diversification.
Overall, having a single “bar” across multiple very different programs and outcome measures is an attractive feature because equalizing marginal returns across different programs is a requirement for optimizing the overall allocation of resources
Prior to 2019, we used a “100x” bar based on the units above, the scalability of direct cash transfers to the global poor, and the roughly 100x ratio of high-income country income to GiveDirectly recipient income. As of 2019, we tentatively switched to thinking of “roughly 1,000x” as our bar for new programs, because that was roughly our estimate of the unfunded margin of the top charities recommended by GiveWell
We’re also updating how we measure the DALY burden of a death; our new approach will accord with GiveWell’s moral weights, which value preventing deaths at very young ages differently than implied by a DALY framework. (More)
This post focuses exclusively on how we value different outcomes for humans within Global Health and Wellbeing; when it comes to other outcomes like farm animal welfare or the far future, we practice worldview diversification instead of trying to have a single unified framework for cost-effectiveness analysis. We think it’s an open question whether we should have more internal “worldviews” that are diversified over within the broad Global Health and Wellbeing remit (vs everything being slotted into a unified framework as in this post).
Speaking about Open Philanthropy’s portfolio rather than about criminal justice, instead of strict worldview diversification, one could compare these different cause areas as best one can, strive to figure out better comparisons, and set the marginal impact of grants in each area to be roughly equal. This would better approximate expected value maximization, and it is in fact not too dissimilar to (part of the) the original reasoning for worldview diversification. As explained in the original post, worldview diversification makes the most sense in some contexts and under some assumptions: diminishing returns to each cause, and similar marginal values to more funding.
But somehow, I get the weak impression that worldview diversification (partially) started as an approximation to expected value, and ended up being more of a peace pact between different cause areas. This peace pact disincentivizes comparisons between giving in different cause areas, which then leads to getting their marginal values out of sync.
Instead, I would like to see:
further analysis of alternatives to moral diversification,
more frequent monitoring of whether the assumptions behind moral diversification still make sense,
and a more regular rebalancing of the proportion of funds assigned to each cause according to the value of their marginal grants [13].
On Open Philanthropy’s Openness
After a shallow investigation and reading a few of its public writings, I’m still unsure why exactly Open Philanthropy invested a relatively large amount into this cause area. My impression is that there are some critical details about this that they have not yet written about publicly.
Open Philanthropy’s Rationality
I used to implicitly model Open Philanthropy as a highly intelligent unified agent to which I should likely defer. I now get the impression that there might be a fair amount of politicking, internal division, and some suboptimal decision-making.
I think that this update was larger for me than it might be for others, perhaps because I initially thought very highly of Open Philanthropy. So others who started from a more moderate starting point should make a more minor update, if any.
I still believe that Open Philanthropy is likely one of the best organizations working in the philanthropic space.
Systems that could improve Open Philanthropy’s decision-making
While writing this piece, the uncomfortable thought struck me that if someone had realized in 2017 that criminal justice was suboptimal, it might have been difficult for them to point this out in a way which Open Philanthropy would have found useful. I’m also not sure people would have been actively incentivized to do so.
Once the question is posed, it doesn’t seem hard to design systems that incentivize people to bring potential mistakes to Open Philanthropy’s attention. Below, I consider two options, and I invite commenters to suggest more.
Red teaming
When investing substantial amounts in a new cause area, putting a large monetary bounty on red teams seems a particularly cheap intervention. For instance, one could put a prize on the best red teaming, and a larger bounty on a red teaming output, leading to a change in plans. The recent Criticism Contest is a one-off example which could in theory address Open Philanthropy.
Forecasting systems
Per this recent writeup, Open Philanthropy has predictions made and graded by each cause’s officer, who average about 1 prediction per $1 million moved. The focus of their prediction setup seems to be on learning from past predictions, rather than on using prediction setups to inform decisions before they are made. And it seems like staff tend to make predictions on individual grants, rather than on strategic decisions.
This echoes the findings of a previous report on Prediction Markets in the Corporate Setting: organizations are hesitant to use prediction setups in situations where this would change their most important decisions, or where this would lead to social friction. But this greatly reduces the usefulness of predictions. And in fact, we do know that Open Philanthropy’s prediction setup failed to avoid the pitfalls outlined in this post.
Instead, have a forecasting system which is not restricted to Open Philanthropy staff, which has real-money bets, and which a focuses on using predictions to change decisions, rather than on learning after the fact. Such a system would ask things such as:
whether a key belief underlying the favourable assessment of a grant will later be estimated to be false
whether Open Philanthropy will regret having donated a given grant, or
whether Open Philanthropy will regret some strategic decision, such as going into a cause area, or having set-up such-and-such disbursement schedule,
These questions might be operationalized as:
“In year [x], what probability will [some mechanism] assign to [some belief]?”
“In year [x], what will Open Philanthropy’s best estimate of the value for grant [y] be?” + “In year [x], what will be Open Philanthropy’s bar for funding be?”.
Or, even simpler still, asking directly or “in year [x], will Open Philanthropy regret having made grant [y]?”,
“in year [x], will Open Philanthropy regret having made decision [y]?”,
There would be a few challenges in creating such a forecasting system in a way that would be useful to Open Philanthropy:
It would be difficult to organize this at scale.
If open to the public, and if Open Philanthropy was listening to them, it might be easy and desirable to manipulate them.
If structured as a prediction market, it might not be worth it to participate unless the market also yielded interest.
If Open Philanthropy had enough bandwidth to create a forecasting system, it would also have been capable of monitoring the criminal justice reform situation more closely (?)
It would be operationally or legally complex
Prediction markets are mostly illegal in the US
In 2018, the best way to structure this may have been as follows: Open Philanthropy decides on a probability and a metric of success and offers a trusted set of advisors to bet against the metric being satisfied. Note that the metric can be fuzzy, e.g., “Open Phil employee X will estimate this grant to have been worth it”.
With time, advisors who can predict how Open Philanthropy will change its mind would acquire more money and thus more independent influence in the world. This isn’t bullet-proof—for instance, advisors would have an incentive to make Open Philanthropy be wrong so that they can bet against them—but it’d be a good start.
Note that the pathway to impact of making monetary bets wouldn’t only be to change Open Philanthropy’s decisions—which past analysis suggests would be difficult—but also to transfer wealth to altruistic actors that have better models of the world.
The TarasBob method for maximizing predictive accuracy
In July 2022, there still aren’t great forecasting systems that could deal with this problem. The closest might be Manifold Markets, which allows for the fast creation of different markets and the transfer of funds to charities, which gives some monetary value to their tokens. In any case, because setting up such a system might be laborious, one could instead just offer to set such a system up only upon request.
I am also excited about a few projects that will provide possibly scalable prediction markets, which are set to launch in the next few months and could be used for that purpose. My forecasting newsletter will have announcements when these projects launch.
Conclusion
Open Philanthropy spent $200M on criminal justice reform, $100M of which came after their own estimates concluded that it wasn’t as effective as other global health and development interventions. I think Open Philanthropy could have done better.
And I am left confused about why Open Philanthropy did not in fact do better. Part of this may have been their unique approach of worldview diversification. Part of this may have been the political preferences of their funders. And part of this may have been their more optimistic Fermi estimates. I oscillate between thinking “I, a young grasshopper, do not understand”, and “this was clearly suboptimal from the beginning, and obviously so”.
Still, Open Philanthropy did end up parting ways with their criminal justice reform team. Perhaps forecasting systems or red teams would have accelerated their decision-making on this topic.
Acknowledgements
Thanks to Linch Zhang, Max Ra, Damon Pourtahmaseb-Sasi, Sam Nolan, Lawrence Newport, Eli Lifland, Gavin Leech, Alex Lawsen, Hauke Hillebrandt, Ozzie Gooen, Aaron Gertler, Joel Becker and others for their comments and suggestions.
This post is a project by the Quantified Uncertainty Research Institute (QURI). The language used to express probabilities distributions used throughout the post is Squiggle, which is being developed by QURI.
Appendix: Incorporating savings and the cost of recidivism.
// First section: Same as before
initialPrisonPopulation = 1.8M to 2.5M # Data for 2022 prison population has not yet been published, though this estimate is perhaps too wide.
reductionInPrisonPopulation = 0.25 to 0.75
badnessOfPrisonInQALYs = 0.2 to 6 # 80% as good as being alive to 5 times worse than living is good
accelerationInYears = 5 to 50
probabilityOfSuccess = 0.01 to 0.1 # 1% to 10%.
estimateQALYs = initialPrisonPopulation * reductionInPrisonPopulation * badnessOfPrisonInQALYs * accelerationInYears * probabilityOfSuccess
cost = 2B to 20B
costEffectivenessPerQALY = cost / estimateQALYs
// New section: Costs and savings
numPrisonersFreed = initialPrisonPopulation * reductionInPrisonPopulation * accelerationInYears * probabilityOfSuccess
savedCosts = numPrisonersFreed * (14k to 70k)
savedQALYsFromCosts = savedCosts / 50k
probabilityOfRecidivism = 0.3 to 0.7
numIncidentsUntilCaughtAgain = 1 to 10 // uncertain; look at what percentage of different types of crimes are reported and solved.
costPerIncident = 1k to 50k
lostCostsFromRecidivism = numPrisonersFreed * probabilityOfRecidivism * costPerIncident
lostQALYsFromRecidivism = lostCostsFromRecidivism/50k
costPerQALYIncludingCostsAndIncludingRecidivism = truncateLeft(cost / (estimateQALYs + savedQALYsFromCosts - lostQALYsFromRecidivism), 0)
// ^ truncateLeft needed because division is very numerically unstable.
// Display
// costPerQALYIncludingCostsAndIncludingRecidivism
// ^ increase the number of samples to 10000 and uncomment this line
A review from Open Philanthropy on the impacts of incarceration on crime concludes by saying that “The analysis performed here suggests that it is hard to argue from high-credibility evidence that at typical margins in the US today, decarceration would harm society”. But “high-credibility evidence” does a lot of the heavy lifting: I have a pretty strong prior that incentives matter, and the evidence is weak. In particular, the evidence provided is a) mostly at the margin, and b) mostly using evidence based on short-term change. So I’m slightly convinced that for small changes, the effect in the short term—e.g., within one generation—is small. But if prison sentences are marginally reduced in length or in quantity, I still end up with the impression that crime would marginally rise in the longer term, as crimes become marginally more worth it. Conversely, if sentences are reduced more than in the margin, common sense suggests that crime will increase, as observed in, for instance, San Francisco (note:or not; see this comment and/or this investigation.)
Footnotes
[0]. This number is $138.8 different than the $138.8M given in Open Philanthropy’s website, which is probably not up to date with their grants database.
[1]. Note that this paragraph is written from my perspective doing a postmortem, rather than aiming to summarize what they thought at the time.
[2]. Note that restorative justice is normally suggested as a total replacement for punitive justice. But I think that pushing back punitive justice until it is incentive compatible and then applying restorative justice frameworks would also work, and would encounter less resistance.
[3]. Subjective estimate based on the US having many more guns, a second amendment, a different culture, more of a drug problem.
[4]. Subjective estimate; I think it would take 1-2 orders of magnitude more investment than the already given $2B.
[5]. Note that QALYs refers to a specific construct. This has led people to come up with extensions and new definitions, e.g., the WALY (wellbeing-adjusted), HALY (happiness-adjusted), DALY (disability-adjusted), and SALY (suffering-adjusted) life years. But throughout this post, I’m stretching that definition and mostly thinking about “QALYs as they should have been”.
[6]. Initially, Squiggle was making these calculations using monte-carlo simulations. However, operations multiplying and dividing lognormals can be done analytically. I extracted the functionality to do so into Simple Squiggle, and then helped the main Squiggle branch compute the model analytically.
Simple Squiggle does validate the model as producing an interval of $1.3k to $290k. To check this, feed `1000000000 * (2 to 20) / ((1000000 * 1.5 to 2.5) * 0.25 to 0.75 * 0.2 to 6 * 5 to 50 * 0.01 to 0.1 * 0.5 to 1 )` into it
[7]. To elaborate on this, as far as I understand, to estimate the impact of incarceration, the reports’ best source of evidence are randomized trials or natural experiments, e.g., harsher judges randomly assigned, arbitrary threshold changes resulting from changes in guidelines or policy, etc. But these methods will tend to estimate short-term changes, rather than longer term (e.g., intergenerational) changes.
And I would give substantial weight to lighter sentencing in fact making it more worth it to commit crime. See Lagerros’ Unconscious Economics.
This topic also has very large number of degrees of choice (e.g., see p. 133 on lowballing the cost of murder on account of it being rare), which I am inclined to be suspicious about.
The report has a “devil’s advocate case”. But I think that it could have been much harsher, by incorporating hard-to-estimate long-term incentive changes.
[8]. Excerpt, with some light editing to exclude stutters:
With a lot of hedging and assumptions and guessing, I think that we can show that we were at around 250x, versus GiveWell, which is at more like 1000x [9]. So according to Open Philanthropy, if you’re just like, what’s the place where I can put my dollar that does the most good, you should give to GiveWell, I think.
That said, I would say well, first of all, if you feel that now’s the time, now’s a particular unique and important time to be working on this when there is a lot of traction, that puts a thumb on the scale more towards this. Deworming was very important 10 years ago, will be very important in 10 years. I think that’s different than this issue, where you have this moments where we can actually make a lot of change, where a boost of cash is good.
And then second, that there is a lot that’s not captured in that 250x.
And then third, that 250x is based on the assumption that a year of freedom from prison is worth $50k, and a year of freedom from jail is worth $100k. I think a jail bed gone empty for a year could be worth $250k, for example.
So, I’m telling you this, I don’t say this to normal people, I have no idea what I’m talking about. But for EA folks, I think we’re closer to 1000x than I’ve been able to show thus far. But if you want to be like “I’m helping the most that I can be certain about” yeah, for sure, go give your money to deworming, that’s still probably true.
[9]: “1000x” (resp. 250x) refers to being 1000 times (resp. 250 times) more cost-effective than giving a dollar to someone with $50k of annual income; see here.
[10]. As I was writing this, it featured campaigns calling for common carriers to drop Fox, and for Amazon and Twitch to carry out racial equity audits. But these have since cycled through.
[11]. It rose from $216k in 2016 to $415k in 2019. Honestly I’m not even sure this unjustified; he could probably be a very highly paid political consultant, and a high salary is in fact a strong signal that his funders think that he shouldn’t be.
[12]. This excludes considerations around how much to donate each year.
[13]. A side effect of spinning off Just Impact with a very sizeable initial endowment is that the careers of the Open Philanthropy officers involved appear to continue progressing. Commenters pointed out that this might make it easier to hire talent. But coming from a forecasting background which has some emphasis in proper scoring rules, this seems personally unappealing.
[14]. Technically, according to the shape of the values of their grants and the expected future shape, not just the values of the marginal grant.
I also considered suggesting a ruthless Hunger Games-style fight between the representatives of different cause areas, with the winner getting all the resources regardless of diminishing returns. But I concluded that this was likely not possible in practice, and also that the neartermists would probably be in better shape.
A Critical Review of Open Philanthropy’s Bet On Criminal Justice Reform
Epistemic status: Dwelling on the negatives.
Hello AstralCodexTen readers! You might also enjoy this series on estimating value, or my forecasting newsletter. And for the estimation language used throughout this post, see Squiggle.—Nuño.
Summary
From 2013 to 2021, Open Philanthropy donated $200M to criminal justice reform. My best guess is that, from a utilitarian perspective, this was likely suboptimal. In particular, I am fairly sure that it was possible to realize sooner that the area was unpromising and act on that earlier on.
In this post, I first present the background for Open Philanthropy’s grants on criminal justice reform, and the abstract case for considering it a priority. I then estimate that criminal justice grants were distinctly worse than other grants in the global health and development portfolio, such as those to GiveDirectly or AMF.
I speculate about why Open Philanthropy donated to criminal justice in the first place, and why it continued donating. I end up uncertain about to what extent this was a sincere play based on considerations around the value of information and learning, and to what extent it was determined by other factors, such as the idiosyncratic preferences of Open Philanthropy’s funders, human fallibility and slowness, paying too much to avoid social awkwardness, “worldview diversification” being an imperfect framework imperfectly applied, or it being tricky to maintain a balance between conventional morality and expected utility maximization. In short, I started out being skeptical that a utilitarian, left alone, spontaneously starts exploring criminal justice reform in the US as a cause area, and to some degree I still think that upon further investigation, though I still have significant uncertainty.
I then outline my updates about Open Philanthropy. Personally, I updated downwards on Open Philanthropy’s decision speed, rationality and degree of openness, from an initially very high starting point. I also provide a shallow analysis of Open Philanthropy’s worldview diversification strategy and suggest that they move to a model where regular rebalancing roughly equalizes the marginal expected values for the grants in each cause area. Open Philanthropy is doing that for its global health and development portfolio anyways.
Lastly, I brainstorm some mechanisms which could have accelerated and improved Open Philanthropy’s decision-making and suggest red teams and monetary bets or prediction markets as potential avenues of investigation.
Throughout this piece, my focus is aimed at thinking clearly and expressing myself clearly. I understand that this might come across as impolite or unduly harsh. However, I think that providing uncertain and perhaps flawed criticism is still worth it, in expectation. I would like to note that I still respect Open Philanthropy and think that it’s one of the best philanthropic organizations around.
Open Philanthropy staff reviewed this post prior to publication.
Index
Background information
What is the case for Criminal Justice Reform?
What is the cost-effectiveness of criminal justice grants?
Why did Open Philanthropy donate to criminal justice in the first place?
Why did Philanthropy keep donating to criminal justice?
What conclusions can we reach from this?
Systems that could have optimized Open Philanthropy’s impact
Conclusion
Background information
From 2013 to 2021, Open Philanthropy distributed $199,574,123 to criminal justice reform [0]. In 2015, they hired Chloe Cockburn as a program officer, following a “stretch goal” for the year. They elaborated on their method and reasoning on The Process of Hiring our First Cause-Specific Program Officer.
In that blog post, they described their expansion into the criminal justice reform space as substantially a “bet on Chloe”. Overall, the post was very positive about Chloe (more on red teams below). But the post expressed some reservations because “Chloe has a generally different profile from the sorts of people GiveWell has hired in the past. In particular, she is probably less quantitatively inclined than most employees at GiveWell. This isn’t surprising or concerning—most GiveWell employees are Research Analysts, and we see the Program Officer role as calling for a different set of abilities. That said, it’s possible that different reasoning styles will lead to disagreement at times. We think of this as only a minor concern.” In hindsight, it seems plausible to me that this relative lack of quantitative inclination played a role in Open Philanthropy making comparatively suboptimal grants in the criminal justice space [1].
In mid-2019, Open Philanthropy published a blog post titled GiveWell’s Top Charities Are (Increasingly) Hard to Beat. It explained that, with GiveWell’s expansion into researching more areas, Open Philanthropy expected that there would be enough room for more funding for charities that were as good as GiveWell’s top charities. Thus, causes like Criminal Justice Reform looked less promising.
In the months following that blog post, Open Philanthropy donations to Criminal Justice reform spike, with multi-million, multi-year grants going to Impact Justice ($4M), Alliance for Safety and Justice ($10M), National Council for Incarcerated and Formerly Incarcerated Women and Girls ($2.25M), Essie Justice Group ($3M), Texas Organizing Project ($4.2M), Color Of Change Education Fund ($2.5M) and The Justice Collaborative ($7.8M).
Initially, I thought that might be because of an expectation of winding down. However, other Open Philanthropy cause areas also show a similar pattern of going up in 2019, perhaps at the expense of spending on Global Health and Development for that year:
In 2021, Open Philanthropy spun out its Criminal Justice Reform department as a new organization: Just Impact. Open Philanthropy seeded Just Impact with $50M. Their parting blog post explains their thinking: that Global Health and Development interventions have significantly better cost-effectiveness.
What is the case for Criminal Justice Reform?
Note: This section briefly reviews my own understanding of this area. For a more canonical source, see Open Philanthropy’s strategy document on criminal justice reform.
There are around 2M people in US prisons and jails. Some are highly dangerous, but a glance at a map of prison population rates per 100k people suggests that the US incarcerates a significantly larger share of its population than most other countries:
Outlining a positive vision for reform is still an area of active work. Still, a first approximation might be as follows:
Criminals should be punished in proportion to an estimate of the harm they have caused, times a factor to account for a less than 100% chance of getting caught, to ensure that crimes are not worth it in expectation. This is in opposition to otherwise disproportionate jail sentences caused by pressures on politicians to appear tough on crime. In addition, criminals then work to provide restitution to the victim, if the victim so desires, per some restorative justice framework [2].
In a best-case scenario, criminal justice reform could achieve somewhere between a 25% reduction in incarceration in the short-term and a 75% reduction in the longer term, bringing the incarceration rate down to only twice that of Spain [4], while maintaining the crime rate constant. Say that $2B to $20B, or 10x to 100x the amount that Open Philanthropy has already spent, would have a 1 to 10% chance of succeeding at that goal [5].
What is the cost-effectiveness of criminal justice grants?
Estimation strategy
In this section, I come up with some estimates of the impact of criminal justice reform, and compare them with some estimates of the impact of GiveWell-style global health and development interventions.
Throughout, I am making the following modelling choices:
I am primarily looking at the impact of systemic change
I am looking at the first-order impacts
I am using subjective estimates
I am primarily looking at the impact of systemic change because many of the largest Open Philanthropy donations were aiming for systemic change, and their individual cost-effectiveness was extremely hard to estimate. For completeness, I do estimate the impacts of a standout intervention as well.
I am looking at the first-order impacts on prisoners and GiveWell recipients, rather than at the effects on their communities. My strong guess is that the story the second-order impacts would tell—e.g., harms to the community from death or reduced earnings in the case of malaria, harms from absence and reduced earnings in the case of imprisonment)—wouldn’t change the relative values of the two cause areas.
After presenting my estimates, I discuss their limitations.
Simple model for systemic change
Using those what I consider to be optimistic assumptions over first-order effects, I come up with the following Squiggle model:
That model produces the following distribution:
mean(cost)/mean(estimateQALYs)is equal to $8160/QALYThis model estimates that criminal justice reform buys one QALY [6](quality-adjusted life year) for $76k, on average. But the model is very uncertain, and its 90% confidence interval is $1.3k to ~$290k per QALY. It assigns a 50% chance to it costing less than ~$19k. For a calculation that instead looks at more marginal impact, see here.
EDIT 22/06/2022: Commenters pointed out that the mean of
cost / estimateQALYsin the chart above isn’t the right quantity to look at in the chart above.mean(cost)/mean(estimateQALYs)is probably a better representation of “expected cost per QALY. That quantity is $8160/QALY for the above model. If one looks at1/mean(estimateQALYs/cost), this is $5k per QALY. Overall I would instead recommend looking at the 90% confidence intervals, rather at the means. See this comment thread for discussion. I’ve added notes below each model.Simple model for a standout criminal justice reform intervention
Some grants in criminal justice reform might beat systemic reform. I think this might be the case for closing Rikers, bail reform, and prosecutorial accountability:
Rikers is a large and particularly bad prison.
Bail reform seems like a well-defined objective that could affect many people at once.
Prosecutorial accountability could get a large multiplier over systemic reform by focusing on the prosecutors in districts that hold very large prison populations.
For instance, for the case of Rikers, I can estimate:
mean(cost)/mean(estimateQALYs)is $837/QALYSimple model for GiveWell charities
Against Malaria Foundation
Using a similar estimation for the Against Malaria Foundation:
mean(costPerLife)/mean(valueOfSavedLife)is $245/QALYIts 90% confidence interval is $90 to ~$800 per QALY, and I likewise validated this with Simple Squiggle. Notice that this interval is disjoint with the estimate for criminal justice reform of $1.3k to $290k.
GiveDirectly
One might argue that AMF is too strict a comparison and that one should instead compare criminal justice reform to the marginal global health and development grant. Recently, my colleague Sam Nolan quantified the uncertainty in GiveDirectly’s estimate of impact. He arrived at a final estimate of ~$120 to ~$960 per doubling of consumption for one year.
The conversion between a doubling of consumption and a QALY is open to some uncertainty. For instance:
GiveWell estimates it about equal based on the different weights given to saving people of different ages—a factor of ~0.8 to 1.3, based on some eye-balling from this spreadsheet.
GiveWell recently updated their weighings to give a DALY (similar to a QALY) a value of around ~2 doublings of income.
Commenters pointed out that few people would trade half their life to double their income, and that for them a conversion factor around 0.2 might be more appropriate. But they are much wealthier than the average GiveDirectly recipient.
Using a final adjustment of 0.2 to 1.3 QALYs per doubling of consumption (which has a mean of 0.6 QALYs/doubling), I come up with the following model an estimate:
mean(costPerDoublingOfConsumption)/mean(qalysPerDoublingOfConsumption)is $690/QALYThis has a 90% confidence interval between $160 and $2700 per QALY.
Discussion
My estimate for the impact of AMF ($90 to $800 per QALY) does not overlap with my estimate for systemic criminal justice reform ($1.3k to $290k per QALY). I think this is informative, and good news for uncertainty quantification: even though both estimates are very uncertain—they range 2 and 3 orders of magnitude, respectively—we can still tell which one is better.
When comparing GiveDirectly ($160 and $2700 per QALY; mean of $900/QALY) against one standout intervention in the space ($200 to $19K per QALY, with a mean of $5k/QALY), the estimates do overlap, but GiveDirectly is still much better in expectation.
EDIT 22/06/2022. Using the better mean, the above paragraph would be: When comparing GiveDirectly ($160 and $2700 per QALY; mean of $690/QALY) against one standout intervention in the space ($200 to $19K per QALY, with a mean of $837/QALY), the estimates do overlap, but GiveDirectly is still better in expectation.
One limitation of these estimates is that they only model first-order effects. GiveWell does have some estimates of second-order effects (avoiding malaria cases that don’t lead to death, longer-term income increases, etc.) However, for the case of criminal justice interventions, these are harder to estimate. Nonetheless, my strong sense is that the second-order effects of death from malaria or cash transfers are similar to or greater than the second-order effects of temporary imprisonment, and don’t change the relative value of the two cause areas all that much.
Some other sources of model error might be:
QALYs being an inadequate modelling choice: QALYs intuitively have a bound of 1 QALY/year, and might not be the right way to think about certain interventions.
I ignored the cost to the US of keeping someone in prison, as opposed to how that money could have been spent otherwise
I didn’t model the increased productivity of someone outside prison
I didn’t estimate recidivism or increased crime from lower incarceration
I didn’t estimate the cost of pushback, such as lobbying for opposite policies
My estimates of the cost of reform were pretty optimistic.
Of these, I think that not modelling the cost to the US of keeping someone in prison, and not modelling recidivism are one of the weakest aspects of my current model. For a model which tries to incorporate these, see the appendix. So overall, there is likely a degree of model error. But I still think that the small models point to something meaningful.
We can also compare the estimates in this post with other estimates. A lengthy report commissioned by Open Philanthropy on the impacts of incarceration on crime mostly concludes that marginal reduction in crime through more incarceration is non-existent—because the effects of reduced crime while prisoners are in prison are compensated by increased crime when they get out, proportional to the length of their sentence. But the report reasons about short-term effects and marginal changes, e.g., based on RCTs or natural experiments, rather than considering longer-term incentive landscape changes following systemic reform. So for the purposes of judging systemic reform rather than marginal changes, I am inclined to almost completely discount it [7]. That said, my unfamiliarity with the literature is likely one of the main weaknesses of this post.
Open Philanthropy’s own initial casual cause estimations are much more optimistic. In a 2020 interview with Chloe Cockburn, she mentions that Open Philanthropy estimates criminal justice reform to be around 1/4th as valuable as donations to top GiveWell charities, but that she is personally higher based on subjective factors [8].
For illustration, here are a few grants that I don’t think meet the funding bar of being comparable to AMF or GiveDirectly, based on casual browsing of their websites:
$600k: Essie Justice Group — General Support
$500k: LatinoJustice — Work to End Mass Incarceration
$261k: The Soze Agency — Returning Citizens Project
$255k: Mijente — Criminal Justice Reform
$200k: Justice Strategies — General Support
$100k: ReFrame Mentorship — General Support
$100k: Cosecha, general support. (part 1, part 2)
$10k: Photo Patch Foundation — General Support
The last one struck me as being both particularly bad and relatively easy to evaluate: A letter costs $2.5, about the same as deworming several kids at $0.35 to $0.97 per deworming treatment. But sending a letter intuitively seems significantly less impactful.
Conversely, larger grants, such as, for instance, a $2.5M grant to Color Of Change, are harder to casually evaluate. For example, that particular grant was given to support prosecutorial accountability campaigns and to support Color Of Change’s work with the film Just Mercy. And because the grant was 50% of Color of Change’s budget for one year, I imagine it also subsidized its subsequent activities, such as the campaigns currently featured on its website [10], or the $415k salary of its president [11]. So to the extent that the grant’s funds were used for prosecutorial accountability, they may have been more cost-effective, and to the extent that they were used for other purposes, less so. Overall, I don’t think that estimating the cost-effectiveness of larger grants as the cost-effectiveness of systemic change would be grossly unfair.
Why did Open Philanthropy donate to criminal justice in the first place?
Epistemic status: Speculation.
I will first outline a few different hypotheses about why Open Philanthropy donated to criminal justice, without regard to plausibility:
The Back of the Envelope Calculation Hypothesis
The Value of Information Hypothesis
The Leverage Hypothesis
The Strategic Funder Hypothesis
The Progressive Funders Hypothesis
The “Politics is The Mind Killer” Hypothesis
The Non-Updating Funders Hypothesis
The Moral Tension Hypothesis
I obtained this list by talking to people about my preliminary thoughts when writing this draft. After outlining them, I will discuss which of these I think are most plausible.
The Back of the Envelope Calculation Hypothesis
As highlighted in Open Philanthropy blog posts, early on, it wasn’t clear that GiveWell was going to find as many opportunities as it later did. It was plausible that the bar could have gone down with time. If so, and if one has a rosier outlook on the tractability and value of criminal justice reform, it could plausibly have been competitive with other areas.
For instance, per Open Philanthropy’s estimations:
That is, Open Philanthropy’s lower bound for funding criminal justice reform was $500 to $1,000 per year of prison/jail avoided. Per this lower bound, criminal justice reform would be roughly as cost-effective as GiveDirectly. But this bound is much more optimistic than my estimates of the cost-effectiveness of criminal justice reform grants above.
The Value of Information Hypothesis
In 2015, when Open Philanthropy hadn’t invested as much into criminal justice reform, it might have been plausible that relatively little investment might have led to systematic reform. It might have also been plausible that, if found promising, an order of magnitude more funding could have been directed to the cause area.
Commenters in a draft pointed out a second type of information gain: Open Philanthropy might gain experience in grantmaking, learn information, and acquire expertise that would be valuable for other types of giving. In the case of criminal justice reform, I would guess that the specific cause officers—rather than Open Philanthropy as an institution—would gain most of the information. I would also guess that the lessons learnt haven’t generalized to, for instance, pandemic prevention funding advocacy. So my best guess is that the information gained would not make this cause worth it if it otherwise would not have been. But I am uncertain about this.
The Leverage Hypothesis
Even if systemic change itself is not cost-effective, criminal justice reform and adjacent issues attract a large amount of attention anyway. By working in this area, one could gain leverage, for instance:
Leverage over other people’s attention and political will, by investing early in leaders who will be in a position to channel somewhat ephemeral political wills.
Leverage over the grantmaking in the area, by seeding Just Impact
The Strategic Funder Hypothesis
My colleagues raised the hypothesis that Open Philanthropy might have funded criminal justice reform in part because they wanted to look less weird. E.g., “Open Philanthropy/the EA movement has donated to global health, criminal justice reform, preventing pandemics and averting the risks of artificial intelligence” sounds less weird than “...donated to global health, preventing pandemics and averting the risks of artificial intelligence”.
The Progressive Funders Hypothesis
Dustin Moskovitz and Cari Tuna likely have other goals beyond expected utility maximization. Some of these goals might align with the mores of the current left-wing of American society. Or, alternatively, their progressive beliefs might influence and bias their beliefs about what maximizes utility.
On the one hand, I think this would be a mistake. Cause impartiality is one of EA’s major principles, and I think it catalyzes an important part of what we’ve found out about doing good better. But on the other hand, these are not my billions. On the third hand, it seems suboptimal if politically-motivated giving were post-hoc argued to be utility-optimal. If this was the case, I would really have appreciated if their research would have been upfront about this.
The “Politics is The Mind Killer” Hypothesis
In the domain of politics, reasoning degrades, and principal-agent problems arise. And so another way to look at the grants under discussion is that Open Philanthropy flew too close to politics, and was sucked in.
To start, there is a selection effect of people who think an area is the most promising going into it. In addition, there is a principal-agent problem where people working inside a cause area are not really incentivized to look for arguments and evidence that they should be replaced by something better. My sense is that people will tend to give very, very optimistic estimates of impact for their own cause area.
These considerations are general arguments, and they could apply to, for instance, community building or forecasting, with similar force. Though perhaps the warping effects would be stronger for cause areas adjacent to politics.
The Moral Tension Hypothesis
My sense is that Open Philanthropy funders lean a bit more towards conventional morality, whereas philosophical reflection leans more towards expected utility maximization. Managing the tension between these two approaches seems pretty hard, and it shouldn’t be particularly surprising that a few mistakes were made from a utilitarian perspective.
Discussion
In conversation with Open Philanthropy staff, they mentioned that the first three hypotheses —Back of the Envelope, Value of Information and Leverage—sounded most true to them. In conversation with a few other people, mostly longtermists, some thought that the Strategic Funders and the Progressive Funders hypothesis were more likely.
I would make a distinction between what the people who made the decision were thinking at the time, and the selection effects that chose those people. And so, I would think that early on, Open Philanthropy leadership mainly was thinking about back-of-the-envelope calculations, value of information, and leverage. But I would also expect them to have done so somewhat constrainedly. And I expect some of the other hypotheses—particularly the “progressive funders hypothesis”, and the “moral tension hypothesis”—to explain those constraints at least a little.
I am left uncertain about whether and to what extent Open Philanthropy was acting sincerely. It could be that criminal justice reform was just a bet that didn’t pay off. But it could also be the case that some factor put the thumb on the scale and greased the choice to invest in criminal justice reform. In the end, Open Philanthropy is probably heterogenous; it seems likely that some people were acting sincerely, and others with a bit of motivated reasoning.
Why did Open Philanthropy keep donating to criminal justice?
Epistemic status: More speculation
The Inertia Hypothesis
Open Philanthropy wrote about GiveWell’s Top Charities Are (Increasingly) Hard to Beat in 2019. They stopped investing in criminal justice reform in 2021, after giving an additional $100M to the cause area. I’m not sure what happened in the meantime.
In a 2016 blog post explaining worldview diversification, Holden Karnofsky writes:
Under some assumptions explained in that post, namely that the amounts given to each cause area are balanced to ensure that the values of the marginal grants to each area are similar, worldview diversification would be approximately optimal even from an expected value perspective [12]. My impression is that this monitoring and rebalancing did not happen fast enough in the case of criminal justice reform.
Incongruous as it might ring to my ears, it is also possible that optimizing the allocation of an additional $100M might not have been the most valuable thing for Open Philanthropy’s leadership to have been doing. For instance, exploring new areas, convincing or coordinating with additional billionaires or optimizing other parts of Open Philanthropy’s portfolio might have been more valuable.
The Social Harmony Hypothesis
Firing people is hard. When you structured your bet on a cause area as a bet on a specific person, I imagine that resolving that bet as a negative would be awkward [14].
The Soft Landing Hypothesis
Abruptly stopping funding can really be detrimental for a charity. So Open Philanthropy felt the need to give a soft roll-off that lasts a few years. On the one hand, this is understandable. But on the other hand, it seems that Open Philanthropy might have given two soft landings, one of $50M in 2019, and another $50M in 2021 to spin-off Just Impact.
The Chessmaster Hypothesis
There is probably some calculation or some factor that I am missing. There is nothing disallowing Open Philanthropy from making moves based on private information. In particular, see the discussion on information gains above. Information gains are particularly hard for me to estimate from the outside.
What conclusions can we reach from this?
On Open Philanthropy’s Observe–Orient–Decide–Act loops
Open Philanthropy took several years and spent an additional $100M on a cause that they could have known was suboptimal. That feels like too much time.
They also arguably gave two different “golden parachutes” when leaving criminal justice reform. The first, in 2019, gave a number of NGOs in the area generous parting donations. The second, in 2021, gave the outgoing program officers $50 million to continue their work.
This might make similar experimentation—e.g., hiring a program officer for a new cause area, and committing to it only if it goes well—much more expensive. It’s not clear to me that Open Philanthropy would have agreed beforehand to give $100M in “exit grants”.
On Moral Diversification
Open Philanthropy’s donations to criminal justice were part of its global health and development portfolio, and, thus, in theory, not subject to Open Philanthropy’s worldview diversification framework. But in practice, I get the impression that one of the bottlenecks for not noticing sooner that criminal justice reform was likely suboptimal, might have had to do with worldview diversification.
In Technical Updates to Our Global Health and Wellbeing Cause Prioritization Framework, Peter Favaloro and Alexander Berger write:
Speaking about Open Philanthropy’s portfolio rather than about criminal justice, instead of strict worldview diversification, one could compare these different cause areas as best one can, strive to figure out better comparisons, and set the marginal impact of grants in each area to be roughly equal. This would better approximate expected value maximization, and it is in fact not too dissimilar to (part of the) the original reasoning for worldview diversification. As explained in the original post, worldview diversification makes the most sense in some contexts and under some assumptions: diminishing returns to each cause, and similar marginal values to more funding.
But somehow, I get the weak impression that worldview diversification (partially) started as an approximation to expected value, and ended up being more of a peace pact between different cause areas. This peace pact disincentivizes comparisons between giving in different cause areas, which then leads to getting their marginal values out of sync.
Instead, I would like to see:
further analysis of alternatives to moral diversification,
more frequent monitoring of whether the assumptions behind moral diversification still make sense,
and a more regular rebalancing of the proportion of funds assigned to each cause according to the value of their marginal grants [13].
On Open Philanthropy’s Openness
After a shallow investigation and reading a few of its public writings, I’m still unsure why exactly Open Philanthropy invested a relatively large amount into this cause area. My impression is that there are some critical details about this that they have not yet written about publicly.
Open Philanthropy’s Rationality
I used to implicitly model Open Philanthropy as a highly intelligent unified agent to which I should likely defer. I now get the impression that there might be a fair amount of politicking, internal division, and some suboptimal decision-making.
I think that this update was larger for me than it might be for others, perhaps because I initially thought very highly of Open Philanthropy. So others who started from a more moderate starting point should make a more minor update, if any.
I still believe that Open Philanthropy is likely one of the best organizations working in the philanthropic space.
Systems that could improve Open Philanthropy’s decision-making
While writing this piece, the uncomfortable thought struck me that if someone had realized in 2017 that criminal justice was suboptimal, it might have been difficult for them to point this out in a way which Open Philanthropy would have found useful. I’m also not sure people would have been actively incentivized to do so.
Once the question is posed, it doesn’t seem hard to design systems that incentivize people to bring potential mistakes to Open Philanthropy’s attention. Below, I consider two options, and I invite commenters to suggest more.
Red teaming
When investing substantial amounts in a new cause area, putting a large monetary bounty on red teams seems a particularly cheap intervention. For instance, one could put a prize on the best red teaming, and a larger bounty on a red teaming output, leading to a change in plans. The recent Criticism Contest is a one-off example which could in theory address Open Philanthropy.
Forecasting systems
Per this recent writeup, Open Philanthropy has predictions made and graded by each cause’s officer, who average about 1 prediction per $1 million moved. The focus of their prediction setup seems to be on learning from past predictions, rather than on using prediction setups to inform decisions before they are made. And it seems like staff tend to make predictions on individual grants, rather than on strategic decisions.
This echoes the findings of a previous report on Prediction Markets in the Corporate Setting: organizations are hesitant to use prediction setups in situations where this would change their most important decisions, or where this would lead to social friction. But this greatly reduces the usefulness of predictions. And in fact, we do know that Open Philanthropy’s prediction setup failed to avoid the pitfalls outlined in this post.
Instead, have a forecasting system which is not restricted to Open Philanthropy staff, which has real-money bets, and which a focuses on using predictions to change decisions, rather than on learning after the fact. Such a system would ask things such as:
whether a key belief underlying the favourable assessment of a grant will later be estimated to be false
whether Open Philanthropy will regret having donated a given grant, or
whether Open Philanthropy will regret some strategic decision, such as going into a cause area, or having set-up such-and-such disbursement schedule,
These questions might be operationalized as:
“In year [x], what probability will [some mechanism] assign to [some belief]?”
“In year [x], what will Open Philanthropy’s best estimate of the value for grant [y] be?” + “In year [x], what will be Open Philanthropy’s bar for funding be?”.
Or, even simpler still, asking directly or “in year [x], will Open Philanthropy regret having made grant [y]?”,
“in year [x], will Open Philanthropy regret having made decision [y]?”,
There would be a few challenges in creating such a forecasting system in a way that would be useful to Open Philanthropy:
It would be difficult to organize this at scale.
If open to the public, and if Open Philanthropy was listening to them, it might be easy and desirable to manipulate them.
If structured as a prediction market, it might not be worth it to participate unless the market also yielded interest.
If Open Philanthropy had enough bandwidth to create a forecasting system, it would also have been capable of monitoring the criminal justice reform situation more closely (?)
It would be operationally or legally complex
Prediction markets are mostly illegal in the US
In 2018, the best way to structure this may have been as follows: Open Philanthropy decides on a probability and a metric of success and offers a trusted set of advisors to bet against the metric being satisfied. Note that the metric can be fuzzy, e.g., “Open Phil employee X will estimate this grant to have been worth it”.
With time, advisors who can predict how Open Philanthropy will change its mind would acquire more money and thus more independent influence in the world. This isn’t bullet-proof—for instance, advisors would have an incentive to make Open Philanthropy be wrong so that they can bet against them—but it’d be a good start.
Note that the pathway to impact of making monetary bets wouldn’t only be to change Open Philanthropy’s decisions—which past analysis suggests would be difficult—but also to transfer wealth to altruistic actors that have better models of the world.
In July 2022, there still aren’t great forecasting systems that could deal with this problem. The closest might be Manifold Markets, which allows for the fast creation of different markets and the transfer of funds to charities, which gives some monetary value to their tokens. In any case, because setting up such a system might be laborious, one could instead just offer to set such a system up only upon request.
I am also excited about a few projects that will provide possibly scalable prediction markets, which are set to launch in the next few months and could be used for that purpose. My forecasting newsletter will have announcements when these projects launch.
Conclusion
Open Philanthropy spent $200M on criminal justice reform, $100M of which came after their own estimates concluded that it wasn’t as effective as other global health and development interventions. I think Open Philanthropy could have done better.
And I am left confused about why Open Philanthropy did not in fact do better. Part of this may have been their unique approach of worldview diversification. Part of this may have been the political preferences of their funders. And part of this may have been their more optimistic Fermi estimates. I oscillate between thinking “I, a young grasshopper, do not understand”, and “this was clearly suboptimal from the beginning, and obviously so”.
Still, Open Philanthropy did end up parting ways with their criminal justice reform team. Perhaps forecasting systems or red teams would have accelerated their decision-making on this topic.
Acknowledgements
Thanks to Linch Zhang, Max Ra, Damon Pourtahmaseb-Sasi, Sam Nolan, Lawrence Newport, Eli Lifland, Gavin Leech, Alex Lawsen, Hauke Hillebrandt, Ozzie Gooen, Aaron Gertler, Joel Becker and others for their comments and suggestions.
This post is a project by the Quantified Uncertainty Research Institute (QURI). The language used to express probabilities distributions used throughout the post is Squiggle, which is being developed by QURI.
Appendix: Incorporating savings and the cost of recidivism.
Epistemic status: These models are extremely rough, and should be used with caution. A more trustworthy approach would use the share of the prison population by type of crime, the chance of recidivism for each crime, and the cost of new offenses by type. Nonetheless, the general approach might be as follows:
A review from Open Philanthropy on the impacts of incarceration on crime concludes by saying that “The analysis performed here suggests that it is hard to argue from high-credibility evidence that at typical margins in the US today, decarceration would harm society”. But “high-credibility evidence” does a lot of the heavy lifting: I have a pretty strong prior that incentives matter, and the evidence is weak. In particular, the evidence provided is a) mostly at the margin, and b) mostly using evidence based on short-term change. So I’m slightly convinced that for small changes, the effect in the short term—e.g., within one generation—is small. But if prison sentences are marginally reduced in length or in quantity, I still end up with the impression that crime would marginally rise in the longer term, as crimes become marginally more worth it. Conversely, if sentences are reduced more than in the margin, common sense suggests that crime will increase, as observed in, for instance, San Francisco (note:or not; see this comment and/or this investigation.)
Footnotes
[0]. This number is $138.8 different than the $138.8M given in Open Philanthropy’s website, which is probably not up to date with their grants database.
[1]. Note that this paragraph is written from my perspective doing a postmortem, rather than aiming to summarize what they thought at the time.
[2]. Note that restorative justice is normally suggested as a total replacement for punitive justice. But I think that pushing back punitive justice until it is incentive compatible and then applying restorative justice frameworks would also work, and would encounter less resistance.
[3]. Subjective estimate based on the US having many more guns, a second amendment, a different culture, more of a drug problem.
[4]. Subjective estimate; I think it would take 1-2 orders of magnitude more investment than the already given $2B.
[5]. Note that QALYs refers to a specific construct. This has led people to come up with extensions and new definitions, e.g., the WALY (wellbeing-adjusted), HALY (happiness-adjusted), DALY (disability-adjusted), and SALY (suffering-adjusted) life years. But throughout this post, I’m stretching that definition and mostly thinking about “QALYs as they should have been”.
[6]. Initially, Squiggle was making these calculations using monte-carlo simulations. However, operations multiplying and dividing lognormals can be done analytically. I extracted the functionality to do so into Simple Squiggle, and then helped the main Squiggle branch compute the model analytically.
Simple Squiggle does validate the model as producing an interval of $1.3k to $290k. To check this, feed `1000000000 * (2 to 20) / ((1000000 * 1.5 to 2.5) * 0.25 to 0.75 * 0.2 to 6 * 5 to 50 * 0.01 to 0.1 * 0.5 to 1 )` into it
[7]. To elaborate on this, as far as I understand, to estimate the impact of incarceration, the reports’ best source of evidence are randomized trials or natural experiments, e.g., harsher judges randomly assigned, arbitrary threshold changes resulting from changes in guidelines or policy, etc. But these methods will tend to estimate short-term changes, rather than longer term (e.g., intergenerational) changes.
And I would give substantial weight to lighter sentencing in fact making it more worth it to commit crime. See Lagerros’ Unconscious Economics.
This topic also has very large number of degrees of choice (e.g., see p. 133 on lowballing the cost of murder on account of it being rare), which I am inclined to be suspicious about.
The report has a “devil’s advocate case”. But I think that it could have been much harsher, by incorporating hard-to-estimate long-term incentive changes.
[8]. Excerpt, with some light editing to exclude stutters:
With a lot of hedging and assumptions and guessing, I think that we can show that we were at around 250x, versus GiveWell, which is at more like 1000x [9]. So according to Open Philanthropy, if you’re just like, what’s the place where I can put my dollar that does the most good, you should give to GiveWell, I think.
That said, I would say well, first of all, if you feel that now’s the time, now’s a particular unique and important time to be working on this when there is a lot of traction, that puts a thumb on the scale more towards this. Deworming was very important 10 years ago, will be very important in 10 years. I think that’s different than this issue, where you have this moments where we can actually make a lot of change, where a boost of cash is good.
And then second, that there is a lot that’s not captured in that 250x.
And then third, that 250x is based on the assumption that a year of freedom from prison is worth $50k, and a year of freedom from jail is worth $100k. I think a jail bed gone empty for a year could be worth $250k, for example.
So, I’m telling you this, I don’t say this to normal people, I have no idea what I’m talking about. But for EA folks, I think we’re closer to 1000x than I’ve been able to show thus far. But if you want to be like “I’m helping the most that I can be certain about” yeah, for sure, go give your money to deworming, that’s still probably true.
[9]: “1000x” (resp. 250x) refers to being 1000 times (resp. 250 times) more cost-effective than giving a dollar to someone with $50k of annual income; see here.
[10]. As I was writing this, it featured campaigns calling for common carriers to drop Fox, and for Amazon and Twitch to carry out racial equity audits. But these have since cycled through.
[11]. It rose from $216k in 2016 to $415k in 2019. Honestly I’m not even sure this unjustified; he could probably be a very highly paid political consultant, and a high salary is in fact a strong signal that his funders think that he shouldn’t be.
[12]. This excludes considerations around how much to donate each year.
[13]. A side effect of spinning off Just Impact with a very sizeable initial endowment is that the careers of the Open Philanthropy officers involved appear to continue progressing. Commenters pointed out that this might make it easier to hire talent. But coming from a forecasting background which has some emphasis in proper scoring rules, this seems personally unappealing.
[14]. Technically, according to the shape of the values of their grants and the expected future shape, not just the values of the marginal grant.
I also considered suggesting a ruthless Hunger Games-style fight between the representatives of different cause areas, with the winner getting all the resources regardless of diminishing returns. But I concluded that this was likely not possible in practice, and also that the neartermists would probably be in better shape.