This is in response to Sarah Constantin's recent post about intellectual dishonesty within the EA community.
I roughly agree with Sarah's main object level points, but I think this essay doesn't sufficiently embody the spirit of cooperative discourse it's trying to promote. I have a lot of thoughts here, but they are building off a few existing essays. (There's been a recent revival over on Less Wrong attempting to make it a better locus for high quality discussion. I don't know if it's especially succeeded, but I think the concepts behind that intended revival and very important)
- Why Our Kind Can't Cooperate (Eliezer Yudkowsky)
- A Return to Discussion (Sarah Constantin)
- The Importance of [Less Wrong, OR another Single Conversational Locus] (Emphasis mine) (Anna Salamon)
- The Four Layers of Intellectual Conversation (Eliezer Yudkowsky)
I think it's important to have all three concepts in context before delving into: - EA has a lying problem (Sarah Constantin)
I recommend reading all of those. But here's a rough summary of what I consider the important bits. (If you want to actually argue with these bits, please read the actual essays before doing so, so you're engaging with the full substance of the idea)
- Intellectuals and contrarians love to argue and nitpick. This is valuable - it produces novel insights, and keeps us honest. BUT it makes it harder to actually work together to achieve things. We need to understand how working-together works on a deep enough level that we can do so without turning into another random institution that's lost it's purpose. (See Why Our Kind... for more)
- Lately, people have tended to talk on social media (Facebook, Tumblr, etc) rather than in formal blogs or forums that encourage longform discussion. This has a few effects. (See A Return to Discussion for more)
- FB discussion is fragmented - it's hard to find everything that's been said on a topic. (And tumblr is even worse)
- It's hard to know whether OTHER people have read a given thing on a topic.
- A related point (not necessarily in "A Return to Discussion" is that social media incentives some of the worst kinda of discussion. People share things quickly, without reflection. People read and respond to things in 5-10 minute bursts, without having time to fully digest them.
- Having a single, long form discussion area that you can expect everyone in an intellectual community to have read, makes it much easier to building knowledge. (And most of human progress is due, not to humans being smart, but being able to stand on the shoulders of giants). Anna Salamon's "Importance of a Single Conversational Locus" is framed around x-risk, but I think it applies to all aspects of EA: the problems the world faces are so huge that they need a higher caliber of thinking and knowledge-building than we currently have in order to solve.
- In order to make true intellectual progress, you need people to be able to make critiques. You also need those critics to expect their criticism to in turn be criticized, so that the criticism is high quality. If a critique turns out to be poorly thought out, we need shared, common knowledge of that so that people don't end up rehashing the same debates.
- And finally, (one of) Sarah's points in "EA has a lying problem" is that, in order to be different from other movements and succeed where they failed, EA needs to hold itself to a higher standard than usual. There's been much criticism of, say, Intentional Insights for doing sketchy, truth-bendy things to gain prestige and power. But that plenty of "high status" people within the EA community do things that are similar, even if to a different degree. We need to be aware of that.
I would not argue as strongly as Sarah does that we shouldn't do it at all, but it's worth periodically calling each other out on it.
Cooperative Epistemology
So my biggest point here, is that we need to be more proactive and mindful about how discussion and knowledge is built upon within the EA community.
To succeed at our goals:
- EA needs to hold itself to a very high intellectual standard (higher than we currently have, probably. In some sense anyway)
- Factions within EA needs to be able to cooperate, share knowledge. Both object level knowledge (i.e. how cost effective is AMF?) and meta/epistemic knowledge like:
- How do we evaluate messy studies
- How do we discuss things online so that people actually put effort into reading and contributing the discussion.
- What kinds of conversational/debate norms lead people to be more transparent.
- We need to be able to apply all the knowledge to go out and accomplish things, which will probably involve messy political stuff.
I have specific concerns about Sarah's post, which I'll post in a comment when I have a bit more time.
The post does raise some valid concerns, though I don't agree with a lot of the framing. I don't think of it in terms of lying. I do, however, see that the existing incentive structure is significantly at odds with epistemic virtue and truth-seeking. It's remarkable that many EA orgs have held themselves to reasonably high standards despite not having strong incentives to do so.
In brief:
The incentive structure of the majority of EA-affiliated orgs has centered around growth metrics related to number of people (new pledge signups, number of donors, number of members), and money moved (both for charity evaluators and for movement-building orgs). These are the headline numbers they highlight in their self-evaluations and reports, and these are the numbers that people giving elevator pitches about the orgs use ("GiveWell moved more than $100 million in 2015" or "GWWC has (some number of hundreds of millions) in pledged money"). Some orgs have slightly different metrics, but still essentially ones that rely on changing the minds of large numbers of people: 80,000 Hours counts Impact-Adjusted Significant Plan Changes, and many animal welfare orgs count numbers of converts to veganism (or recruits to animal rights activism) through leafleting.
These incentives don't directly align with improved epistemic virtue! In many cases, they are close to orthogonal. In some cases, they are correlated but not as much as you might think (or hope!).
I believe the incentive alignment is strongest in cases where you are talking about moving moderate to large sums of money per donor in the present, for a reasonable number of donors (e.g., a few dozen donors giving hundreds of thousands of dollars). Donors who are donating those large sums of money are selected for being less naive (just by virtue of having made that much money) and the scale of donation makes it worth their while to demand high standards. I think this is related to GiveWell having relatively high epistemic standards (though causality is hard to judge).
With that said, the organizations I am aware of in the EA community hold themselves to much higher standards than (as far I can make out) their donor and supporter base seems to demand of them. My guess is that GiveWell could have been a LOT more sloppy with their reviews and still moved pretty similar amounts of money as long as they produced reviews that pattern-matched a well-researched review. (I've personally found their review quality improved very little from 2014 to 2015 and much more from 2015 to 2016; and yet I expect that the money moved jump from 2015 to 2016 will be less, or possibly even negative). I believe (with weaker confidence) that similar stuff is true for Animal Charity Evaluators in both directions (significantly increasing or decreasing review quality won't affect donations that much). And also for Giving What We Can: the amount of pledged money doesn't correlate that well with the quality or state of their in-house research.
The story I want to believe, and that I think others also want to believe, is some version of a just-world story: in the long run epistemic virtue ~ success. Something like "Sure, in the short run, taking epistemic shortcuts and bending the truth leads to more growth, but in the long run it comes back to bite you." I think there's some truth to this story: epistemic virtue and long-run growth metrics probably correlate better than epistemic virtue and short-run growth metrics. But the correlation is still far from perfect.
My best guess is that unless we can get a better handle on epistemic virtue and quantify quality in some meaningful way, the incentive structure problem will remain.
I suspect that a crux of the issue about the relative importance of growth vs. epistemic virtue is whether you expect most of the value of the EA community comes from novel insights and research that it does, or through moving money to the things that are already known about.
In the early days of EA I think that GiveWell's quality was a major factor in getting people to donate, but I think that the EA movement is large enough now that growth isn't necessarily related to rigor -- the largest charities (like Salvation Army or YMCA) don't seem to be particular... (read more)