Biases in person-assessment
Tyler Alterman, then a community-builder for the effective altruism movement, was working on tools to help EA organizations evaluate prospective hires and collaborators — in particular, a test for genuine domain expertise, and a way to correct for the bias that comes from people applying to multiple programs at once. Alyssa, drawing on her hiring experience at Google, MetaMed Research, and Humanity+, replies that explicit expertise tests are useful but down the priority list; the bigger problems are subtler social biases. The list of biases below is the bulk of her response, followed by notes on how to build absolute (rather than relative) effectiveness scales, and on whom in EA it’s actually worth recruiting heavily.
Detailed assessments of expertise are very useful elsewhere, but they’re fairly far down our current priority list. I have a fair amount of experience with person-assessment from my work at Google, MetaMed, and Humanity+; both the successes and the failures; both doing it myself and watching others. In my view, some bigger issues than making an explicit mistake about expertise are: Unfair biases towards people in Our Tribe: self-identified “rationalists”, “effective altruists”, Eliezer fans, CFAR alumni, and so on.
Unfair biases towards people who are Against The Establishment. This takes many forms, but eg. at MetaMed we saw (correctly) that famous doctors and hospitals often screwed up, and so concluded (incorrectly) that we should work with anyone who was “on the same side”. As you might expect, this group turned out to have a high proportion of con artists, quacks, incompetents, etc.
Unfair biases towards romantic partners. Most people know that, if they’re dating Bob, they shouldn’t decide whether to hire Bob. But they’ll often say nice things about Bob, promote Bob’s work, introduce Bob to their friends, and so on, which often results in their friends hiring Bob despite Bob being unqualified. (It also adds lots of extra drama when relationships become troubled.)
Disproportionate reliance on social proof. Discarding social proof entirely would be dumb, of course. But there are entire organizations that exist specifically to hack social proof, and they’re sometimes very effective. They’ll recommend each other at strategically chosen times, bribe people into recommending them, write sets of books which all recommend each other, write their own Wikipedia pages, write their own Amazon reviews, create their own fake media outlets which write stories about each other, etc., it’s all very well-planned.
Insufficient reliance on past behavior. The best predictor of future behavior is past behavior, for both good qualities and bad qualities. To quote Eliezer, “People who have donated before, donate again; people who plan to donate next year, will, the next year, be planning to donate next year.”
Overly broad interpretations of favored fields. Eg., very few people among the CFAR alumni are experts in science. If CFAR was using an expertise test, my guess is they’d likely conclude (correctly) that many alumni were experts in math, and then (incorrectly) assume this was easily transferable to science. In fact, the number of IMO winners who become good scientists is a rounding error, and even among IPhO winners, far more go into math than science. A related form of this is the stack fallacy .
Unfair biases towards skilled, prolific writers. To some extent this is justified, since writing is a valuable and widely-applicable skill. But eg. the fourth-largest donor to MIRI is a guy called Loren Merritt, and I think very few people realize this, since he never writes anything.
Time-value bias. Suppose you hold an event which delivers $200 in consumer surplus (maybe a bit less or more, depending on personal preferences). Who’s most likely to come? The people whose time is worth the least — they have nothing to lose by attending, rather than playing video games. People whose time is worth more are likely to not bother, or duck out early. People whose time is really valuable won’t come at all. So when you evaluate people you’ve met, you’re likely to choose someone ineffective — because they hang around a lot (mere-exposure effect), and because the most effective people aren’t around for long enough to make the comparison.
On multiple-application bias, the only way to really tackle that is developing guesstimated absolute effectiveness scales along various metrics, and then applying those to both the applicant pool and the acceptee pool. (“Absolute” means relative to the whole US or world population, not just a small sub-population.) This is a bit tricky, but it’s well worth doing. This often comes down to an economics-style “bag of tricks” about where it’s possible to get decent data. Some quick examples: Data on income is pretty well-known; at the low end from Census and IRS data, and at the high end from “rich lists” like the Forbes 400. (Important to adjust this for age, but that’s fairly straightforward.)
Data on intelligence is pretty well-known, through an array of standardized test scores. Lots of things correlate strongly with IQ; the most common one we use is SAT taken at age 12.
Data on credentials can be obtained from Wolfram Alpha, and various studies about education. Eg., if someone has a physics Ph.D. from Harvard, how common is that?
Startup valuations can be compared to data collected from CrunchBase, and from various VC firms. Anything about “number of companies” funded, advised, etc. should be ignored as meaningless. (It’s like a college saying “we have 600 people on our football team” — sure, but are any of them any good?)
Fame, for both people and companies, can be evaluated using a sort of “distance ladder”. At the very top, one can look at Wikipedia articles rank-ordered by traffic. In the middle, one can compare the traffic of an article to the traffic of top articles, and extrapolate the appropriate place on the power-law curve. At the bottom, one can count Google or Google News hits, and compare them to those of Wikipedia figures.
Of course, there are many key attributes of a group that can’t be easily measured, and it would be very dumb to ignore those. But the ones that can be measured are important, I think, to make sure that you don’t fool yourself.
On bugging people: with good reasons I’d bug everyone on the original list, with a few exceptions for people I didn’t know well enough and would get introductions to from Vassar. I think good reasons for bugging boil down to gains from trade: you give them something that makes their life better, and they also give you something that makes your life better. Some quick examples, using a stand-in I’ll call Bob: There’s a job that needs to be filled. Bob is a great fit, and is better-qualified than any of the current applicants. At the same time, the job is significantly better than Bob’s current job.
Bob’s company is looking for investors. You know a reliable, reputable investor, and Bob’s company is a better bet than anything in their current portfolio.
Bob is bored with his social life. You know a group of people who’d love Bob, and who Bob would consider more fun than his current friends.
So, there are two main problems to solve in order to bug everyone we’d ideally want to bug. The first is not having anything worthwhile to trade with (or, at least, nowhere near enough for all trading partners). This is essentially a liquidity problem. If Bob’s labor is worth $5 million a year, but he’s willing to sell it for $200K a year, that’s a fantastic deal… unless you don’t have the $200K, in which case it’s useless. (This can be either literal or metaphorical.)
The second problem is if we can’t make any use of what we receive from Bob, because there are other bottlenecks standing in the way. At this point, my guess is the biggest bottleneck is the lack of competent, well-run organizations, since such organizations are the generators by which labor and capital get converted into utility. Of organizations I’m well connected to, I’d say that MIRI, FHI, Wave, LightSail, Nanotronics, and OPP are competent and well-run, and maaaaybe 80K and EAO although those are still embryonic. I’d also count Ed Boyden’s group at MIT, AltspaceVR, Nomiku, Bellroy, and Alcor, although my ties to them are peripheral. After that, it tapers off pretty fast. (This isn’t a judgement about people — eg. I still think Max Tegmark is a friendly, hard-working, and very helpful genius — but about organizations as a whole.)