Suppose you see someone flip a coin five times, and get five heads. What’s the chance that the next time they flip the coin, it’s also heads?
Naive probability would say one half, by assuming it’s a fair coin. In the real world, it’s a little more than a half: it might be an unfair coin, a party trick, or something like that.
But there’s another question to consider, even if the coin is totally fair and no tricks are pulled: Why are you watching this coin be flipped? If you’re watching the coin be flipped live in front of you, it’s no big deal. But if you’re watching a video you found on the internet, you’re way more likely to see the next coin come up heads.
This is postselection: The reason you’re watching this video is probably because there’s a long streak of heads flipped in this video. This isn’t just any video, it’s a video of someone flipping a lot of heads. This isn’t just any coin, it’s a coin that’s going to flip heads.
The degree of postselection, the amount of bias towards something impressive happening, is based on how filtered your viewing environment is. The real world, and live events, are essentially unfiltered. If the coin flipping video is made by a friend or someone you know, it’s probably only moderately filtered: it would only take a somewhat impressive occurrence to bring the video to your attention.
But if the coin flipping video is just out there on the general internet, a stupendous amount of filtering has happened to bring this video to your attention. Out of millions of people flipping millions of coins, you’re watching this one. As a result, one-in-a-million coincidences, like flipping 20 heads in a row, are commonplace.
Implications of Postselection
Now, instead of coin flipping, you see a very offensive tweet that someone sent. You want to figure out if the tweet is reflective of the sender’s inherent nature (i.e. they’re an awful person) or not reflective (i.e. a failed joke, a typo, out of context, etc.).
This question closely parallels the coin flipping question. If we see a coin flipped heads twenty times, do we lean towards a plugged coin (inherent nature) or random chance (not inherent nature).
To answer this question, we must examine the level of postselection that has occurred in bringing the offensive tweet to our attention. If the tweet has gone viral and spread to millions of people, an extreme degree of postselection has occured, and every aspect of the tweet is likely to be perfect for the spread.
A good example of this is a tweet sent by Justine Sacco. To briefly summarize, Sacco wrote an incredibly offensive tweet in a failed attempt at a joke. Compounding factors include:
She worked as a communications director, and bringing down someone of her high status was as a motivating reason for one of the earlier retweeters.
She sent the tweet just before getting on an 11 hour flight, so she could not take down the post before it spread everywhere.
In this case, an incredible degree of postselection occurred. Her tweet was seen by a large fraction of the people on Twitter. It was rebroadcast through the wider media system, reaching people like me who don’t even use Twitter. And Sacco originally only had 170 followers. This represents at least a million-to-one postselection factor, and probably significantly more.
As a result, we shouldn’t be surprised that a bunch of rare events happened in a row. Is it rare for someone to misjudge their joke incredibly badly? Yes. Is it rare to specifically do so in a highly offsensive manner? Yes. Is it rare for that misjudged joke to be made on a public, easily spreadable platform? Yes. Is it rare for the person who did so to also have a high-profile, communications-based profession? Yes. Is it rare for the person who did so to be unable to communicate for a long time afterwards? Yes. Each of these rare events might be a 10-to-1 or 100-to-1 rare event. And yet, the postselection power is sufficient to overcome this, and bring this case to wide attention. We should therefore not be surprised if this case was truly the result of many rare events stacking up, and not reflective of the inherent nature of the person in question.
Now, I will mention an alternative scenario: Where little postselection has occurred, so we can draw conclusions more immediately about inherent nature.
For an example of an extremely offensive comment with little postselection, take the case of Travis Woo. While he has made many offensive comments, the case I’m thinking of (and which I linked) involved Woo conducting an hour-long stream discussing the merits of Adolf Hitler and Mein Kampf. I heard about this because Woo was a content creator for a Magic: the Gathering website, Channel Fireball, which I followed regularly. There were perhaps twenty regular posters on that site, and perhaps twenty sites which I followed as regularly as Channel Fireball, across all topics. As a result, very little postselection had occurred, roughly a factor of 400-to-1, to bring the stream to my attention. I therefore concluded that this stream was likely reflective of Woo’s inherent nature, not an artifact of postselection.
Further evidence has supported the “inherent nature” conclusion, including Woo’s further negative consequences for organizing a Facebook group that was home to a variety of misogynist and racist content, Magic for Bad.
An interesting angle on these two stories: Both Sacco and Woo were immediately fired from their jobs by their employers for their offensive comments. I thing this was a reasonable decision, because for their employers, little postselection occurred. However, if the typical person who heard about each story used the framework I’ve described here, they would have extended Sacco the benefit of the doubt, and not extended Woo that benefit. In both cases, so far as I can tell from later evidence, that decision would have been correct.
To navigate the modern world, and draw accurate inferences about emotionally laden topics, we must keep in mind the mechanism that brought some piece of information to our attention. Has no postselection occurred? Mild postselection? Heavy? Extreme? These distinctions make a major difference in how dramatic a conclusion we should draw about based on that evidence.