Inside the deranged mind of Twitter's 'For You' algorithm
What you learn when your X account is hacked.
Update: X was able to restore my account by associating it with a different email address. I appreciate their help. I can’t promise that my tweets won’t be stupid and annoying, but at least they’ll be coming from me.
On Saturday, right as I was finishing up a fantasy basketball auction1, I got an email saying the address associated with my Twitter account had been changed. For what it’s worth, this wasn’t coming completely out of nowhere — there had been some weird attempts to access my account from England and Russia earlier in the week, and I had two-factor authentication on and had recently changed my password. Shortly thereafter, a spammy post was issued from my account advertising some sort of sale on a MacBook, and then the account was locked. It could have been worse, I guess — at least whoever stole the account didn’t post porn or a blockchain address.
I’m posting this in part as a warning — please ignore anything coming from the @NateSilver538 Twitter account until I’ve given the all-clear, because it isn’t coming from me.
So far, I haven’t been able to regain access. Twitter — officially X — has been completely unhelpful. There are automated forms one can fill out — but a first attempt to regain access was rejected even after I sent in additional documentation proving my identity. Meanwhile, a second one got literally stuck in the mail:
Mail Delivery Subsystem <xxxxx>
1:13 PM (1 hour ago)
to me
**********************************************
** THIS IS A WARNING MESSAGE ONLY **
** YOU DO NOT NEED TO RESEND YOUR MESSAGE **
**********************************************
The original message was received at Sun, 22 Oct 2023 05:45:39 -0700
from xxxx [x.x.x.x]
----- Transcript of session follows -----
... while talking to aspmx.l.google.com.:
>>> DATA
<<< 450-4.2.1 The user you are trying to contact is receiving mail at a rate that
<<< 450-4.2.1 prevents additional messages from being delivered. Please resend your
<<< 450-4.2.1 message at a later time. If the user is able to receive mail at that
<<< 450-4.2.1 time, your message will be delivered.
I’m not sure whether at any point, any request has been evaluated by any actual human being. Some parts of Twitter have held together better than people predicted after Elon Musk’s major layoffs, but the customer service team apparently isn’t one of those functions.2 A text message request for help to a Twitter insider also went ignored. I’m sure I could get through to them if I caused enough of a ruckus. But you’d think at some point that someone would reach out to a 3.3-million follower account that has obviously been hacked — a Twitter Blue account3 that is followed by Elon Musk — even if it was to say “due to our security protocols, this might take a couple of days to sort out”. It’s a weekend, I guess.
There has been one blessing in disguise, though: I’ve learned a lot about how the Twitter ‘For You’ algorithm works.
I should mention that I actually kind of like the ‘For You’ tab that Twitter serves up to @NateSilver538 and sometimes use it instead of my chronological feed. It knows that I really like the NBA and poker, for instance, more than you’d gather from how often I tweet about those subjects. Sometimes ‘For You’ can even be smart in uncanny ways. It tends to know, for instance, who I find to be an enjoyable hate-read versus who I just find annoying.
But the algorithm has a lot of data on me and I have some pretty consistent preferences. If you’re just starting out though? Hoo boy. ‘For You’ nudges you into some very weird corners.
In attempting to regain access to my account, I inadvertently created a new X account. The account follows Elon Musk and only Elon Musk. It has never tweeted, replied, retweeted or liked a tweet. All the algorithm knows is my name, email address and that I follow Elon.4 So what do I get from ‘For You’?
There is very little pure text (~10% of all posts). A large majority of posts (~70%) are videos. The remaining ~20% contain still images. There’s a lot of, uh, this:
There are even fewer links to external content — maybe 2-5% of posts. This is in line with what Elon has said publicly. The Twitter algorithm really discourages links to outside sources.
I haven’t seen a single ad. Not unless the very frequent references to the new Playstation game Spider-Man 2 is a viral marketing campaign.
The algorithm either quickly imputed that I’m a guy, or it has intrinsically very bro-y preferences. It’s very Barstool Sports. Lots of video game content, lots of sports, actually not too much smut though some stuff that’s smut-adjacent. You’re very discernibly trapped in the brain of a 26-year-old heterosexual male.
‘For You’ quickly goes down rabbit holes. The algorithm is convinced that I’m absolutely fascinated by a seemingly blown call in the Iowa-Minnesota football game last night. Literally, something like 20 percent of my feed is taken up by this one play:
I honestly don’t think I’ve done much to encourage the algo — maybe I lingered to watch the video of the play the first time I saw it? This is not a particularly huge sports development, by the way; neither Iowa nor Minnesota are ranked. This looks like an example of an overfit algorithm.
There is not that much politics, but most of the political content is low-quality and strongly right-wing, often with a conspiratorial angle. These Tweets are occasionally interspersed with low-quality partisan left-wing content, like “dunk” tweets from Democratic members of Congress. But around 80 percent of the political content, which is maybe 10 percent of the feed overall, has a right-wing bent. Sometimes, you’ll be scrolling through endless Iowa-Minnesota and Spider-Man 2 content and then boom, there’s a vax-skeptical post:
I grabbed this next example because it’s a rare instance of a long text-only post. Since the algorithm strongly punishes text, it must really like the content of this one to make up for it. And what is that content? It’s literally a laundry list of right-wing conspiracy tropes:
Finally, there’s some content that’s conspiratorial in ways that are just plain … weird? I’m not sure exactly what’s going on with Bill Gates and McDonald’s french fries. That Spider-Man game sure does look fun, though!
There’s very little hard news. Meta’s Threads has been criticized for de-emphasizing news and politics content. But honestly X’s ‘For You’ tab is kind of the same way. Would you know from scrolling ‘For You’ that there’s a war going on between Israel and Hamas? I think so, but only in sort of a meme-ified way — I saw one post, for instance, that showed some sort of bodega with the produce arranged in the shape of a Palestine flag. You’d have even less awareness of something like the Republican speakership crisis — and what you would see about it is dunks and memes, not analysis. There’s very little content of any kind from any professional news organization.
I don’t know how ‘For You’ would be different if I’d followed no one instead of Elon, or if I’d signed up from an email account with a different name. It does have some information to work off, I guess. But man, it really doesn’t take much to send it off into conspiracy land, and I imagine that these preferences still exert an influence even as X does learn more about you. Even if this is just what new users are being exposed to, it could make a pretty big difference because of path-dependence.
I also can’t tell you much about whether the algorithm is optimizing for engagement, versus optimizing for Whatever Elon Wants. And I can’t tell you whether it’s doing a good job of optimizing or a poor one. In general, my bias is to think it’s pretty hard to design good algorithms — and that a lot of dubious algorithmically-tailored content5 reflects attempts to exploit badly-designed algorithms rather than cater to authentic consumer preferences. But ideally I’d like to talk with someone at X about this, and right now I can’t even get them to restore my hacked account.
If you do know how to help out, drop me a line. And more importantly, subscribe to this newsletter, even if it’s just the free version. I thought I’d have a little bit more time to ween people off of relying on Twitter for reading this newsletter — but at this point who the hell knows.
I’m going to have a great team so long as Joel Embiid, Zion Williamson, Kristaps Porzingis, Cade Cunningham, Ja Morant and Paul George all stay healthy. In other words, I’m screwed.
Since I’m no longer associated with FiveThirtyEight, I’ve also been trying for months to change my user handle from @NateSilver538 to something else, but to no avail.
Although not one that I was paying for. Some months back, Twitter gifted Blue status to everyone with 1m+ followers.
And I guess it knows that I’m located in the US and use — no apologies! — an Android phone. It’s also now seen about a half an hour of scrolling behavior and what I tend to hover over and what I don’t — but I grabbed most of the examples below as soon as I could before it began amplifying those preferences.
I've just never had any reason to switch. Google integration is useful in various ways, really like the Galaxy series cameras. Had an iPhone for work as a second phone at a couple of points and didn't quite fall in love. I *am* a PC to MacBook convert FWIW (the Fake Nate tweeting from my hacked account was right about that lol). But also don't mind diversifying technology providers a bit.
They almost got me with the 'free fries.' Thanks, Dom.
Seriously though, there ought to be a law or something where if your company is of a certain size or importance, then a customer has to be able to talk to a human being at some point. And when did we opt in to all the voice command bs when you're on the line? I much preferred pressing a keypad to barking into the phone like an imbecile. "Urinary Incontinence!" "Urinary Incontinence!" I'm sorry, could you repeat that?