I tracked discussion, leaked from inside Facebook, of a single algorithm weight to show how the company ignored warnings about how it would make the site angrier – and then backtracked.

The “Facebook Files” included hundreds of screenshotted internal discussions about the guts of Facebook’s News Feed. I collated the scattered discussions of “reaction weights” – how much the algorithm used likes, loves, “anger” and other emoji reactions in choosing how broadly to show a post. The anger emoji was initially weighted highly. Employees had warned this would incentivize ragebait content… and it did. My investigation showed how Facebook ignored warnings that it was making a mistake, before gradually dialing down the angry emoji’s weight to zero (exactly, as those employees had recommended beforehand).