Thread by @cfiesler

string.replace("sexual", "political")

Not that I was *surprised* to see this study about predicting "political orientation," but since I've been talking about the "gaydar" (sigh) algorithm from the same researcher for a while now, here's some reflection.
nature.com/articles/s4159…

Given criticism of the previous paper (which if you're not familiar is here: psyarxiv.com/hv28a/ ) I was genuinely expecting to see an ethical considerations section by the end of this paper (since that criticism pretty much constructed it exactly!). There is not one.

There is a lengthy "author notes" document linked to from the article that includes FAQs (like "physiognomy????") and twice warns to not "shoot the messenger" so I guess that's the ethics statement.

To be fair to the author, here is a partial argument for why this study is not a problem. But even if this were sufficient, I am disappointed that @nature did not require this kind of reflection to be IN THE PAPER and not in an attached google doc.

One can argue about whether "let's see if we can" is a strong scientific motivation or contribution, especially if the science is very similar. This is not my field so I won't speculate. (But like, if someone does cat people vs dog people next do they get a paper in Nature?)

I also suppose one could characterize "it's important to know this is possible because threat to civil liberties!" as testing vulnerabilities in a system, kind of like cybersecurity research, but like... where's your patch?

And I kind of hate to point this out, but the 2013 paper about Facebook Likes cited above (pnas.org/content/110/15…) somehow did not prevent Cambridge Analytica. (I don't know if there's any actual evidence that it inspired it instead, but not outside the realm of possibility?)

And I'll let others who are more familiar with this type of research speak to this, but the findings have a ton of caveats, it doesn't perform THAT much better than chance anyway, and there was a replicability problem with the last one. arxiv.org/abs/1902.10739

and as @doctorow pointed out, there's a science communication issue with this. Because a lot of headlines are like OMG COMPUTERS KNOW THAT YOU'RE REPUBLICAN!1!1!! without explanation of those caveats.

https://twitter.com/doctorow/status/1350116084779675648?s=20

and because it's me, you might be wondering "and also where did all those faces come from?!?!" and interestingly, whereas for the sexual orientation study they "obtained" (scraped I assume) public dating profiles, for this one an unnamed dating site "provided" the data

and say what you want about TOS' relevance for research ethics (I have: cfiesler.medium.com/spiders-and-cr…), if I make a dating profile I am signing up to find love, not have my data USED TO TRAIN FACIAL RECOGNITION SYSTEMS

I'll also note that I legitimately do not like calling out very specific papers/researchers like this, but in this case, they 100% knew exactly what the criticism would be, so I assume they're prepared for it.

But also that's why I expected all this reflection in the paper. :(