To look hottest in selfies, you should have long-hair, be female and use a filter that decreases contrast, making the skin appear flawless and a little glowy. If you want to appear intelligent, on the other hand, you should avoid the colors pink, purple and red, as well as showing too much skin.

This is according to the results of two separate studies that looked at how computers judge photographs of humans. (H/T Adam Harvey.)

https://twitter.com/adamhrv/status/750612450794438656

https://twitter.com/adamhrv/status/750610992464293888

In the first study (which made the rounds on Twitter this week but is actually from last fall), researchers looked at two million selfies, using information such as how many likes and followers the user had to train a neural network to recognize what's a "good" selfie and what's a "bad" one.

A good selfie, the algorithm determined, basically looks like this:

This is who a computer program thinks is hot.

The factors that most attractive selfies had in common seemed to be that the selfie-taker was female with long, flowing locks. The composition mattered, too. In the good selfies, the face occupied about a third of the image, giving the viewer a close-up of the face. The images were also often over-saturated, with a filter that faded the image out and decreased the contrast. Basically, A.I.'s image of the perfect selfie looked an awful lot like Maxim's view of the perfect woman.

Advertisement

In the second study, published on Arvix last week, Cambridge researchers wanted to see whether humans or computers would be better at judging a person's intelligence based on a Facebook profile picture.

First the researchers had 1,122 Facebook users take an intelligent test. Then humans and a computer program looked at those users' profile photos and guessed how smart they were. The computer program's algorithm had been trained primarily on academic texts having to do with image aesthetics, image perception and image recognition literature.

The computer program tended to be better at more accurately guessing a person's actual intelligence based on the image—it was less likely to be fooled by factors that did not actually correlate with intelligence, like the presence of glasses or books or highly chromatic images. Here's how the computer decided whether someone was smart:

  • They avoided the colors pink, purple and red in their profile images.
  • They usually had fewer colors overall in their profile photo, and particularly liked the color green.
  • They showed less skin.
  • They tend to take clearer, less blurry photos.
  • They tended to have fewer people in their profile photo.

"Most intelligent people in our dataset understand that a profile picture is most effective with a single person," wrote the researchers.

Advertisement

The researchers conducted the experiment because of previous studies showing that job candidates with the most attractive Facebook profile images obtained 39 percent more job interview invitations. Knowing that humans are biased toward hotties, the researchers conclude it may be best to let an algorithm instead review candidates' Facebook profiles.

The studies show two uses of artificial intelligence with seemingly opposite results. In one, the biases of the humans liking all those selfies on social media were merely transferred to the A.I. algorithm when those human preferences were used to train it. The machine was not only a reflection, but an exaggeration of man—in this case, apparently a man with a penchant for long-haired blonde women. In the second study, though, the machine seemed better than humans at discerning intelligence from a photo, managing to avoid biases.

All this serves as a reminder of just how subjective "objectivity" can wind up being. Allow machines to be programmed in man's image and they will only exacerbate our own biases. But approached differently, they can be used to break them down.