In 2016, the startup app, Blinq, made waves with its artificial intelligence attractiveness detector. Users of the new site could get a rating on a six-point scale from “hmm” to “godlike” in terms of how hot they were.

The company created the website (no longer active) and allowed users to upload photos. The Blinq AI would then run the photo through its algorithm, guess the age of the person, and give a hotness rating. While there were issues with the AI, the website became very popular and garnered massive media attention.

Blinq’s app was a dating app similar to Tinder. Users could vote yes or no on a series of faces uploaded to the app. Using location beacons, Blinq was able to match users based on proximity and shared intrigue. A Swiss app, most of Blinq’s users were Swiss, and before it was acquired, the app had a relatively small user base.

Using neural networks, Blinq hoped to differentiate itself from the competition (Tinder, OkCupid) by learning user preferences over time, eventually serving up only good matches to its users.

The System Behind Blinq AI’s Hotness Calculator

The Blinq app and have since been acquired, so the resources are no longer available online for you to try for yourself. However, the basic concepts behind the Blinq AI were fairly straightforward. Developed in partnership with University ETH Zurich, the AI relied on three main stages of calculation in order to determine hot or not.

First, upon uploading the photo, the AI would identify what part of the photo contained a face. Next, the AI broke down the face into component parts, measuring the proportions of the face as it did. Finally, the AI made a prediction about age and hotness based on what it had learned from previous entries. While users on the internet had fun feeding the AI strange pictures and it sometimes made wildly incorrect guesses, the Blinq AI is still a useful lesson in the challenges of AI, and how it works.

Face Detection vs. Face Identification

An important distinction to make about Blinq’s AI is between face detection and face identification. Face identification is a machine-learning problem we haven’t fully solved yet. It involves supplying an AI with a photo and then asking it to identify the person based on a database of past photos. When Facebook tries to suggest friends you should tag in a photo, that’s facial identification. It turns out that AI is still not that good at doing this kind of comparison. Blinq’s AI did not attempt to identify faces or match them with pictures of the same person.

Face detection, on the other hand, is a well-established branch of AI, and one that Blinq’s AI used. It involves finding a face within a picture. AIs accomplish this through a series of tests to recognize facial features like eyes, noses, mouths, and ears. The next time you use a smartphone to take a picture of someone, you’ll notice that most models automatically focus the camera on the faces it finds. This is face detection at work. Blinq’s AI detected the faces in photos then cropped out the other features surrounding the face in order to run its analysis.

Teaching the Blinq AI What Hotness Is

Before Blinq could give hotness ratings, it had to learn what hot means. This is harder than you might imagine. While humans learn to recognize faces and beauty naturally, the Blinq AI had to learn about beauty mathematically. It received more than 100,000 images in its database to study the proportions of the users’ faces.

The developers then supplied more than 20 million ratings of those images to the algorithms. Using those 20 million ratings, the AI actually learned what proportions and facial features make a face hot.

Human Bias Becomes AI Bias

The difficulty with such a system, and why it had so many problems when launched to the world, is 100,000 data points is not very much. While it may seem like a large sample, it’s not a representative sample. Remember that the Blinq app (where the data came from) is based in Switzerland, and the user ratings are Swiss ratings. As such, the AI did a poor job of judging attractiveness in other parts of the world.

One post on Hacker News found that Denzel Washington, a black actor known for his good looks, was only found to be mildly attractive, and the issue persisted with other faces from different races. This is an important reminder that AI is only as good as the data you give it. If your data is biased, the AI will be, too.

What Happened to Blinq?

In the end, the founders of Blinq put the company up for sale in October 2016. A few months later, Blinq and were acquired by APG|SGA, an outdoor advertising company interested in Blinq’s model for beacon networks and location-based marketing. Blinq never fully developed the AI, and it’s not likely any other dating apps will, so we won’t know the full potential of AI matchmaking any time soon.

Image Source: Adobe Stock