Member-only story
Google’s A.I.-Powered Camera Was Destined to Fail
A program just can’t tell what’s ‘interesting’
Google just pulled the plug on its Clips camera, a tiny device that was supposed to capture photos of the little snippets of your day that would otherwise slip by.
Maybe it would snap a picture of your baby smiling as you went to pick them up, or your dog wagging its tail as you walked in the door. But when reviewers and consumers got their hands on the device, that fairy tale didn’t pan out.
“I wore the camera on the collar of my shirt while playing with my daughter hoping that it was capturing all her smiles at 7 weeks old. Logged into the app, nothing. I was crushed,” wrote one reviewer on Best Buy’s website.
Google promised to use A.I. to capture interesting moments in our lives, and that was the exact problem: There’s no one definition of interesting.
As a user, you had to spend a lot of time showing Google what you wanted it to take pictures of. Engadget’s Cherlynn Low wrote last year that she trained the Clips’ A.I. on her Google Photos history, ostensibly a diary of her visual preferences, and still couldn’t get it to reliably photograph her dog or friends.
We know what A.I. is good at doing: Making a decision with a clearly correct answer. If there’s a dog in an image, Google can identify it. The dog is a fact. But for Google to figure out when your dog is doing something you want to see again, like the guilty face it makes after digging through the garbage, it needs a ton of outside information that won’t be the same for every user.
In other words, A.I. doesn’t work when it has to quantify something qualitative. That spans from whether a person is trustworthy to whether a photo is good or bad.
“Basically, we’re trying to teach English to a two-year-old by reading Shakespeare instead of ‘Go, Dog. Go!’”
Google originally hoped to rectify this issue by training the camera on images from professional photographers so images taken by the camera would resemble the framing and composition of the pros, but according to a 2018 Google blog post, that proved to be too ambiguous of a goal.
