Microsoft’s new AI bot is way less racist, still pretty confused
Microsoft’s slow, stumbling climb toward The Rise Of The Machines continues today, as another one of the tech giant’s vaunted AI bots catches the mocking ire of the social-media hordes. On the plus side, the company’s latest experiment, CaptionBot, has yet to be locked away in online juvie after becoming a drug-addicted virtual teen racist. On the other hand, it does seem to think Michelle Obama is a cell phone, which suggests a certain technocentric bias of its own:
The bot is the latest in a long line of image recognition programs the company has tested on the public, including self-esteem damager How Old Do I Look? and the vital dog recognition application What Dog? This time, though, Microsoft is going for the holy grail of image recognition, tasking the bot with verbally identifying the contents of a pic. And even though CaptionBot is pretty upfront about its own work-in-progress nature, people on Twitter have been quick to pass around some of its more embarrassing failures.
Of course, it wouldn’t be fair to CaptionBot to make fun of its capabilities without testing them out on our own, so we also dipped into The A.V. Club Newswire’s extensive archive of header pictures for some experimenting of our own. Here are the results:
On the plus side, the program’s almost as good at celebrity recognition as it is at imagining fictitious cellphones, or overestimating the happiness in people’s expressions.
It’s also capable of recognizing Resting Kanye Face. But naked William H. Macy leaves it as speechless as the rest of us:
Ultimately, though, it comes through when it really counts:
That’ll do, CaptionBot. That’ll do.