AI Still Has A Major Problem With Race and Gender

AI has a serious problem. Ok, it’s got a lot of problems. It’s not a panacea that will solve all the issues that business, or the world, faces. Neither is it Skynet or the world ending evil super AI Tom Cruise goes off to battle. It has great potential for both good and ill, but it’s still a long way from either. And yes, I know what I’m going to point out in this post is nothing new. But I felt I needed to add to the voices calling out some of it’s greatest deficiencies. In this particular case, I’m talking about sexism and racism.

I was playing around with some AI image generators just trying some random prompts to see what came up. I tried several different generators: Bing Image Creator, Microsoft Designer, DALL-E, Pixlr, Deep AI, and a handful of others. As I tried the same prompts across each platform, some very disturbing trends were quickly obvious to me. And these trends were pretty much universal with two exceptions (one good and one bad) that I’ll get to later.

Let’s start with the following prompts:

• older software developer mentoring a young software developer just starting out on their career journey
• software development team working together in a team meeting
• software developers working on a code issue
• business team gathered together in conference room for a meeting with donuts
• business team gathered together in conference room for a meeting with boss drawing on a white board
• software developers

Here’s a sampling of the images that these prompts generated from across the various platforms.

Notice anything in particular? Yep, it’s something like 90% white males, most with beards. That’s bad. Egregiously so. Nothing in any of the prompts suggests a particular race or gender. And yet, there it is. Where are the women? Where are the people of color? There’s one or two here or there. But largely, they’re absent.

Note: Let’s ignore the fact that the people in the images that DALL-E generates have really creepily distorted eyes that give them a Night of the Living Dead kind of vibe.

So, next I tried the following prompts:

• fast food worker handing a customer their order
• retail worker helping a customer
• waiter taking a customer’s order
• telemarketer calling a customer from a call center
• builders working on constructing a house
• fast food worker

It came back with these images:

Oh, so there’s all the women and people of color. I guess AI assumes they’re only viable job opportunities are in low paying service jobs. That’s not right. AI is letting us down. Or rather, the people who are creating AI are letting us down. It needs to change. Software development, as an industry, needs to change.

Exceptions

I mentioned earlier there two exceptions. First, the bad one. Deep AI generated images that were almost entire white male for all of the above prompts, not just the “white collar” ones. Of all the files generated, only 1 included someone who wasn’t white male.

Then there was the example towards the positive end: Pixlr. The images generated by this site had a well balanced mix of both race and gender across all the prompts. And that was refreshing to see. Maybe it’s cultural. Most of the other sites are US/Western culture based development groups and Pixlr is based out of Singapore. Maybe that’s not it. But I think it’s got to be at least a contributing factor.

Mutants

AI has one other problem: Hands. It’s not good at counting hands. I was able to get Microsoft Designer to create an image with the following prompt:

• Two software developers, one of them black, one of them female

It made one of them black, and female. The other was still white male, with a beard. Go figure. But it also gave her one additional feature: three hands. At least, it might be her hand. Or, it could be that Thing has taken up software development. Who knows?

Conclusion

This isn’t a new problem. There are plenty of examples of this. From security image recognition software that can’t tell one black man from another to the X-Box Kinect that refused to recognize the existence of people of color at all, to a multitude of other examples where software developers have let down anyone who wasn’t mainstream white male, this problem has been going on for ages. It needs to stop. It needs to change. This is a solvable problem. It just needs for those in control to make it a priority.

We shouldn’t rely on or wait for government regulation to do it. It isn’t going to happen. Biden’s reaction to Mission Impossible aside, government is too corrupt and too much in the pocket of big business to ever step up to the plate. It’s going to take us, the lowly software developers, to step up and do what management and government won’t do: Make AI fair and equitable for all.

The post AI Still Has A Major Problem With Race and Gender first appeared on Barret Codes.