Google DeepMind's guardrails "were clearly wrong” after Gemini showed historically inaccurate depictions of people of color, VP says

Transcript

00:00you guys created a model that it was actually very hard

00:02to get it to generate images of white people,

00:05even in historically appropriate context for that.

00:07And it was quite embarrassing for you.

00:09How do you overcome issues like that?

00:11Because I think that's one of the problems now

00:14is that people want a model that you can just say,

00:16okay, well, be diverse when that's appropriate.

00:18And when it's not appropriate,

00:21be correct to the historical context,

00:23but you can't just tell the model that apparently.

00:25So how do you overcome this problem?

00:27First of all, thanks, Jeremy, for asking.

00:29It really highlights the fact

00:31that these challenges are difficult, right?

00:34First of all, we want the tools,

00:37the generative AI tools to be relevant to people

00:41all around the world.

00:42So if you ask a model to create images of, let's say,

00:47a nurse in a hospital or a school teacher

00:51or a board member of a company,

00:53it has to be relevant to people all around the world.

00:56Google is a global company.

00:57We have people using our tools on every continent.

01:02But the guardrails that we put in in that case

01:06were clearly wrong.

01:07We came out and we actually said that.

01:10We then pulled the image generation for Gemini off

01:15so that we can improve it and put it back out

01:17in a way that's more acceptable to more people.

01:20These are cutting-edge tools.

01:22You put them out there.

01:23People find them really useful sometimes.

01:26Sometimes they find them offensive, right?

01:28And I think this is a healthy thing

01:32that the whole AI community

01:33is really grappling with right now.