31
Google pauses Gemini’s ability to generate AI images of people after diversity errors
(www.theverge.com)
Battlefield V Moment
this post was submitted on 22 Feb 2024
31 points (97.0% liked)
Technology
1934 readers
242 users here now
Which posts fit here?
Anything that is at least tangentially connected to the technology, social media platforms, informational technologies and tech policy.
Rules
1. English only
Title and associated content has to be in English.
2. Use original link
Post URL should be the original link to the article (even if paywalled) and archived copies left in the body. It allows avoiding duplicate posts when cross-posting.
3. Respectful communication
All communication has to be respectful of differing opinions, viewpoints, and experiences.
4. Inclusivity
Everyone is welcome here regardless of age, body size, visible or invisible disability, ethnicity, sex characteristics, gender identity and expression, education, socio-economic status, nationality, personal appearance, race, caste, color, religion, or sexual identity and orientation.
5. Ad hominem attacks
Any kind of personal attacks are expressly forbidden. If you can't argue your position without attacking a person's character, you already lost the argument.
6. Off-topic tangents
Stay on topic. Keep it relevant.
7. Instance rules may apply
If something is not covered by community rules, but are against lemmy.zip instance rules, they will be enforced.
Companion communities
[email protected]
[email protected]
Icon attribution | Banner attribution
If someone is interested in moderating this community, message @[email protected].
founded 1 year ago
MODERATORS
What did people expect, that a text-to-image model would be able to understand centuries of human stupidity and the politics it have created?
I am not tech savvy but I expect that there is probably a large set of images that the model was trained on. Based on my experience, almost all images that could be labeled “German soldier, 1943” would show people with similar physical characteristics to the first of the four images. I guess my thesis is that I’m not sure that the historical context is necessary when the source/training data is likely fairly homogeneous. To get outputs such as they show in the example suggests that either the training data was not reliable or someone behind the scenes has their thumb on the scale to push the model toward racial/gender diversity even when that doesn’t match the input.
The context is that earlier image generating AI took some flak for generating images that didn't have much diversity, even if that was true of the training data. For instance, you might have asked one to generate images of a doctor treating a child in a doctor's office, and it would generate almost all male doctors, maybe even white male.
So in response, some of them are programmed to generate a range of diverse subjects when creating images of people. Lots of racial and gender variation. But they didn't stop to think that some historical groups didn't have much diversity, and it's a mistake to artificially create it, as in the example of Nazi soldiers.
So yes, a "thumb on the scale" of sorts.
someone behind the scenes has their thumb on the scale to push the model toward racial/gender diversity
I thought it obviously was the case here.
And it is understandable, these models work by putting the most likely pixel after another until it ends with a picture, and the average picture is filled with the average stereotype (bad or good). And pushing it away from that will of course affect everything in the model. The computer cant handle every exception we need to put into it for it to be clean and sellable. Theres gonna be many more ai generated pictures like this but shocking in a different way, in the future. Probably even worse.
I’m just floored that evidently nobody involved appears to have had the barest inkling that this sort of weird and inappropriate cross-pollination would have a good chance of occurring.
OK but that just makes it clear you don't understand how generative AI works?
I understand how ML and neural nets work. I’m saying that outputs of LLMs are rather stochastic in nature, and cannot realistically be fully controlled in a deterministic fashion if you want to make it pretend to be an AGI.
Battlefield V Moment