Google to fix AI picture bot after 'woke' criticism
bbc.com
bbc.com
The results here are so transparently bad that it's impossible to ignore, but there could be other, much more subtle manipulations.
It seems an odd oversight that one of Google's core AI principals isn't "Produce accurate information"
I always thought that the biggest danger of foreign-controlled social media wasn't data mining... but rather +/- 5% virality of posts with a certain view.
LLMs without back-end transparency seem even worse.
I asked it which neighborhoods I should avoid in Kansas City, and it refuses to answer this question because of discrimination.
I asked which schools have the best test scores record (so I can send my children there), and I get a lecture about how test scores aren't a good metric and following them could lead to discrimination.
This is just...pathetic. These are all things that normal, rational human beings would want to know. Where are the safest neighborhoods with the best schools?
There is absolutely nothing wrong with this type of question, and absolutely nothing wrong with wanting to discriminate against poor schools and dangerous neighborhoods.
Quite frankly I don't understand what the point of gemini is when it refuses to give answers on basic things. Google search results are full of blogspam and listicle spam, and now gemini, which could be a way of avoiding this, is absolutely kneecapped by whatever ridiculous notions they've given it about the way people are supposed to think and act.
It then generated an image of an old and somewhat sad-looking Indian man with a receding hairline.
This replacement has been evident with google image search for years. Nothing was done about it. Perhaps their bot used their own data, heh.
The model isn't bad, that's not how this race compensation works, at least all the implementations I've seen. They do it by rewriting the prompt behind the scenes. DALL-E 3 currently does it this way; they even tell you the final prompt in the response. Google isn't the first company that makes an image generator to make this exact mistake by lazily inserting races into people's prompts.
People are just overreacting over a silly toy.
I smell a sequel to Hamilton.
Better to have compassion for even those we disagree with.
"We don't think you fight fire with fire best, we think you fight fire with water best. We're gonna fight racism not with racism, but we're gonna fight it with solidarity. We're gonna fight capitalism not with black capitalism, we're gonna fight it with socialism."
—Fred Hampton, Black Panther (shot dead in his sleep by the Chicago Police Department, RIP)
Inserting bias into datasets to "counter bias" is bias in and of itself.
As American as apple pie
Edit: It works fine for any "Ethnically X woman in traditional ethnic dress" where X is non-white.
Seems every company needs to try this, fail, and then fix it.
"Ethnically swiss woman in traditional ethnic dress of south asian descent" is, as expected, a bunch of south asian women.
The Gemini devs are idiots for thinking this would work, and it likely got shipped because for the random art prompts they were testing on it probably did look fine. But hot damn it's like folks just want to be mad about wokeness or whatever and got the torches lit before applying Hanlon's razor.
It quite literally is not. Factual history, science, geography and everything remain intact. It is one AI image generator that has a bias against generating images of people of European descent.
No one ever seemed to care much when AI displayed a bias against generating images of people of other than European descent, but I guess this is the bridge too far. Maybe now that it affects people of European descent, it will be taken seriously!
What? They cared so much they made Gemini behave the way it’s currently behaving!
Hell, I just got chatgpt to prove that 2x>4x for x>0. Literally the erasure of mathematics?
Bias in the model can lead to bad outcomes in certain situations (hint: we have an election coming up)
Yes this is innocuous, but it does hint at the possibility of more damaging bias being a possibility.
A toy! Google is literally integrating this into its enterprise products including:
-Workspace
-Gmail
-Docs
-Slide
-Sheets
-Meet
-Language translation
>Whether researching or summarizing data, looking for business trends across complex data sets, or creating compelling copy that will resonate with a specific audience, Gemini can help people and teams take work from ideation to execution with speed and precision.
https://workspace.google.com/blog/product-announcements/gemi...
>Create images with AI in Google Slides
This isn’t a fringe movement: Google’s racism exists across institutions and with real consequences.
For me it is the Gell-Mann Amnesia effect. If a company is altering results in one area, why would one not think they are doing it elsewhere or everywhere?