Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Google to pause Gemini image generation of people after issues

theverge.com

Post reply on HN
Print view View original

881–890 of 1001 posts

Re: Google to pause Gemini image generation of people after issues

#881

Post by feoren »

freilanzer wrote: post #684
feoren wrote: post #657

People are not understanding what Gemini is for . This is partly Google's fault, of course. But clearly historical accuracy is not the point of generative AI (or at least this particular model). If you want an accurate picture of the founding fathers, why would you not go to Wikipedia? You're asking a generative model -- an artist with a particular style -- to generate completely new images for you in a fictional uni…

> If you want an accurate picture of the founding fathers, why would you not go to Wikipedia? You're trying very hard to justify this with a very limited use case. This universe, in which the generated images live, is only artificial because Google made it so.

> You're trying very hard to justify this with a very limited use case

A very limited use case? These are cherry-picked examples. I'm responding to the specific cherry-picking they're doing.

> This universe, in which the generated images live, is only artificial because Google made it so.

No, it's artificial because it's coming from a generative model. If you want your image generator to always be 100% historically accurate, you better train it that way -- Google chose not to. But then don't be annoyed when it can't draw a picture of a dragon. In fact, what would you expect to happen if you asked it to draw "a dragon attacking a German WWII brigade?" Would you lose your mind because some of the German soldiers are Asian women? There's a damn dragon in the picture! What does accuracy even mean at that point?

Re: Google to pause Gemini image generation of people after issues

#882

Post by feoren »

multicast wrote: post #501

We live in times were non-problems are turned into problems. Simple responses should be generated truthfully. Truth which is present in today's data. Most software engineers and CEOs are white and male, almost all US rappers are black and male, most childminder and nurses are female from all kinds of races. If you want the person to be of another race or sex, add it to the prompt. If you want a software engineer from…

> We live in times were non-problems are turned into problems.

This is exactly what everyone who benefits from the status quo always says.

> Most software engineers and CEOs are white and male

55% of Software Engineers are white; 80% are male.[1] So somewhere around 44% of software engineers are white and male. That's not "most". You think it's perfectly fine if 100% of generated images for "Software Engineer" are white males, when ~56% are not in real life? What exactly is your definition of "truth" here?

An unregulated generative model trained on the entire Internet is not going to regurgitate facts, it's going to regurgitate existing beliefs, which is damaging to people who those existing beliefs harm, and to the people who are trying to change those beliefs to actually align better with facts. It is an amplifier of pre-existing perceptions and prejudices; facts have nothing to do with it, except for when they serendipitously line up with common belief. But common beliefs often don't align with the facts -- yes, even yours, as we discovered when you spouted off that "most software engineers are white male" misinformation as if it was some unarguable fact.

[1] https://www.bls.gov/cps/cpsaat11.htm

Re: Google to pause Gemini image generation of people after issues

#883

Post by epivosism »

Dall-E 3 at least exposes the adjusted prompt. Here's an example of it; you can get it if you hit the API directly and look at revised_prompt.

https://twitter.com/eb_french/status/1760763534127010074

At least they show it to us; and you can prepare or attempt to convince the GPT which interprets your prompt into not doing it quite as much (although the example above is where I failed; it seems like it's on to me, because the violation of what I'm asking for is so egregious.)

Re: Google to pause Gemini image generation of people after issues

#884

Post by lupusreal »

janalsncm wrote: post #815
logicalmonster wrote: post #596

Personally speaking, this is a blaring neon warning sign of institutional rot within Google where shrieking concerns about DEI have surpassed a focus on quality results. Investors in Google (of which I am NOT one) should consider if this is the mark of a company on the upswing or downslide. If the focus of Google's technology is identity rather than reality, it is inevitable that they will be surpassed.

Isn’t the fact that Google considers this a bug evidence against exactly what you’re saying? If DEI was really the cause, and not a more broad concern about becoming the next Tay, they would’ve kept it as-is. Weird refusals and paternalistic concerns about harm are not desirable behavior. You can consider it a bug, just like the ChatGPT decoding bug the other day.

They released it like this because people inside Google were too afraid to speak out against it. Only now that people outside the company are shouting that the emperor is naked do they seem to suddenly notice the obvious.

Re: Google to pause Gemini image generation of people after issues

#885

Post by iteratethis »

sega_sai wrote: post #266

That is certainly embarassing. But in the same time, I think it is a debate worth having. What corrections to the training dataset biases are acceptable. Is it acceptable to correct the answer to the query "Eminent scientist" from 95% men, 5% woment to 50%/50% or to the current ratio of men/women in science ? Should we correct the ratio of black to white people in answering a generic question to average across the gl…

Why is bias a problem?

When you prompt "business man" and it outputs a white man, this is quite probably reflective of representation in reality.

Whether this overrepresentation is even a problem at all is debatable as the idea that every job, role or subgroup of people is perfectly diverse or that this even should be the goal isn't just ridiculous, it's demographically impossible.

If you do have a problem with a specific representation in actual reality, reality itself should change. Which it does, it just takes time.

In the meanwhile, just prompt "black business man" if that's what you were after.

Re: Google to pause Gemini image generation of people after issues

#886

Post by rahidz »

iambateman wrote: post #354

I believe this problem is fixable with a “diversity” parameter, and then let the user make their own choice. Diversity: - historically accurate - accurate diversity - common stereotype There are valid prompts for each. “an 1800’s plantation-owner family portrait” would use historically accurate. “A bustling restaurant in Prague” or “a bustling restaurant in Detroit” would use accurate diversity to show accurate sampl…

"turning a big dial taht says "Racism" on it and constantly looking back at the audience for approval like a contestant on the price is right" - dril

Re: Google to pause Gemini image generation of people after issues

#887

Post by iteratethis »

Personally, I'm waiting for the day that Google faces consequences for its actions.

For the longest of times they've had this giant money printer funding what is effectively a playground. An incubator of serial failure but without any consequence.

The trouble is, Google is close to immune to feedback. It's billions of users aren't customers.

Re: Google to pause Gemini image generation of people after issues

#888

Post by rifty »

sega_sai wrote: post #266

That is certainly embarassing. But in the same time, I think it is a debate worth having. What corrections to the training dataset biases are acceptable. Is it acceptable to correct the answer to the query "Eminent scientist" from 95% men, 5% woment to 50%/50% or to the current ratio of men/women in science ? Should we correct the ratio of black to white people in answering a generic question to average across the gl…

I think baseline it should be tuned to be statistically accurate. The problem is that people leave a lot in their prompt to be implied. Some interaction designers use this as an opportunity to infill their own subjective opinion of what should be inferred as a way to 'take care' of the user.

This isn't their only option... they could also just ask for more information.

A good approach here would be ask the user to further clarify what exactly they want before generating a person — "Do you want a random depiction or a specific depiction". A good tool for users is one which helps them be and feel more tactically or predictably in control of it; which means making them aware of its behavioural pitfalls so they can avoid them if they want to.

Re: Google to pause Gemini image generation of people after issues

#889

Post by TheKarateKid »

Satisfy4400 wrote: post #864
TheKarateKid wrote: post #824

Earlier quoted context omitted.

It has been known for a few years now that Google Image Search has been just as inaccurately biased with clear hard-coded intervention (unless it's using a similarly flawed AI model?) to the point where it is flat out censorship. For example, go search for "white American family" right now. Out of 25 images, only 3 properly match my search. The rest are either photos of diverse families, or families entirely with POC…

Just tried the same search and here are my results for the first 25 images: 6 "all" white race families and 5 with at least one white person. Of the remaining 14 images, 13 feature a non-white family in front of a white background. The other image features a non-white family with children in bright white dresses. Can't say I'm feeling too worked up over those results.

I was aware of the white background results, hence my other example query. Both yielded the same result.

7/25 = 0.28 = 28%. That's awful accuracy. Google would be out of business if their general search accuracy had a similar success rate.

Interesting how "black american family" yields results where not a single person in the result is anything but Black. I suppose Google doesn't think that blended families are possible for this query. Where's that 28% precision rate this time?

Re: Google to pause Gemini image generation of people after issues

#890

Post by Jensson »

smsm42 wrote: post #860
logicalmonster wrote: post #596

Personally speaking, this is a blaring neon warning sign of institutional rot within Google where shrieking concerns about DEI have surpassed a focus on quality results. Investors in Google (of which I am NOT one) should consider if this is the mark of a company on the upswing or downslide. If the focus of Google's technology is identity rather than reality, it is inevitable that they will be surpassed.

We are talking about the company that when a shooting happened in 2018, banned all the goods containing substring "gun" (including Burgundy wines, of course), from their shopping portal. They're so big nobody feels like they need to care about anything making sense anymore.

The censorship arm of Google is powerful but not competent. So yeah you get dumb keyword matching returning 0 results. I remember something similar to "girl in miniskirt" returning 0 results on google since someone wrote an article about it. As far as I know the competent engineers doesn't work on this.
Post reply on HN