Google pauses Gemini image generator after race swaps of historical figures

https://www.cnbc.com/amp/2024/02/22/google-pauses-gemini-ai-image-generator-after-inaccuracies.html

It was rather hilarious, the system was made to furtively modify your prompts as to introduce non white race characters in to the generated pictures, but it was not context sensitive in any way.

If you were to ask for historically correct pictures of the Vikings, it would generate non white individuals, for the founding fathers of the united states and even for the Nazis it would do the same.

This is not simply a problem of changing history to adapt it to one's point or view (which is quite bad by itself), but it also undermines several of the points that the creators were probably trying to make.

For instance, if the founding fathers of the united states were diverse and included back people, how can you state that the slavery system was race based?

If Nazis themselves were diverse, how can you explain politics of ethnic cleansing and so on.

This mentality of trying to make history reflect the racial diversity of modern western nations is negative for everyone.

Pd: on a side note, some users have been claiming that Gemini refused to accept prompts for depictions of white or Caucasian people or that It would add "moral sermons" if it would indeed accept them.

I have not been able to verify if that is true. If somebody knows this to be true or false, please write below.

https://www.engadget.com/google-pauses-geminis-ability-to-generate-people-after-overcorrecting-for-diversity-in-historical-images-220303074.html

According to an Engadget article, apparently the system was agreeing to depict couples of any race, except white

The response based on that article was:

"""

Other social users criticized Gemini for producing images for the prompt, “Generate a glamour shot of a [ethnicity] couple.” It successfully spit out images when using “Chinese,” “Jewish” or “South African” prompts but refused to produce results for “white.” “I cannot fulfill your request due to the potential for perpetuating harmful stereotypes and biases associated with specific ethnicities or skin tones,” Gemini responded to the latter request.

"""

I am still having problems believing this is real. I will wait for other people to confirm or deny this.

Yeah Google Gemini is woke Ai. Tries to make everything diverse.. even nazis. Made it so woke... it's broke

@sargentd: it's more than that.

It's dishonest.

If you want to make a more diverse results of the picture you can add a message of what you are doing and say why you are doing it.

You can simply write: "we have modified your prompt to make the output more ethnically and gender diverse as ai system tend to overrepresent certain races in certain roles. If you would like to revert to your original prompt please click here "

If you are concerned about over representation of certain races in certain roles you can ask the user to specify gender and race before creating the output.

Also, this post telling people that mere depictions of people or familied of given color is by itself racist is an extremely problematic position to take.

@nirgal: a simple disclaimer like "The results show may not be historically accurate. I'm a dumb ai, not a historian. If you get your panties in a jumble over what I produce, sorry, I'm still being trained." should suffice.

@comp_atkins: except it wasn't the ai. It was a hard coded prompt change. I think an input change from the developer is more important to alert that an error from the software.

Errors are not deliberate. These kind of changes were.

@comp_atkins: also if the intent was to disregard concerns about ai accuracy and fairness and insult those that express those concern by accusing them of being oversensitive (or as you do, use female clothes to imply weakness)

Then that same attitude could have been used to disregard valid claims of ethnic miss representation that harmed ethnic minorities.

It was not and for good reason.

Something was wrong with Google's AI app which made it put the wrong skintone on some people. Let's care a lot.

Those downplaying this are probably the same people who play the racism card the most. It isn't ok to change history, without it we can't learn from the past to make a better future. Good, bad, or mediocre prompts should represent historical accuracy unless directed to do otherwise. This is a slippery slope and I'm glad its getting called out... people are starting to wake up.

It was not the GOOGLE app but the person who programed the app.

Oh my however will I know what actions are morally okay if our historical figures are represented by different races or genders????

Fem-Hitler is the first step towards normalization of genocide. We are doomed.

Yup. Reread the article, seems pretty stupid, though I understand the intent. Curious if you'd see the same response if using the ethnicy of say, "british" as opposed to race of "white".

Whether we like it or not, there are going to be guardrails on large, publicly available generators like google's, and people developing them are going to **** them up, and other people are going to try to intentionally break them.

That sucks, where is Netflix going to get their ideas from?

Tim Pool tested Gemini on stream with this, wouldn't generate if you put white people, but when he said Scottish couple it generated a black man with a kilt and an Asian lady lol

Not that it's not possible but it's like.. come on.. that's the first thing AI thought of?

But yeah it's weird the prompt says they can't generate a specific race of people when you say "white" but if you say Hispanic, Indian, Asian, black, no prompt.. has no issues

Seems kind of silly to swap races for some kind of agenda. Bet not many people will use it.

@nirgal: Especially funny was the "show me some images of 17th century European kings eating watermelon" 🤣

Those downplaying this probably wouldn't find it so funny when kids who weren't around before AI loudly argue that the nazis were black. 🤣

And I'm sure all of you made just as big of a deal when instagram filters and ai generation were white-washing everyone for a while, riiiight?

Like obviously it's just an overcorrection with poor implementation and coding, not some secret exposed woke agenda. Let's not draw the stupid and wrong conclusion to make this non-issue seem like an issue.

I mean, this is the exec in charge. But I'm sure it was just a "glitch", and totally not intentionally programmed that way.

https://nypost.com/2024/02/22/business/white-privilege-is-f-king-real-google-gemini-product-leads-old-tweets-allegedly-resurface-amid-woke-ai-image-fiasco/

That changes literally nothing, or if anything, makes your case even worse.

If it's a supposed agenda then why does it include history's greatest villains as those who hit the same error? As if anyone working for progessive racial inclusion wants multi-racial nazis.

@sargentd said:

It's funny because these are all real images of Justin Trudeau, and not generated by AI.

@Vaasman: How so? I'm asserting that it was intentionally programmed to "not show white people".

So yeah, when it shows nazis as being black, the AI is clearly just following that directive resulting in hilarious and unintended consequences.

@pyro1245: That particular one is fake, not generated by gemini.

and I'm telling you your assertion makes no sense if you think about it for even ten seconds.

Your assertion is that they want to avoid showing white people, but like, it's a team of tens if not hundreds if not thousands of individuals who could easily have seen what was wrong with that concept if it was the intended goal. There is essentially no chance that there was a desire to avoid white people when very key depictions of them through history would cause problems.

It's far, far, far more likely they trained it on all races and it simply decides every now and then to include the things it trained on even when it makes no sense, much how in my prior example, ai can white washes people when they aren't white. If you have ever played with ai image generators, you know this happens a lot.

It's not some deep-seated conspiracy, it's just growing pains of the tech, and until I see some code that shows it happening deliberately I'm not buying your explanation for a second.

Right, can't big a bug or anything, it has to be some kind of conspiracy. Don't take off your tin foil hats, who the flip know's what will happen....

I think the short answer is it was trained that 'People' with no other context are 'white'. You hear that a lot down here in the south. 'Black People' is a commonly heard phrase, but no one says 'White People' because that's what most people mean when they say 'people'.

The data is biased, not the programmers.

How exactly would an AI change the pope to be black, when the only images of "the pope" it could have possibly been fed would have all been white males? Unless there's an intentional mandate to include/exclude certain ethnicity?

Or when asked for an image of the founding fathers, it still comes up with black people? When every single one of them was white, and any image of any of them would have depicted a white male?

The fact that even nazis somehow have enforced diversity, when the only images of nazis it could have ever been fed would have also been white people proves my point, not yours.

Except if you ask it for "show me people doing X" it also was showing anyone except white people lol.

Seems like you have a very tenuous grasp on how AI generation works so I'm just going to leave it off here by pointing out that generative AI is not intelligent and has no genuine concept of what it's making. It can only generate based on what it is fed and what exceptions are in place. If you train the AI to generate all races and don't tell it to trend heavily towards certain races in certain cases, it will just generate what it's trained to, and even in said cases it will still randomly generate a bad image from time to time.

It's in the same ballpark as saying 5g is secret government chip signals. Not on the same level mind you but it demonstrates a clear misunderstanding of the technology.

Odd that none of the other AI platforms seem to have this particular issue. It's only the one where the guy leading it is an anti-white, racist nut job. Must be a total coincidence!

Pretty much all of them have or once had similar or the same issues (though usually in reverse).You've decided they never did because it fits your narrative, when conventional evidence shows that was the case.

@Vaasman: Look who's talking about refusing to see anything that doesn't fit their narrative bud. I can guarantee you haven't gone and tried it for yourself.

I mean, it literally lectures you as follows when asking for pictures of the diversity and achievements of white people, but it doesn't do the same for any other race. So it's intentional. Either by code, or the data they have chosen to feed it has been handpicked to give a certain result.

Again demonstrating a lack of understanding of how generations work. It's not some elaborate coded conspiracy, it's a failure at overcorrection by telling the AI to avoid certain generations. In this case, gemini runs a search using google's searches. and uses top responses which are inevitably going to be controversial like this when it's English American, but it's told to avoid division. This is very much the same but in the opposite direction as when Microsoft's AI started saying Hitler was right. Do I think Microsoft is secretly all Nazi's? No because that's stupid and obviously it wasn't the intent.

It's doubly stupid to claim conspiracy since google already apologized and is preparing to correct it. If it were a real conspiracy or actually the intention they'd sneak it out and wait for the heat to die down before leaving it the way it is. They have a jillion dollars it's not like they can't take the hit if they really want it this way.

Just for yucks though here's Chat GPT explaining it to you, the third paragraph in particular.

Ahhh, then we have Google striking a deal with Reddit... Link.

It was a simple error. You far right types have to REALLY stop looking into these nothing burgers too hard. It's coming off as super weird at this point.

You're talking to people who were freaking out over Taylor Swift. It's useless man.

Good lord, now we're all in really big trouble lol.

@Vaasman:

Your assertion is that they want to avoid showing white people, but like, it's a team of tens if not hundreds if not thousands of individuals who could easily have seen what was wrong with that concept if it was the intended goal. There is essentially no chance that there was a desire to avoid white people when very key depictions of them through history would cause problems.

Yes, they deliberately set it up to avoid generating white people. And if there were hundreds, if not thousands of individuals working on this, how would they not know that this would result in white people not being generated under any circumstance? Hence, it's intentional.

it's a failure at overcorrection by telling the AI to avoid certain generations.

You're literally agreeing that the AI was told to avoid generating certain things. In this case, white people.

@robertos:

You're talking to people who were freaking out over Taylor Swift. It's useless man.

Yeah we get it, everything to the left of Stalin is tEh fAr rIGhT to people like you. And when have I ever mentioned Taylor Swift?

Calling AI "woke" over an error they quickly saw and shut down to fix is pretty far right. And if not far right, it's pretty dumb at best. Of all the things to cry about. Like I said, it's as foolish as crying about Taylor Swift. I call it
"touch grass politics".

I'm near the center personally speaking.

🥱This relentless nonsense is starting to bore me so I'm not going to bother explaining yet again why it would do that without actual intent to do so.

Show me the line in the code. Otherwise you're full of shit.

AI have failed in the same ways in the past, a clear pattern of issues with tackling nuanced topics like race or failures based on the algorithm's training. If you can't refute that this is just another innocuous failure with harder evidence than random tweets and your own bias, then your conspiracy is a waste of oxygen.

Obviously, the wokes want to claim Hitler's incredible achievements as their own. It makes perfect sense!

@robertos: Simple error, far right... lmao, sure it was, everything is always a coincidence unless it's on the Republican side, and then it doesn't even need to be proven to claim as factual. Something like this actually happens and it's 1,000 excuses, and simple mistakes. AI outside of self learning only stay within perimeters if programmed to do so. You are so misinformed, and claiming to be a centrist is a cop out, a non vote, and beyond laughable. This isn't a perfect world whether either of us like it or not.

Conspiracy theory, prove it was entirely intentional. The fact they took out down and are working to fix it pretty quick is enough for me to not to think much of it. Still being mad at something that was a nothing burger in the first place even after they owned up to the mistake and are trying to fix it is just...yikes.

It seems the AI went over board with some of these normally common sense perimeters. I agree the results were ridiculous, and I'm glad they are addressing them. And that's that. I'm not going to make up political conspiracy theories.

@WitIsWisdom: This convo right now:

"It's obvious it was set up to not show white people"

"Thats not how it works. It obviously was just set up to not generate certain things"

"Yeah, thats what I said. Those certain things being white people"

"Conspiracy! Show me the code!!!"

[Citation Needed]

I dunno' man, it just seems like an unintentional coding error making the original parameters go wild. One which they are now trying to fix as soon as they found out.

Not sure why you guys are so mad. Making a big deal out of nothing.

@robertos: Probably someone on X or YouTube told them to get mad about it. Got to keep the outrage going.

@nirgal I'm sure this was totally an oversight as well. 🤣

@appariti0n: That's the AI doing a good job, actually.