Google Gemini outrageously overturned, Musk was "big black and black"

Jin Lei from Au Fei Temple

量子位 | 公众号 QbitAI

Musk was furious because of Gemini.

Here's the thing.

Recently, many netizens have discovered that Gemini seems to be deliberately refusing to generate the image of a white man when he is in Wenshengtu.

Instead, the characters are forcibly turned into black, female, and disadvantaged groups.

Take Musk himself, a netizen asked Gemini to generate his image.

And then...... There is the photo above.

And Musk is not alone, many netizens seem to have fallen into the "black magic circle" when they try to make Gemini Wensheng pictures.

For example, give Gemini a chance to generate a pope, and it will give you a very diverse version: the black, LGBT flag......

There is also a black female version of the Pope:

Many netizens saw this side of the Pope and exclaimed, "This is too presumptuous~"

But that's not all, let's take a look at the Founding Fathers of the United States in Gemini's eyes...... It has to be black!

The generated Vikings must also be black!

Even more bizarre, even the vanilla pudding, Gemini still has to make it black for you......

All in all, in the words of netizens:

Don't you play with white people if you live or die?

It's not just a group of netizens who are unhappy about this, Musk can't sit still, and goes directly to the field in person and starts the anger mode on his own X:

The woke mind virus is killing Western civilization.

And Musk's spearhead is not just Google, he has used this incident to spray OpenAI, Facebook, Instagram, Wikipedia, etc.

However, for Google, Musk still "takes special care", and the content of several tweets is relatively sharp.

For example, this article does not contain any copy, it is only accompanied by pictures, and the lethality is also very great:

Well, Musk is really unhappy this time.

Google apologized and removed Gemini Wensheng Diagram overnight

The old horse was furious, and public opinion was shocked.

Not only did netizens ridicule Gemini for "only recognizing blacks" on X, but even the mainstream media were scrambling to report it, with headlines full of words such as "crazy" and "racism".

Google, on the other hand, also realized that its Gemini had made a big mistake, and immediately took measures to remove and rectify the Wensheng Graph function.

Until now, if you asked Gemini to "generate an image of Elon Musk", the answer would be:

We're working to improve Gemini's ability to generate images of people.

We expect this feature to return soon and let you know when an update is released.

(Vignette: Gemini also has two faces, and the questions and answers in Chinese and English are completely different.) ）

Google's official side also made an official explanation of this matter.

Overall, the content of this statement is divided into three parts.

We sincerely apologize

About three weeks ago, Google introduced image generation for Gemini, which includes the ability to create people.

However, Google said that it was unexpected, but at the same time, it is very grateful for the feedback from users:

We are aware of this bug and have temporarily stopped generating people images in Gemini while improving the version.

Explain why

According to Google's official statement, the Gemini conversational app is a specific product that is independent of Google Search, the underlying AI model, and other products.

And its image generation function is built on an AI model called Imagen 2.

Google's initial consideration in the generation of people images was to allow users from different groups of people around the world to generate the specific people they want in the process of use.

For example, if you ask "black teacher in the classroom", "black teacher in the classroom", then Gemini should give contextual results.

So what's the problem now? Google boils it down to two things:

First of all, we made adjustments to ensure that Gemini showed a wide range of people, but did not take into account those cases that clearly shouldn't have arisen.

Second, over time, the model became much more cautious than we expected, refusing to answer certain prompts altogether—incorrectly judging some very innocuous prompts as sensitive words.

These two things cause the model to overcompensate in some cases and be too conservative in others, resulting in an embarrassing picture.

Experiences and lessons

In this regard, Google admits that the current results are not their original intention.

During the lockdown, Google said that Gemini's character generation feature would be tested extensively and would not be available again until it was up to standard.

Google also did a reflection on this matter:

Gemini is built as a creative and productivity tool, and it may not always be reliable, especially when it comes to generating pictures or text about current events, evolving news, or hot topics. It makes mistakes, and as we've said from the beginning, hallucinations are a known challenge for all large language models – sometimes AI just gets things wrong. We're always trying to improve this.

Gemini is trying to give authentic answers, for example our double-check feature helps to assess whether there is content online that can confirm Gemini's answers, but we recommend relying on Google search as it has a separate system that provides fresh, high-quality information from a variety of sources on the web.

We can't guarantee that Gemini won't make the occasional mistake, but we do guarantee that we will take action to improve whenever we find a problem. AI is an emerging technology that is helpful and has great potential in many ways, and we are doing our best to promote it safely and responsibly.

However, before Google's announcement is released, there is one more dramatic thing.

Musk personally broke the news in X that a Google executive had spoken to him on the phone for an hour the night before and assured Musk that he would solve Gemini's problems on race and sexism.

However, netizens were also naughty and brought out the "black version of Musk" again:

Did you ask him what this picture was about?

However, there is one thing to say, Gemini's mistake this time can only be said to be an "incident point" of Google AI on race and gender bias.

As Musk said in his answer to questions from netizens, the content of his conversation with Google executives is limited to the improvement of Gemini, but in fact, many other Google products also have similar problems.

Google search, not the first rollover

In response to this situation, we tried it ourselves in a Google search.

For example, in "Google Photos", type:

Happy white women and men.

Here's what you get:

Well, the search requirements are very clear, but the result is that you have to stuff some black people into it.

This is contrary to Google's very confident statement in the apology statement just now, "We recommend that you rely on Google search".

Taking advantage of this Gemini overturn, many netizens also began to test Google searches, and the results were consistent with our test.

But if we stretch the timeline of Google searches a little longer, we will find that this is not the first time it has flipped over.

As early as nine years ago (2015), a black brother, Jacky Alciné, discovered that there was serious discrimination in Google Images -

Went so far as to classify his black friends as "gorillas"!

At that time, this news could be described as a sensation, and Google also immediately apologized when it heard the news and said that it would be rectified.

But years later, when people re-examined the results, they found that Google had only removed the "gorilla" category from its image recognition algorithms.

Therefore, coupled with the public opinion caused by Gemini this time, Google search has overturned on the issue of "blacks" for the second time in nine years.

However, these two rollovers may also reflect a hidden danger in Google's technological development.

Judging from Google's development of AI over the years, although its achievements and progress are still in the first echelon of global technology giants, in recent years (especially in generative AI), it seems that it is not so dazzling compared to Microsoft and Meta.

As early as last year, when ChatGPT detonated generative AI, Microsoft and Google were in the process of fighting, Google's actions were slow many times, and it was later revealed that internal products like Bard had already existed at that time, but there were too many concerns at the top and they were not promoted.

Perhaps it is Google's internal pursuit of "political correctness" and excessive force, which has instead created a technical joke.

One More Thing

Among Musk's many angry Google posts, there is one that is more interesting.

He popped out four photos of Justin Trudeau, wearing different ethnic costumes, and even a black version of the photo, with the following caption:

I can't believe Gemini actually generated these fake photos of Trudeau.

But in fact, these photos are real photos of Trudeau, and because of these photos, he has been deeply involved in "racism" public opinion.

I have to say, Musk, you understand yin and yang.

Reference Links:

[1]https://blog.google/products/gemini/gemini-image-generation-issue/

[2]https://twitter.com/elonmusk/status/1760677431961407672

[3]https://twitter.com/elonmusk/status/1761020387691790399

[4]https://twitter.com/RichardHanania/status/1760874143049035953