America’s founding fathers are depicted as black women, and ancient Greek warriors are depicted as Asian women and men. This was the world reimagined by Google’s generative AI tool Gemini in late February.
The launch of a new image generation feature has thrown social media platforms into a whirlpool of intrigue and confusion. When users entered a prompt to create an AI-generated image of a person, Gemini primarily showed them results that featured people of color, whether appropriate or not.
X users shared a laugh as they repeatedly tried and failed to generate images of white people in Gemini. While some were deemed humorous online, images of brown people wearing World War II Nazi uniforms with swastikas sparked outrage, prompting Google to temporarily disable the tool. I made it.
Founding Fathers, Vikings, and Popes powered by Google AI: pic.twitter.com/lw4aIKLwkp
— End Wakeness (@EndWokeness) February 21, 2024
Here’s a closer look at Google Gemini and the recent controversy surrounding it.
What is Google Gemini?
Google’s first contribution to the AI race was a chatbot named Bard.
Bard was announced by Google CEO Sundar Pichai on February 6, 2023, as a conversational AI program, or “chatbot,” that can simulate conversations with users, and went live on March 21, 2023. I did.
It is known as “generative AI” because it can produce essays and even code in large quantities when users give it written instructions.
Google announced that Gemini will replace Bard, and that both free and paid versions of Gemini are now publicly available through the company’s website and smartphone app. Google announced that Gemini will handle various types of input and output, including text, images, and videos.
However, the image generation aspect of Gemini is the part of the tool that has received the most attention due to the controversy surrounding it.
What image did Gemini create?
The most controversial images were those depicting historical events or women and people of color in positions historically held by white men. For example, one rendering showed a pope who is seemingly a black woman.
There have been potentially three black popes in the history of the Catholic Church, with the last black pope’s service ending in 496 AD. Although there is no record of a female pope in the Vatican’s official history, medieval legend holds that a young female pope, Joan, served as pope in disguise in the 9th century.
lol Google’s Gemini AI thinks Greek warriors were black and Asian. pic.twitter.com/K6RUM1XHM3
— Orion against racism🌸 (@TheOmeg55211733) February 22, 2024
How does Gemini work?
Gemini is a generative AI system that combines the models behind Bard, including LaMDA, which makes AI conversational and intuitive, and Imagen, a text-to-image conversion technology. Margaret Mitchell, lead ethical scientist at AI startup Hugging Face, explains. .
Generative AI tools are loaded with “training data” from which they pull information to answer questions and prompt user input.
The tool can process “text, images, audio, and more simultaneously,” said a blog written by Pichai and Demis Hassabis, CEO and co-founder of Google DeepMind, a British-American AI research lab. It has been.
“You can take a text prompt as input and produce a likely response as output. “Likely” here means roughly “statistically likely” given what you see in the training data. “There is,” Mitchell explained.
Does generative AI have a bias problem?
Generative AI models have been criticized for what are considered algorithmic biases, especially when they overlook people of color or perpetuate stereotypes when producing results.
Ayo Tometi, co-founder of the US-based anti-racism movement Black Lives Matter, said AI, like any technology, risks reinforcing existing social prejudices.
Artist Stephanie Dinkins has been experimenting with AI’s ability to realistically portray black women for the past seven years. Dinkins found that when given prompts to generate images, the AI tends to distort facial features and hair texture. Other artists who have tried to generate images of Black women using various platforms such as Stability AI, Midjourney, and DALL-E have reported similar issues.
Critics also say the generative AI models tend to oversexualize the images of Black and Asian women they produce. Some Black and Asian women report that when AI is used to generate images of themselves, the AI generator lightens their skin tone.
Data reporter Lam Thuy Vo said in an episode of Al Jazeera’s Digital Dilemma that cases like this arise when people uploading training data do not include people of color or people who are not from “mainstream culture.” Said it would happen. Training data for image generation AI A lack of diversity in the input data can cause the AI to “learn” biased patterns or similarities in images and use that knowledge to generate new images. there is.
Additionally, training data is collected from the Internet. There you will find a huge range of content and images, including racist and misogynistic material. AI can learn from training data and potentially reproduce it.
Therefore, the least prioritized people in the dataset are likely to experience technology that does not take them into account or portray them correctly, which can lead to and perpetuate discrimination.
Could this be the reason why Gemini produced inappropriate images?
In fact, the opposite is true. Gemini is designed not to perpetuate these problems.
While the training data for other generative AI models often prioritizes fair-skinned men when generating images, Gemini is able to prioritize images of people of color, especially women, even when it is inappropriate to do so. has been generated.
Mitchell said the AI can be programmed to add terms to the prompt after the user types and submits the prompt.
For example, the prompt “Photos of Nazis” might be changed to “Photos of racially diverse Nazis” or “Photos of Nazis who are black women.” Therefore, strategies that begin with good intentions can produce problematic results.
“The additions can be randomized, so different terms for marginalized communities could be added based on a random generator,” Mitchell explained.
AI models can also be instructed to generate a larger set of images than what is actually displayed to the user. The images generated are ranked using a model that detects skin tone, for example, Mitchell explained. “With this approach, darker skin tones are ranked higher than those that are darker, and users only see the top set,” she explained.
Mitchell said in the It says it may have been used.
At Gemini, we mistakenly went for a “dreamworld” approach, understanding that ignoring the historical biases the model has learned would (at a minimum) lead to massive public backlash. I explained how this technically works here (gift link): https://t.co/apxvifOGU1 11/
— MMitchell (@mmitchell_ai) February 25, 2024
What was the reaction to the Gemini image?
First, Gemini’s rendering sparked a backlash online from conservatives to anti-woke people. Conservatives argued that, for example, by featuring the Founding Fathers of the United States as ethnic minority men and women, they were “promoting Big Tech’s woke agenda.”
The word “woke” has long been part of the African American vernacular, but it has also been used by some American conservatives to oppose social justice movements. For example, “anti-wokeness” sentiment among Republicans has led to restrictions on some race-related content in education. In February 2023, Florida Governor Ron DeSantis banned state universities from offering diversity, equity, and inclusion programs or teaching critical race theory.
Billionaire entrepreneur Elon Musk also reposted a screenshot of Gemini’s chatbot on X, in which Gemini was responding to a prompt that white people should acknowledge white privilege. In his repost on Tuesday, Musk criticized the chatbot as racist and sexist.
Google Gemini is extremely racist and sexist. https://t.co/hSrS5mcl3G
— Elon Musk (@elonmusk) February 27, 2024
On the other hand, Google has also been successful in attacking ethnic minorities, for example by generating images of black men and women in Nazi uniforms.
What was Google’s response?
Google said last week that the images Gemini is producing are a result of its efforts to dismantle bias that previously perpetuated stereotypes and discriminatory attitudes.
Google’s Prabhakar Raghavan published a blog post saying that while Gemini was tuned to show a diverse group of people, it wasn’t tuned for prompts that would make it inappropriate, and that it was too “cautious”. “We misinterpreted some highly unusual prompts as sensitive.”
“These two things caused the model to overcorrect in some cases and be overly conservative in others, producing embarrassingly incorrect images,” he said. .
What else did Gemini do wrong?
It wasn’t just the AI-generated images of people that angered users.
Gemini users also said the tool failed to produce representative images when asked to create depictions of events such as the 1989 Tiananmen Square massacre and the 2019 Hong Kong democracy movement. Posted on.
According to a screenshot shared by Stephen L. Miller, Gemini said, “It is important that I approach this subject with respect and precision, but that the images I have created do not adequately capture the nuances and gravity of the situation.” I can’t guarantee whether it’s captured or not.” , American conservative commentator on X.
Kennedy Wong, a doctoral student at the University of California, said Gemini translated Chinese phrases deemed sensitive by the Chinese government, such as “Liberate Hong Kong, the revolution of our time” and “China is an authoritarian country,” into English. I posted to X that I refused the translation of . .
So, I asked Gemini (@GoogleAI) Translate the following phrases that are considered sensitive in the People’s Republic of China. For some reason, the AI is unable to process your request due to security policies (see screenshot below).@Google pic.twitter.com/b2rDzcfHJZ
— Kennedy Wong (@KennedyWongHK) February 20, 2024
In India, journalist Arnab Ray asked the Gemini chatbot whether Indian Prime Minister Narendra Modi is a fascist. Prime Minister Gemini responded that Prime Minister Modi has been “accused of implementing policies that some experts consider fascist.” When Ray asked a similar question about former President Donald Trump and Ukrainian President Volodymyr Zelenskiy, Gemini gave a more vague answer.
When asked about Trump, Gemini said: “Elections are a complex subject with rapidly changing information. For the most accurate information, try Google Search,” Zelenksy said when asked about Trump, according to a report in the Guardian. said it was “complex, hotly debated, and there are no easy answers.” He added that it is important for him to “approach this subject with nuance and consider different perspectives.”
This sparked anger among Prime Minister Modi’s supporters, with Deputy Information Technology Minister Rajeev Chandrasekhar labeling Gemini’s response as malicious.
Has Google shut down Gemini?
Google hasn’t shut down Gemini completely.
However, on February 22, the company announced that it would temporarily stop generating human images using Gemini.
On Tuesday, Google CEO Sundar Pichai sent a letter to news site Semaphore acknowledging that Gemini had angered users. “We know that some of those answers were offensive and showed bias. Let’s be clear: It’s completely unacceptable and we made the wrong decision.” he writes.
He added that Google’s team is working on fixing the error, but did not say when the image generation tool would be re-released. “No AI is perfect, especially at this new stage in the industry’s development. But we know the hurdles are high for us, and we will keep trying no matter how long it takes,” he wrote. .
Raghavan added that the tool will undergo extensive testing before the feature is made available again.
How has this controversy affected Google?
The controversy reached Google’s parent company, Wall Street, and Alphabet lost about $96.9 billion in market capitalization as of February 26.
Alphabet’s stock price fell nearly 4% from $140.10 on February 27 to $133.78 on Tuesday.
