Nazi soldiers of color or the Pope imagined as a woman: last week, Google suspended its image creation tool Gemini after it generated embarrassing results, in some cases refusing to represent white people or generating images historically false.
The tool, which allows you to create images from a simple text query, has also portrayed the Founding Fathers of the United States or early 19th century U.S. senators as people of color, which is incorrect.
Elon Musk, also behind an AI tool called Grok, mocked Google's chatbot by making fun of one of its images which depicted George Washington as a black man.
“The woke virus is killing Western civilization,”
the American billionaire writes on X.
The controversy swelled again after Gemini claimed that it was
“difficult to say”
whether Hitler or Elon Musk had had a more negative impact on society, as evidenced by screenshots of conversations with users.
“The response here is appalling and inappropriate.
We are implementing an update so that Gemini no longer displays the response
,” announced a Google spokesperson cited by The Telegraph.
These serial controversies pushed Sundar Pichai, the CEO of Google, to address his teams.
In an email revealed by the American media Semafor, he deplores Gemini's embarrassing responses and commits to making structural changes to resolve the problem.
“I know some of the responses offended our users and showed bias.
To be clear, this is completely unacceptable and we were wrong
,” he wrote in the message.
“Our teams are working around the clock to resolve these issues. We are already seeing a substantial improvement in the responses,”
he adds.
To avoid further slippage, Sundar Pichai promises
"structural changes, updated guidelines, robust evaluations and 'red teaming'"
, a practice aimed at discovering the flaws of a product or computer program by testing its security.
The CEO also reminded his teams of the reputational risks of such a controversy:
“Our mission, which consists of organizing the world's information and making it universally accessible and useful, is sacrosanct.
We have always sought to provide users with useful, accurate and unbiased information in our products.
That's why people trust them.
We need to take this approach for all of our products, including our new artificial intelligence products.”
“No AI is perfect”
Most companies that offer AI tools put safeguards in place to prevent these language models from replicating or amplifying bias.
By wanting to get around this bias on its new Gemini AI, available since the beginning of February in the United States, Google got entangled in an opposite problem.
Gemini product director Jack Krawczyk said Google
“designs its image generation tools to reflect the diversity of Google users around the world, consistent with our AI Principles.”
Also read From Dall-E to ChatGPT, how OpenAI democratized a new part of artificial intelligence
Dall-E, the image generation tool from OpenAI, the parent company of ChatGPT, is regularly accused of mainly creating images of white people in positions of high responsibility or of depicting people of color in stereotypical roles .
Its new version, Dall-E 3, has also been improved in this sense in order to reduce bias
“for the sake of diversity.”
While the web giant is known for its caution before putting new products on the market, Google has had to change its habits with the emergence of generative AI.
Gemini, formerly named Bard, was launched to compete with ChatGPT.
“No AI is perfect, especially at this stage of development of the industry, but the bar is high for us
,” Sundar Pichai also wrote in his message.