Google CEO Sundar Pichai told employees in an internal memo that historically inaccurate images and text generated by Google's Gemini AI "offend our users and display bias."
Last week, Google suspended Gemini's ability to generate images after it was widely discovered that the model's images were racially disparate, including Nazi-era German soldiers, non-white American Founding Fathers, and even inaccurate depictions of the race of Google's own co-founders. While Google later apologized for the "mistake" and said it was working to re-enable image generation in the coming weeks, Tuesday's memo was the first time the Google CEO responded broadly to the controversy.
In the memo, first reported by Semafor, Pichai said the company "has been working around the clock" to resolve "problematic text and image replies in the Gemini app." He didn't say Google has solved the problem. "No one is perfect, especially in this nascent stage of industry development, but we know the demands placed on us will be high and we will continue to work hard no matter how long it takes," he wrote.
You can read the full memo from Sundar Pichai to Google employees below:
Hello everyone!
I'd like to talk about a recent issue with text and image replies in the Gemini app (formerly Bard). I know some of its responses offended our users and showed bias, which is completely unacceptable and was a mistake on our part.
Our team has been working around the clock to resolve these issues. We've seen substantial improvements to various prompts. No Al is perfect, especially at this emerging stage of the industry's development, but we know the standard is high for us and we will keep working hard no matter how long it takes. We will re-examine what happened and make sure we address these issues at scale.
Our mission to organize the world's information and make it universally accessible and useful is sacrosanct. We always strive to provide users with useful, accurate and unbiased information in our products. That's why people trust our products. All our products, including emerging Al products, must adhere to this philosophy.
We will take a clear set of actions, including structural changes, updated product guidance, improved release processes, rigorous evaluation and red teaming, and technical advice. We are reviewing these issues comprehensively and will make necessary changes.
While learning lessons, we should also build on the products and technologies released in the AI field over the past few weeks. This includes some of the foundational progress we've made on the underlying models, such as our 1 million long context window breakthrough and our open model, both of which have been well received.
We know what it takes to create great products that are used and loved by billions of people and businesses, and with our infrastructure and research expertise, we have an excellent springboard to realize the AI wave. Let’s focus on what matters most: building useful products that users can trust.