Recently, we introduced a new image generation feature for the Gemini conversational app (formerly Bard), allowing users to create images of people.
Unfortunately, this feature did not meet expectations, generating inaccurate and even offensive images. We appreciate user feedback and apologize for the shortcomings of this feature.
Acknowledging the error, we have temporarily suspended the generation of people’s images on Gemini while we enhance and refine the feature.
What Went Wrong
Gemini’s image generation feature operates independently from our other products and utilizes an AI model known as Imagen 2.
While developing this feature, we aimed to avoid pitfalls encountered with image generation technologies in the past, such as creating inappropriate or biased content. Our goal was to provide diverse and inclusive image results suitable for all users.
However, the model failed to deliver accurate responses in certain scenarios, either showing an array of images where unnecessary or being overly cautious and omitting responses completely.
These issues caused the model to produce misleading and inappropriate images at times.
Moving Forward
Our intention was never to withhold images of specific groups or inaccurately depict historical events. We have disabled the people image generation feature to make significant enhancements before reintroducing it, following thorough testing.
Gemini is designed as a creativity and productivity tool, with occasional limitations in generating content, especially concerning current events or sensitive topics. While mistakes may occur, we are committed to addressing and resolving any issues promptly.
We advise using Google Search for up-to-date and reliable information, as it offers a more comprehensive coverage of various topics.
While occasional errors may happen, we are dedicated to ensuring the responsible and safe deployment of AI technology, recognizing its potential benefits and challenges.
