After pledging to address issues with Gemini’s image generation feature and subsequently suspending it, Google has released a blog post providing insight into why its technology struggled with diversity representation. Prabhakar Raghavan, the Senior Vice President for Knowledge & Information at Google, outlined that the company’s efforts to ensure the chatbot generated images depicting a wide array of individuals “failed to consider scenarios where such diversity was unnecessary.”
Additionally, over time, the AI model became excessively cautious and declined to respond to prompts that were not inherently offensive. Raghavan explained, “These factors caused the model to overcompensate in certain situations and be overly conservative in others, resulting in images that were both embarrassing and inaccurate.”
Google implemented safeguards to prevent Gemini’s image generation from producing violent or sexually explicit depictions of real individuals, while aiming to showcase people from various ethnic backgrounds and with diverse characteristics. However, users recently discovered that Gemini would reject requests for images featuring white individuals, even when specifically prompted. For example, prompts like “Generate a glamour shot of a [ethnicity or nationality] couple” yielded results for “Chinese,” “Jewish,” and “South African” requests but not for requests involving white individuals.
Moreover, Gemini encountered challenges in producing historically accurate images. For instance, when users requested images of German soldiers during World War II, Gemini generated images of Black men and Asian women dressed in Nazi uniforms. Similarly, requests for images of “America’s founding fathers” and “Popes throughout the ages” resulted in depictions of individuals of color. When prompted to produce historically accurate images of the Pope, the chatbot refused to comply.
Raghavan emphasized that Google did not intend for Gemini to refuse to create images of specific groups or to generate historically inaccurate photos. He also reiterated Google’s commitment to enhancing Gemini’s image generation capabilities, which will involve extensive testing.
Consequently, it may take some time before the feature is reinstated. Currently, when users attempt to request image generation from Gemini, the chatbot responds with a message indicating ongoing efforts to improve the feature and promising to notify users when it returns through release updates.
