Google is reintroducing its AI image generation capability, which had previously been disabled, following a significant backlash. The updated feature will be available to English-language users of Google’s AI chatbot, Gemini Advanced, and will utilise the company’s newest image generation model, Imagen 3.
In February, Google had removed the functionality to generate images of people due to issues with the tool’s output. Google acknowledged that its Gemini app’s image generation feature produced inaccurate and sometimes offensive images. The company admitted that the tool’s attempt to ensure diverse representation led to overcompensation and excessive caution, resulting in problematic outputs. Google is now working on significant improvements to the feature, including extensive testing, to better align it with user expectations and ethical standards. The company emphasized its commitment to refining the AI while acknowledging that, despite best efforts, occasional errors may still occur.
The controversy erupted after users highlighted problematic images produced by the tool, such as depictions of a Black woman as a US founding father and Black and Asian individuals as Nazi-era German soldiers. These images fueled claims of bias and were widely criticised on social media, attracting the attention of high-profile figures, including X (formerly Twitter) owner Elon Musk, who had previously voiced concerns about the perceived biases of AI tools. Google CEO Sundar Pichai admitted that some of the generated content was unacceptable and acknowledged that the company had “got it wrong.”
“We’ve upgraded our creative image generation capabilities, and, over the coming days, we’re bringing our latest image generation model, Imagen 3 to Gemini Apps and expanding its availability for users in all languages,” Dave Citron, a senior director at Google, stated in a blog post.
The new image generation model, Imagen 3, introduces several significant improvements. It offers enhanced image quality, producing highly detailed and accurate visuals based on textual prompts, and supports a diverse range of artistic styles. The model incorporates advanced safeguards to prevent the generation of inappropriate content, ensuring that it avoids creating images of identifiable individuals, minors, or scenes with excessive violence or sexual content. Google says that users have greater control over their creative process, with the ability to request modifications if the initial image does not meet their expectations.
