Highlights:

  • In response to ChatGPT, Bard launched Gemini last year. After several feature updates, Google renamed the chatbot on Feb. 8. This update introduced an image generation tool, which led to the Gemini controversy.
  • Gemini’s image generation capabilities are based on a large language model family of the same name developed by the search giant’s Google DeepMind unit.

Google LLC Chief Executive Officer Sundar Pichai has informed employees that the company will implement “structural changes” to rectify the errors in Gemini’s image generation feature.

Pichai outlined the plan in an internal memo distributed earlier, as reported by Semafor very recently. The executive also condemned the error, which resulted in Gemini producing misleading images of racially diverse people, as “completely unacceptable.” Google verified the memo.

Gemini is the new name of the Bard chatbot, which the company launched last year as a response to ChatGPT. Google renamed the chatbot on Feb. 8 following a series of feature updates rolled out over the preceding days. One of those updates introduced an image generation tool, which laid the groundwork for the recent Gemini controversy.

Approximately a week ago, users of X started noticing that some of the images generated by the chatbot inaccurately portrayed racially diverse characters in historical settings.
In response, Google temporarily disabled Gemini’s capability to generate images of people. Prabhakar Raghavan, the company’s Senior Vice President of knowledge and information, announced earlier that the feature will be reactivated once the issue is resolved.

Pichai gave more information about the troubleshooting effort in the memo sent to staff members. Pichai wrote, “Our teams have been working around the clock to address these issues. We’re already seeing a substantial improvement on a wide range of prompts.”

According to sources familiar with the matter, Bloomberg reported that Google only equipped Gemini with the ability to generate images of people after undergoing multiple internal reviews. However, the massive search engine didn’t test “all the ways that the feature might deliver unexpected results.”

Pichai clarified that Google will tackle the problem in more ways than technical troubleshooting. In the memo from this week, he stated that the business plans to make “structural changes” in addition to launching “updated product guidelines, improved launch processes, robust evals,” and additional improvements.

The effort will also prioritize red teaming. That’s a practice in which a company’s engineers seek out user output, such as chatbot prompts, that could potentially lead to application malfunctions. Last July, Google revealed that it had established a team dedicated to identifying issues in its AI products.

Gemini’s image generation capabilities are based on a large language model family of the same name developed by the search giant’s Google DeepMind unit. The unit also created the Imagen 2 model, which powers the chatbot’s image-generation features in Gemini. Demis Hassabis, DeepMind’s CEO, revealed earlier that Gemini’s ability to generate images of people could be reactivated within a few weeks.