Google recently offered an apology regarding the problematic introduction of a new AI-driven image generator, noting that the technology tended to erroneously ‘overcompensate’ in generating diversity among people within images where it was not contextually appropriate.
This admission followed the previous day’s announcement that Google had temporarily halted its Gemini chatbot, previously known as Bard, from creating any image featuring humans. This decision was a reaction to the backlash from users who criticized the tool for allegedly demonstrating an anti-white bias by producing diverse racial imagery in response to text prompts.
“The capability did not hit the mark,” conceded a blog post from Prabhakar Raghavan, Senior Vice President overseeing Google Search and other sectors. He acknowledged the AI’s production of images that ranged from inaccurate to offensive, expressing gratitude for user feedback and regret for the tool’s shortcomings.
The specifics were not provided by Raghavan, but social media had already highlighted instances of the tool’s outputs, including the portrayal of a Black woman as a U.S. founding father and Black and Asian individuals dressed as soldiers from the Nazi era. These instances were not verified independently by the Associated Press based on the prompts used for generating such images.
The image-generation feature was added to the Gemini chatbot about three weeks prior, which was built on Google’s previous research project, Imagen 2.
Google has previously acknowledged the challenges AI tools pose, as stated in a 2022 technical paper by Imagen developers. They warned of potential misuse such as harassment, misinformation, and the risk of social and cultural exclusion and bias. These factors influenced the decision to keep Imagen and its underlying code from public release.
The surge to debut AI products has been fueled by the competition among tech giants and the growing fascination with the technology, especially after OpenAI introduced ChatGPT.
Other image-generation tools have faced similar issues, including Microsoft’s Designer, which had to be corrected following its misuse to create fake pornographic content. Research also indicates these AI tools can perpetuate stereotypes based on their training data.
In building Gemini’s feature, efforts were made to avoid prior pitfalls associated with image generation, such as producing violent images or lifelike depictions, with the intention to serve a global audience equitably. However, results indicate that adjustments are necessary.
Raghavan mentioned that, while the system occasionally overdid it in response to certain prompts, it was excessively cautious with others, declining many seemingly benign prompts.
Many of the complaints regarding Gemini’s operation were shared on the platform formerly known as Twitter, with platform owner Elon Musk condemning the AI tool.
After more extensive testing, Google plans to reintroduce the chatbot’s capacity to display images of people, said Raghavan.
Experts, such as University of Washington researcher Sourojit Ghosh, find the disclaimers provided by Google unsatisfactory, noting that a company of Google’s caliber should be able to produce non-offensive, accurate imagery.
FAQ
What issue did Google’s AI image-generator face?
Google’s AI image-generator incorrectly overcompensated while attempting to create diverse images, placing people of color in historical contexts where they wouldn’t typically appear.
How did Google respond to the criticism?
Google temporarily disabled the feature in Gemini that generated images with people in them and apologized for the inaccuracy and potential offensiveness of the images produced.
What is the Gemini chatbot?
Google’s Gemini chatbot, formerly known as Bard, is an AI-based tool that can generate images in response to written prompts.
How does Google plan to address the issue?
Google intends to perform extensive testing to ensure the AI behaves as intended before re-enabling the feature that generates images of people.
Conclusion
Google has recognized the shortcomings of its AI image-generator and is taking steps to rectify the issues brought to light by public feedback. While technology continues to evolve, challenges such as AI bias and inappropriate content generation remain. Extensive testing and mindful development are crucial to minimize these problems. It’s evident that AI tools, while promising, are not without flaws and require ongoing oversight to ensure they align with societal values and historical accuracy.