Google and Meta should consider providing insensitivity training for their artificial intelligence models.
Both companies are facing criticism after unveiling their most recent AI language models, with many believing that the tech giants may have designed their models to be overly sensitive.
A few weeks back, Google rolled out an update to its Gemini web app that introduced an image generation tool.
Soon after, users began to notice an issue: the image generator was refusing to create specific content.
One instance involved a user requesting the generator to produce an image of a white cowboy.
However, Gemini declined, stating that emphasizing a character’s ethnicity could perpetuate negative stereotypes and assumptions.
After the user requested a picture of a black cowboy, the generator promptly generated the images.
AI Glitches Hit Google and Meta
Apologizing for the issue in a blog post, Google’s senior vice president mentioned that the model in Gemini became overly cautious, leading to misinterpretation of prompts.
At the moment, the image generator is not accessible while Google is working on a solution.
At Meta, developers are currently working on resolving a similar issue.
Researchers and senior leadership at the company believe that Llama 2, the company’s flagship AI model, may not be effective when handling sensitive topics, as per a report from The Information.
The design of Llama 2 was focused on avoiding controversial topics, potentially limiting its capability to address various inquiries.
An article by The Information mentioned that Meta is planning to assign an internal team member to focus on tone and safety training, aiming to improve responses before the release of Llama 3 in July.
