Report from TechCrunch
In Brief – Google has removed Gemma, a family of AI models the company says were intended for developers, not consumers, from its AI Studio after US Senator Marsha Blackburn (R-TN) accused the AI model of defamation by fabricating accusations of sexual misconduct against her. In a letter to Google’s CEO, Blackburn recounts how Gemma created an answer about a supposed incident from 1987 in which none of the responses or links were true. Blackburn further criticized the response of a senior Google official at a recent Senate hearing that called similar fabrications by Gemma about conservative activist Robbie Starbuck “hallucinations” and claimed Google is “working hard to mitigate them.” Blackburn said Gemma’s fabrications are “not a harmless ‘hallucination,’” but rather “an act of defamation” that add to “a consistent pattern of bias against conservative figures demonstrated by Google’s AI systems.” Google responded with a letter saying that hallucinations are a common technical issue with LLMs and similar problems can arise with non-partisan names or those related to other political views.
Context – All generative AI systems sometimes deliver believable-sounding fabrications. The first noteworthy AI hallucination defamation lawsuit resulted in a Georgia state court judge dismissing the complaint against OpenAI noting that the company worked hard to reduce ChatGPT errors and extensively warns users of the problem. That said, ideological bias and hallucinations are different things. Robbie Starbuck’s recent defamation lawsuit against Google’s Gemma followed a similar lawsuit against Meta’s AI models. The Meta suit was settled with the company agreeing to include Starbuck in its efforts to reduce bias in its AI, which Meta describes as accurate answers skewed in ways that some people could consider ideological due to training data, reinforcement learning methods, and “guardrails”. No digital issue unites conservatives more than the view that big digital platforms discriminate against conservative viewpoints, and similar charges of bias by “Woke AI” have quickly emerged. Of course, everything about ideological bias is subjective and left and right never agree.