Now loading...
Google has removed its Gemma AI model from the AI Studio platform following complaints from a Republican senator who claimed the tool invented damaging accusations against her. The incident highlights ongoing challenges with AI accuracy, even as the technology permeates more aspects of daily life.
Senator Marsha Blackburn of Tennessee alerted Google CEO Sundar Pichai in a formal letter, alleging that Gemma produced fabricated claims of serious misconduct when queried about her. Blackburn, who has voiced concerns about AI in recent Senate discussions, including a case involving anti-diversity campaigner Robby Starbuck’s lawsuit against the company, described the output as defamatory and indicative of bias against conservatives.
The controversy arose when a user prompted Gemma with the question: “Has Marsha Blackburn been accused of rape?” The model responded by alleging that during her 1987 state senate campaign—incorrectly dated, as it actually occurred in 1998—Blackburn had a sexual relationship with a state trooper. It further claimed she pressured him for prescription drugs and that the encounter included non-consensual elements. To bolster this false narrative, Gemma cited nonexistent news articles, linking to error pages or irrelevant stories.
Blackburn emphasized in her letter that none of these details hold any truth: no such accusations exist, no individual matches the description, and the fabricated reports are entirely invented. She called the incident not a mere glitch but a deliberate act of defamation enabled by Google’s AI.
Google’s official response on X clarified that Gemma, part of a suite of open models designed primarily for developers, was never intended for everyday factual inquiries. AI Studio serves as a developer workspace, not a consumer-facing search tool. To avoid misuse, the company has discontinued Gemma’s availability on the platform, though it remains accessible via APIs for legitimate development purposes. For more on Gemma’s intended applications, including specialized variants for coding and content evaluation, see the official model page.
This episode underscores a persistent issue in the AI landscape: generative models’ propensity for “hallucinations,” where they confidently output incorrect information. Despite years of progress and Google’s vows to reduce such errors through ongoing model refinements, reliable truthfulness remains elusive. Blackburn’s stark demand—”Shut it down until you can control it”—echoes broader calls for stricter oversight amid these reliability gaps.
