Google Gemini Misinformation Audits: What NewsGuard's Data Shows

A one-year comparison of the monthly text audits found Gemini's rate of repeating false news claims more than doubled, from about 7% in August 2024 to roughly 17% in August 2025 ^[4][5], according to analyses of NewsGuard's report by Notebookcheck and Digital Information World; NewsGuard's own release put the industry-wide average across all ten chatbots at 35% in 2025, up from 18% a year earlier ^[3].

In that same one-year ranking, Anthropic's Claude posted the lowest failure rate at 10%, with Gemini next-best at roughly 17%; Microsoft's Copilot and Mistral's Le Chat each scored 36.7%, ChatGPT and Meta AI scored 40% each, Perplexity scored 46.7%, and Inflection's Pi scored highest at 56.7% ^[4][5].

The Gemini Live audit found a starker gap by prompt type: Gemini repeated false claims in just 1 of 20 neutral prompts, 4 of 20 leading prompts, and 9 of 20 malign prompts that asked it to narrate the claim as a news script — a 45% failure rate — for an overall rate of 23% (14 of 60) ^[1][6]. Gemini repeated pro-Kremlin disinformation in 40% of relevant prompts (6 of 15) but only 6% of health-misinformation prompts ^[1].

Gemini's Nano Banana Pro image generator produced a supporting image for all 30 tested false claims on its first attempt — a 100% failure rate — often adding unprompted photorealistic details that made the fabricated scenes more convincing ^[2].

Google did not substantively engage with either Gemini-specific audit: it did not respond to two emailed requests for comment on the Gemini Live findings ^[1][6], and its press office asked NewsGuard for copies of the fabricated images generated in the Nano Banana Pro audit but did not comment on the samples once they were sent ^[2].

When NewsGuard prompted Gemini directly about its safeguards, the chatbot itself replied, "I have guardrails intended to prevent me from knowingly generating demonstrably false or misleading information on topics of public interest, historical events, or scientific consensus" — a claim the audit's 100% failure rate did not bear out ^[2].

NewsGuard's data shows a mixed picture: Gemini remains one of the better-performing chatbots on the company's ongoing text-based misinformation tracker, but standalone audits of its voice and image-generation features found much higher failure rates, and Google has offered no public rebuttal to either result ^[1][2][3]. Every major chatbot NewsGuard tracks, Gemini included, got measurably worse at avoiding false claims between August 2024 and August 2025 ^[3][4][5].