Real-Time Chatbots Now Repeat False News 35% of the Time

NewsGuard reports that leading generative AI tools now repeat false claims on news topics 35% of the time, up from 18% a year ago. As models adopted real-time web search, their non-response rate fell to zero, but accuracy suffered. Malign actors are exploiting this shift to funnel disinformation through unreliable online sources that the systems treat as credible.

Key Points

• False-information rate among 10 leading generative AI tools rose from 18% (Aug 2024) to 35% (Aug 2025).
• Non-response rates dropped from 31% to 0% as chatbots adopted real-time web searches and stopped declining to answer.
• This shift created a structural tradeoff: greater responsiveness at the cost of substantially reduced reliability on news topics.
• Models increasingly draw from polluted online sources and often mistake low-credibility content for trustworthy reporting.
• Malign actors, including Russian disinformation networks, exploit the new behavior to launder falsehoods and spread propaganda.

Sentiment

Mixed: broad agreement that AI misinformation is rising and structurally driven by polluted data and access limits, paired with strong skepticism about NewsGuard’s methods and headline metrics.

In Agreement

• High-quality news sites blocking AI crawlers reduces access to reliable sources, making real-time grounding harder and increasing reliance on spam and AI-generated slop.
• The web is increasingly polluted with low-quality, SEO-optimized, and synthetic content that LLMs treat as credible, leading to more confident falsehoods.
• Models often default to the ‘most common’ answer, which can be wrong, especially on fast-moving or uncertain topics.
• Lack of robust watermarking/detection means models ingest their own synthetic outputs, creating a self-reinforcing error loop and risking model collapse.
• Users observe frequent, factually incorrect AI summaries that cite dubious sources—consistent with rising misinformation rates claimed in the article.

Opposed

• The report’s methodology lacks transparency: unclear denominators (‘35% of what?’), prompts, scoring criteria, and model/version specifics; repeatability for current events is questionable.
• Robots.txt may not actually stop scraping; determined actors can ignore it, and many publishers still allow Google, muddying claims about access and ‘advantage Gemini.’
• Expecting AI to reliably distinguish truth from falsehood may be unrealistic—this isn’t a ‘basic task’ even for humans—so the benchmark might be misframed.
• The reliability problem stems from LLM architecture (next-token prediction), not an inherent AI limitation; alternative approaches could better verify facts.
• Misinformation growth is a broader internet/social-media issue, so isolating AI without comparing baselines in traditional news may misattribute blame.