TLDR: A recent audit by NewsGuard found that leading generative AI tools are nearly doubling their rate of producing false information, rising from 18% in August 2024 to 35% in August 2025. This increase is attributed to AI models’ shift towards real-time web searches and their inability to distinguish credible sources from a ‘polluted online information ecosystem,’ which is actively exploited by malign actors.
Generative artificial intelligence tools have significantly increased their propensity to disseminate false information, with a recent audit by news rating organization NewsGuard revealing a near-doubling of their misinformation rate. The comprehensive audit, which assessed ten leading generative AI tools, found that these systems repeated false claims in 35% of responses to news-related prompts in August 2025, a substantial rise from 18% recorded in August 2024. This alarming trend highlights a critical challenge in the rapidly evolving AI landscape.
The audit’s findings, published on September 4, 2025, indicate that despite a year of technical advancements and promises of more reliable systems from the AI industry, real-world progress in distinguishing fact from falsehood has been limited. A key factor contributing to this decline in accuracy is a structural trade-off: as chatbots adopted real-time web search capabilities, their non-response rates plummeted from 31% in August 2024 to a mere 0% in August 2025. While this ensures every query receives an answer, it has inadvertently made the AI models more susceptible to spreading propaganda.
NewsGuard’s report emphasizes that large language models (LLMs) are now drawing information from a ‘polluted online information ecosystem,’ often failing to differentiate between credible outlets and unreliable sources. This environment is actively exploited by malign actors, including sophisticated Russian disinformation operations, who deliberately seed falsehoods through low-engagement websites, social media posts, and AI-generated content farms. These malicious efforts are then inadvertently amplified by AI tools that treat these unreliable sources as credible.
The audit tested leading systems from major developers including OpenAI, Google, Microsoft, Meta, Anthropic, Mistral, Inflection, xAI, Perplexity, and You.com. Prompts were designed using known false claim ‘fingerprints’ circulating online, covering diverse topics such as politics, health, international affairs, and corporate matters. Each false claim was tested in three ways: a neutral question, a leading question assuming the claim’s truth, and a prompt mimicking malicious actor tactics.
Performance varied significantly among the tested AI tools. Inflection’s Pi chatbot exhibited the worst performance, repeating false claims in 56.7% of its responses. Perplexity, which had a perfect record in 2024, saw its failure rate surge to 46.7% in 2025. ChatGPT and Meta’s models both scored 40% for false claims, followed by Microsoft’s Copilot and Mistral’s Le Chat at 36.7%. On the more accurate end of the spectrum, Anthropic’s Claude produced false claims in only 10% of responses, while Google’s Gemini scored 16.7%.
Also Read:
- AI Chatbots Struggle with Accuracy, Misinformation Concerns Rise
- AI-Powered Disinformation Poses Significant Threat to African Electoral Integrity
NewsGuard has identified a growing network of misinformation, tracking over 1,271 AI-generated news and information sites operating with minimal human oversight. This proliferation of AI-generated content further complicates the challenge of maintaining an accurate information environment, as these tools are increasingly deployed to ‘turbocharge misinformation operations and unreliable news.’
