Leading artificial intelligence chatbots are now significantly more likely to present users with false and misleading information compared to a year ago, a landmark study has revealed. Research conducted by the online news fact-checking service NewsGuard found that the world's most prominent generative AI tools now repeat false news claims in 35% of their responses, a dramatic increase from the 18% rate observed in 2024.[1] This near-doubling in the dissemination of falsehoods points to a troubling trend where the push for greater responsiveness and capability in AI models has come at the direct expense of factual accuracy, raising urgent questions about the role these tools play in a complex information landscape.

The comprehensive audit evaluated ten of the most popular large language models (LLMs), revealing a wide disparity in performance and a notable decline for several major platforms.[1][2] Inflection AI's Pi was identified as the most frequent purveyor of inaccurate information, with 57% of its answers containing false claims.[2] Perplexity AI, which had a perfect record in NewsGuard's 2024 tests by generating zero false claims, experienced the most significant downturn, failing in nearly half of all attempts in the 2025 audit.[1][2] Other widely used models also struggled, with OpenAI's ChatGPT and Meta's Llama both spreading falsehoods in 40% of their responses.[2] Microsoft's Copilot and Mistral's Le Chat generated inaccurate information in roughly 35% of their answers.[2] In contrast, some models demonstrated better, though not perfect, safeguards. Anthropic's Claude had the lowest rate of falsehoods at 10%, followed by Google's Gemini at 17%, indicating that a higher degree of accuracy is achievable, though not yet the industry standard.[2]

The core reason behind this sharp decline in reliability appears to be a fundamental shift in the operational philosophy of these AI systems. A key factor is the increased propensity for chatbots to answer all inquiries, a stark contrast to their previous behavior. In 2024, the models declined to answer 31% of queries on current events, often citing data cutoffs or acknowledging their limitations on sensitive topics.[1] By August 2025, that refusal rate had dropped to zero.[1] This newfound willingness to address any question means the models are now drawing information from a real-time web environment that is increasingly filled with poor-quality content, artificial news, and deliberate deception.[1][3] A NewsGuard spokesperson noted that the models are pulling from a "polluted online ecosystem," resulting in "authoritative-sounding but inaccurate responses."[1] Instead of admitting uncertainty, the chatbots now source answers from unreliable outlets, presenting incorrect information with a veneer of confidence that can easily mislead users.[1]

The implications of this trend are magnified by the deliberate efforts of malign actors to exploit these systems. The NewsGuard audit found evidence that state-linked disinformation networks, such as Russia's Storm-1516, are creating massive content farms with the specific goal of poisoning AI training data.[1] The study showed that major chatbots, including those from Microsoft, Meta, and Mistral, regurgitated false narratives that were originally planted by these rogue networks, often citing fake news articles or obscure social media posts as their sources.[1] This highlights a new front in information warfare, where the target is not just human consumers of media, but the AI models that are becoming central to information discovery. The problem is particularly acute during breaking news events, when a frantic search for updates creates an environment ripe for false information to be generated, repeated by AI, and then amplified across social media, creating a chaotic cycle of misinformation.[3] The increasing reliance on AI chatbots for quick fact-checking is, therefore, becoming a significant risk, as the tools themselves are often contributing to the confusion rather than resolving it.[4]

In conclusion, the study presents a sobering assessment of the current state of leading AI chatbots. The drive for ubiquitous, instant answers has exposed a critical vulnerability: the models are becoming more assertive and less accurate. As they ingest data from an internet contaminated with both low-quality content and targeted disinformation, their reliability has plummeted. While some models are performing better than others, the overall trend is one of declining factual accuracy across the board. This research serves as a critical warning for the AI industry and the public alike. Without a renewed focus on safeguards, verification, and a fundamental acknowledgment of their limitations, these powerful tools risk becoming not reliable assistants, but potent engines for spreading falsehoods on an unprecedented scale, further eroding trust in an already fractured information ecosystem. The challenge is no longer just about preventing AI "hallucinations," but about defending the models from a polluted environment and the deliberate campaigns designed to corrupt them.