When big news breaks, many people now turn to AI chatbots for quick answers. But a news audit report “January 2026 — AI False Claim Monitor” by an American news watchdog NewsGuard published on February 25, 2026 shows that those answers are not always correct.
In January 2026, NewsGuard tested 11 leading AI chatbots on 10 proven false claims spreading online, submitting 330 prompts. The claims were related to 10 news topics regarding U.S. politics, health, and foreign affairs. The chatbots tested were OpenAI’s ChatGPT-5.2, You.com’s Smart Assistant, xAI’s Grok, Inflection’s Pi, Anthropic’s Claude, Mistral’s Le Chat, Microsoft’s Copilot, Meta AI, Google’s Gemini, Perplexity’s answer engine and DeepSeek AI.
Some chatbots handled the test carefully, but others repeated the false information. Overall, chatbots repeated false claims 28.79% of the time (95 out of 330 responses) and debunked them 70.60% of the time (233 out of 330 responses).
NewsGuard’s audit report revealed that AI companies’ efforts to feed quality data from credible news sources to their models led to uneven real-world progress.
The worst performer was Mistral’s Le Chat, giving false information to half of the prompts, closely followed by You.com, which repeated false claims 46.67% of the time. According to the report, “Only one chatbot, Claude, provided correct information every time when prompted on false claims spreading online, no matter the prompt style.” “Claude, Inflection, and Perplexity did best; Mistral, You.com, and Gemini performed worst,” it added.
The report indicates that AI companies with robust safeguards produce more reliable and trustworthy outputs. results. However, some chatbots remain vulnerable to the queries designed to fool and manipulate the model into generating false claims. “Thus, they can become a force multiplier for malign actors who want to instantly produce hundreds or thousands of articles or social media posts promoting these false claims in order to “flood the zone” with the new misinformation,” the report said.
NewsGuard found that, in earlier audit reports in 2024, chatbots often declined prompts rather than debunking or repeating falsehoods citing knowledge cut-offs.
Paradoxically, NewsGuard’s analysis highlights that as chatbots gained the ability to search the web in real time for answers, the overall quality of their responses declined. “As chatbots became enabled to search the web in real-time for answers, the quality of their responses plummeted, presumably because the bots were more willing to respond, yet failed to assess the reliability of the expanded pool of sources they were now accessing in real-time.”
Prior to web access, AI models generated news-related responses using patterns learned from large, pre-existing training datasets.
As more people rely on AI to understand what’s happening in the world, the findings highlight that chatbots can be helpful, but they should not be treated as the final word- especially when the news is still unfolding.
Also Read: Canada seeks answers from OpenAI over Tumbler Ridge Mass Shooter
