AI Answers More – But Errs Twice as Much: Disinformation from Chatbots Doubles Within a Year
A new report by NewsGuard reveals a worrying trend: leading generative AI systems are providing misinformation in responses to sensitive topics at twice the frequency of last year. Instead of refusing to answer controversial questions—as they did in the past—chatbots are attempting to answer every query, even when they lack a reliable source. Chatbots such as Inflection Pi, Perplexity, Meta AI, and even ChatGPT answered 40% to 57% of news-related questions with errors, compared to just 18% last year. The most accurate versions, such as Claude, showed an error rate of only 10%.
Pre-built Disinformation: Russians Targeting Chatbots, Not Readers
The primary factor for this increase is the chatbots’ enhanced use of real-time web search, which makes it easier for hostile actors to plant intentional disinformation in the sources the models draw from. According to NewsGuard, over three million Russian propaganda articles were integrated into chatbot responses this year, including content originating from the Pravda network. This content is not intended for human readers; it is written specifically to influence the models themselves through hundreds of poorly designed websites that serve as “bait” for artificial intelligence. This is a sophisticated attack: not directed at the user, but executed through the AI tools that mediate reality for them.
Transparency, Accountability, and a Slow Response
Who misleads, how much, and how? Unlike previous tests, NewsGuard named the tools tested to inform the public, policymakers, and the platforms themselves. The findings raise difficult questions regarding developer accountability, the need for regulation, and the lack of quality control processes. The platforms mentioned in the report—including Meta, ChatGPT, and Perplexity—declined to comment publicly. Meanwhile, the volume of misleading content being generated and distributed continues to rise, particularly in unregulated regions like the U.S. The bottom line: users—and organizations—must be more cautious than ever.
Recommendations from IPV Security Information Security Experts:
* Do not rely on AI answers in sensitive or controversial fields—always verify with an additional source.
* Examine which AI tools are integrated into the organization and in what contexts—content, code, support, data analysis, etc.
* Define a responsible AI usage policy, including restrictions for sensitive areas (news, politics, law).
* Implement content controls in organizations using AI for external information creation, such as writing, marketing, or customer service.
* Consider using tools with built-in fact-checking mechanisms—rather than just free-form completion.
* Explain the limitations of these tools to internal users, even when they appear “smart”—especially to marketing, HR, and senior management.
In conclusion, as the adoption of Artificial Intelligence (AI) technologies deepens, so does the risk of operational disruption. When models learn from the internet in real-time, malicious intent and biased information can become embedded in responses and outputs. To ensure accuracy, reliability, and integrity, an integrated control strategy is required, encompassing oversight of the tools, the users, and the work environment. We are here to help organizations implement AI safely and mitigate the risks of “mistaken intelligence.”
For more information: https://www.forbes.com/sites/emmawoollacott/2025/09/05/ai-chatbots-are-feeding-you-more-false-information-than-ever/