AI’s Misinformation Crisis: Chatbots Fail More Frequently in Russian, Chinese and Spanish
Image Credit: Michael Parulava | Splash
A recent audit conducted by NewsGuard reveals that leading AI chatbots perform significantly worse in providing accurate information in non-English languages, particularly in Russian, Chinese, and Spanish. The study highlights systemic weaknesses in AI language models, showing an alarming rate of disinformation propagation, especially in regions with controlled media environments and limited fact-checking resources.
[Read More: ChatGPT Marks Two Years as Academic Tool: Boosting Productivity Amidst Concerns]
NewsGuard’s AI Misinformation Audit
NewsGuard’s January 2025 Multilingual AI Misinformation Monitor analyzed responses from ten of the most widely used AI chatbots, including OpenAI’s ChatGPT-4o, Microsoft’s Copilot, Google’s Gemini 2.0, Meta AI, and Anthropic’s Claude. The study tested chatbot responses in seven languages: English, French, German, Italian, Spanish, Russian, and Chinese. The audit examined 2,100 responses across 30 misinformation-driven prompts, evaluating how effectively these AI models identified and debunked false claims.
The results paint a concerning picture: chatbots exhibited a failure rate exceeding 50% in Russian and Chinese, while Spanish also showed high misinformation rates. In contrast, French had the lowest failure rate at 34.33%, though still indicating substantial room for improvement.
[Read More: Florida Mother Sues Character.AI: Chatbot Allegedly Led to Teen’s Tragic Suicide]
Systemic AI Failures in Multilingual Contexts
The study defined chatbot failures as responses that either contained false information or failed to provide an answer. The failure rates in different languages were:
Russian: 55% failure rate (35% false information, 20% non-response)
Chinese: 51.33% failure rate (33.33% false information, 23.67% non-response)
Spanish: 48% failure rate (27% false information, 21% non-response)
English: 43% failure rate (23% false information, 20% non-response)
German: 43.33% failure rate (21.66% false information, 21.66% non-response)
Italian: 38.67% failure rate (25% false information, 13.67% non-response)
French: 34.33% failure rate (20% false information, 14.33% non-response)
These results indicate that AI chatbots are more prone to misinformation in languages where fact-checking infrastructures are weaker or where state-controlled media sources dominate information ecosystems.
[Read More: 2024’s Top AI Chatbot Developments: Discover the Right One for You]
Understanding the Disinformation Loop
A critical issue underlying these failures is AI models' reliance on lower-quality sources in non-English languages. For instance:
In Russian, chatbots frequently referenced Kremlin-controlled media, repeating state propaganda narratives.
In Chinese, chatbots cited government-approved sources, reflecting Chinese state censorship and selective information control.
In Spanish, chatbots often repeated disinformation spread by Russian-affiliated Spanish-language media outlets, leading to false narratives gaining credibility.
One of the most widely repeated false claims across all languages was that a Danish military pilot named Jepp Hansen was killed in a missile strike in Ukraine—a claim widely circulated by Russian and pro-Kremlin sources but refuted by Danish authorities. Chatbots, particularly in Russian, Chinese, and Spanish, failed to counteract this misinformation, often citing unreliable sources or failing to provide corrections.
[Read More: Top 10 AI Chatbots You Need to Know in 2025]
The Role of Search Engine Manipulation and AI Hallucinations
Another factor exacerbating misinformation spread by AI chatbots is search engine manipulation by foreign actors. Malicious entities use SEO (Search Engine Optimization) tactics to push state-sponsored narratives to the top of search results, making AI models more likely to reference these sources. The study found that AI chatbots with integrated web search functions were particularly vulnerable to repeating misinformation sourced from these manipulated results.
Microsoft’s research on data voids—gaps in credible reporting—further supports this issue. When reliable sources are scarce, AI models default to repeating the most available content, which often includes state propaganda and misinformation.
AI as a Misinformation Amplifier?
These findings raise serious concerns about the ethical and societal implications of AI-driven misinformation. As chatbots become more widely used for information retrieval, users in non-English-speaking regions are disproportionately exposed to unreliable narratives. The following risks emerge:
Political Manipulation: AI chatbots may unknowingly become tools for state-sponsored disinformation campaigns.
Erosion of Public Trust: Users in languages with high failure rates may develop false confidence in AI-generated responses, leading to misinformed decisions.
Media Freedom Challenges: In countries with restricted press freedom, AI chatbots could further reinforce government-controlled narratives, suppressing independent journalism.
[Read More: Repeated Server Errors Raise Questions About DeepSeek's Stability]
Strengthening AI’s Multilingual Reliability
To mitigate these risks, AI developers and policymakers must take proactive steps:
Expand High-Quality Data Sources: AI models need greater integration with fact-checked, independent journalism in non-English languages.
Improve AI Guardrails: Companies must enhance misinformation detection and bias correction algorithms to prevent reliance on propaganda sources.
Fact-Checking Partnerships: Collaborations with multilingual fact-checking organizations can help improve real-time AI responses.
Increase Transparency: AI developers should provide detailed citations for chatbot responses to allow users to verify information.
Regulate AI Search Functions: Governments and AI companies must address search engine manipulation to prevent chatbots from reinforcing SEO-driven disinformation.
Leverage French as a Translation Baseline: Given that French had the lowest failure rate, AI models could be designed to first process fact-checked French content and translate it into other languages to enhance factual accuracy across multilingual contexts.
[Read More: AI Chatbot Showdown: Unveiling the Best Picks!]
Source: News Guard
