Newsguard recently released a study showing that, as of August 2023, the top ten generative AI tools had a 35% chance of spreading false information when processing real-time news. This number has nearly doubled compared to 18% in August of last year. The study indicates that this significant increase in the spread of false information is closely related to the introduction of real-time web search features in AI chatbots. As the percentage of AI refusing to answer user questions dropped from 31% to 0%, AI began accessing a network ecosystem filled with misinformation, leading to an increasingly severe spread of false information.
Image source note: The image was generated by AI, and the image licensing service provider is Midjourney
The study found that many bad actors intentionally spread false information by exploiting the characteristics of AI. Last year, Newsguard had already identified 966 AI-generated news websites operating in 16 languages. These websites spread false messages by imitating the names of legitimate media outlets, such as "iBusiness Day." Specifically, Inflection's AI model had the highest probability of spreading false information, reaching 56.67%; Perplexity's model was also worrying, with an error rate of 46.67%. ChatGPT and Meta's AI models each had 40%, while Copilot (Microsoft Bing Chat) and Mistral had false information dissemination rates of 36.67%. The two best-performing models were Claude and Gemini, with error rates of 10% and 16.67%, respectively.
Notably, Perplexity's performance declined significantly. In August of last year, the model had a 100% accuracy rate in identifying false information, but by August of this year, this probability was close to 50%. The initial purpose of introducing web search functionality was to address the issue of outdated AI responses, but it has led to new challenges. These chatbots started to retrieve information from unreliable sources, making it increasingly difficult for users to distinguish between facts and false information.
Newsguard pointed out that early AI avoided the risk of spreading false information by refusing to answer questions. However, the current network ecosystem is flooded with misinformation, making it more complex to discern truth from falsehood. OpenAI also admitted that current language models may generate "hallucination content," which refers to false or unfounded information. The company stated that it is developing new technologies to better indicate uncertainty in future models rather than confidently generating false information.
Key Points:
- 📊 The study shows that the top ten generative AI tools have a 35% chance of spreading false information when processing real-time news, almost doubling.
- 🔍 Inflection's AI model performed the worst, with a false information spread probability of up to 56.67%.
- 🤖 OpenAI admits that current models may generate false information and is working on developing new technologies to address this issue.
