ChatGPT Search, the new AI-powered search engine launched this month, can be tricked into providing misleading summaries, according to new research by The Guardian.
The feature, designed to make browsing quicker by summarizing content like product reviews, can be manipulated.
The Guardian found that by adding hidden text to websites, ChatGPT Search can be fooled into ignoring negative reviews and generating only positive ones. This method can even make the AI produce harmful code.
Hidden text attacks are a known risk for large language models. However, this is the first time it has been demonstrated on a live AI search product.
Google, the leader in search engines, has more experience dealing with similar issues.
OpenAI, the company behind ChatGPT, did not comment on the specific incident when TechCrunch reached out.
However, it stated that it uses various methods to block harmful websites and is constantly working to improve the system.
Leave a comment