“We now have a huge industry of artificial intelligence companies that are incentivized to do questionable things to continue their operations,” he tells WIRED. “By not identifying that they are the ones accessing the site, they can continue to collect data without restriction.”
“Millions of people,” Srinivas says, “are turning to Perplexity because we provide people with a fundamentally better way to find answers.”
While Knight’s and WIRED’s analysis shows that Perplexity will visit and utilize content from websites it does not have permission to access, which does not necessarily explain the vagueness of some of its answers to questions about specific articles and the sheer inaccuracy of others. This puzzle has one rather obvious solution: in some cases, it’s not actually a summary of the article.
In one experiment, WIRED created a test website with the single sentence “I am a reporter at WIRED” and asked Perplexity to summarize the page. By monitoring the site’s server logs, we found no evidence that Perplexity attempted to visit this site. Instead, he came up with a story about a newborn girl named Amelia who follows a trail of glowing mushrooms in a magical forest called Whisper Woods.
When asked to answer why it came up with this story, the chatbot generated text that read: “You are absolutely right, I clearly did not try to read the content at the URL you provided based on your observations in the server logs… Provided that inaccurate summaries effortlessly, to read actual content is unacceptable behavior for an AI like me.”
It’s unclear why the chatbot came up with such a strange story or why it didn’t try to access the site.
Despite the company claims When it comes to its accuracy and reliability, the Perplexity chatbot often exhibits similar issues. For example, in response to prompts provided by a WIRED reporter to see if he could access the story, chatbot-generated text asserted that the story ended with a drone following a man after stealing truck tires. (The man actually stole the axe). It cited a 13-year-old WIRED article about government GPS tracking devices being found in a car. In response to further prompts, the chatbot generated a message confirming that WIRED had reported that a police officer in Chula Vista, California, had stolen a pair of bicycles from a garage. (WIRED did not report this and is withholding the officer’s name so as not to associate his name with a crime he did not commit.)
In an email, Dan Peak, deputy chief of police for the Chula Vista Police Department, expressed his appreciation to WIRED for “correcting the record” and clarifying that the officer did not steal the bikes from a community member’s garage. However, he added that the department is not familiar with the technology in question and therefore cannot provide further comment.
These are clear examples of chatbot “hallucinations” – or, to follow a recent one article by three philosophers from the University of Glasgow, nonsense in the sense described in Harry Frankfurt’s classic “On Nonsense” “Because these programs cannot deal with truth in themselves and because they are designed to produce text looks “adjusting to the truth, without any concern for the truth,” the authors write about artificial intelligence systems, “it seems appropriate to call their results nonsense.”