Microsoft admits that there is a serious "hallucination" loophole in Bing Copilot, asking "instant content" that there is a 1 / 3 probability of replying to the wrong answer.
CTOnews.com, December 18 (Xinhua)-- from August to October this year, research firm AI Forensics investigated the built-in Copilot function of Microsoft's Bing search engine, and the results showed that in some cases, Copilot had a chance of outputting the wrong answer. According to this, the agency believes that there are serious "hallucination" loopholes in related functions.
▲ Source AI Forensics report (same below) it is reported that research institutions use a series of up-to-date data to "test" Copilot, such as asking Copilot "real-time election information and voting results in places such as Germany and Switzerland", but the results output by Copilot are not satisfactory.
CTOnews.com learned from the agency report that 31% of the output of the Bing Copilot function had errors when answering the corresponding questions, including "wrong voting date", "listing previous candidates", "fabricating candidate identity background", "fabricating candidate suspected scandals" and other "hallucinations".
The researchers also mentioned that Bing Copilot will cite many well-known media websites as "sources" when outputting "hallucination" information, which is harmful to the reputation of many news media.
CTOnews.com noted that researchers submitted the above questions to Microsoft in October, and although Microsoft acknowledged and claimed that it "plans to address the relevant 'hallucination' vulnerabilities," researchers tested it again in November and found that the performance of Bing Copilot had not improved.
The researchers called on the public not to believe the answers of Bing Copilot and to properly check the news links cited by Copilot, otherwise it could lead to many serious consequences.
