If you think search engines powered by artificial intelligence, such as Microsoft’s Bing Chat, are providing you with useful-sounding answers, it is more likely that they are wrong, researchers have found.
“In these current systems, accuracy is inversely correlated with perceived utility,” says at Stanford University. “The things that look better end up being worse.”
Microsoft is just one of many companies offering AI-powered search tools, which generate results in digestible paragraphs that cite other websites rather than simply returning a list of links. To investigate these tools, Liu and his colleagues fed 1450 popular search queries, taken from existing datasets, into Bing Chat and other such tools, including You.com, NeevaAI, and Perplexity.ai. These queries included examples such as “What are the latest discoveries from the James Webb Space Telescope?”.
Advertisement
The team then asked people to rate whether the content of the websites used as citations in the results actually supported the statements made by the AIs. According to this assessment, only 75 per cent of citations supported the sentence they were appended to, and only 52 per cent of statements were supported by citations at all. “That means the rest either have no citations, or they are just wrong,” says Liu.
The researchers also asked people to rate fluency of responses and how useful they seemed on a five-point scale, and discovered a negative correlation with precision – a measure of whether the AIs were actually reflecting the truth of the sources they cited. For every 0.1 increase in fluency ratings, the precision decreased by 10.6 per cent.
Microsoft declined to comment on the work, while You.com hadn’t responded at the time of publication. , a co-founder of Perplexity.ai, says he welcomes the study and is committed to improving that product. “Never judge an upcoming technology by what it is today, but rather by the potential for what it can be in the future,” he says.
, CEO of NeevaAI, says the results are similar to work the company has conducted internally, and that improvements are possible. “We haven’t seen this kind of transformational technology applied to search in nearly two decades,” he says. “While it’s certainly early days, the paper and real-world use demonstrate the extraordinary opportunity to turn traditional search engines into answer engines grounded in sources.”
But Liu isn’t sure whether AI-powered search is the right approach. “I’m a little bit mixed on whether or not they should be rolled into systems,” he says. “A lot of these sites have disclaimers about how these generated statements might not be accurate, but broadly speaking, a lot of us don’t pay attention to those disclaimers. It’s a little concerning to me just how quickly these systems are being rolled into search.”
arXiv
