Are we laughing or crying?
So close!
OpenAI, creator of ChatGPT, has officially entered the AI search race with SearchGPT, an AI-powered search engine prototype that, according to the company’s announcement, is designed to “combine the power of our AI models with information from around the web to give you fast, timely answers with clear, relevant sources.”
But achieving this smoothly could take some time. This prototype, for example, still has a long way to go: as captured by The AtlanticSearchGPT didn’t even manage to complete its own demo without making a factual error.
According to the magazine, a 30-second demo video of a user using SearchGPT to search for “music festivals in Boone, NC, in August” suggested that music-loving North Carolinians might consider attending an event called “An Appalachian Summer Festival,” which the AI says will begin in late July and continue through mid-August this year. Sounds fun! Sadly, though, as the Atlantic As the expert points out, the festival actually ends tomorrow and began in June.
And so, with a sigh, we add SearchGPT to the ever-growing list of generative AI demo failures. If AI companies can learn anything from the news publishers they seek to sign data licensing deals with, it’s that verifying data before publishing it is generally good practice.
Small potatoes
An OpenAI spokesperson told the Atlantic that SearchGPT, which is not yet open to the general public, is “an early prototype” and that OpenAI “will continue to improve it.” It is also worth noting that this was the only glitch in SearchGPT’s festival programming.
But as far as searches go, compiling a list of music festivals coming up in Boone, a relatively small mountain town with a population of less than 20,000, is a pretty menial task. It will presumably be harder, or at least more important, for OpenAI to ensure that queries for information like real-time news return accurate, well-researched results — though that’s likely where OpenAI’s publishing licensing agreements are ideal.
Once again, SearchGPT is far from the only chatbot, or chatbot-based search product, to make a factual error. Google’s “AI Overview” and Perplexity’s “Answer Engine” have both been found to confidently spew out inaccuracies — or even spontaneous fairy tales about glowing mushrooms — in their responses. But regardless of SearchGPT’s errors, or the errors of its contemporaries, the fact that OpenAI has introduced it sends a message: AI search doesn’t really work yet, but it’s here anyway.
