Google’s New Gmail Tool is Hallucinating Emails that don’t Exist
This article is authored by Maggie Harrison for Futurism.
Google’s new Bard extension will apparently summarize emails, plan your travels, and — oh, yeah — fabricate emails that you never actually sent.
Last week, Google plugged its large language model-powered chatbot called Bard into a bevy of Google products including Gmail, Google Drive, Google Docs, Google Maps, and the Google-owned YouTube, among other apps and services. While it’s understandable that Google would want to marry its newer generative AI efforts with its already-established product lineup, it seems that Google might have moved a little too fast.
According to New York Times columnist Kevin Roose, Bard isn’t the helpful inbox assistant that Google apparently wants it to be — at least yet. In his testing, says Roose, the AI hallucinated entire email correspondences that never took place.
It’s a step above Bing’s AI telling Roose that he should leave his wifeearlier this year, but still, even though Google claims to still be ironing out some bugs, it’s not exactly a promising first step.
Per the column, the hallucinations started when Roose asked Bard to “analyze all of my Gmail and tell me, with reasonable certainty, what my biggest psychological issues are.” While it’s a bit of an odd ask, it’s a simple enough one. Bard got to work quickly, reportedly telling Roose that he tends to “worry about the future,” citing an email, purportedly sent by the writer, in which Roose expressed that he was “stressed about work” and “afraid of failing.”
The problem? Roose never sent that email. The AI had found a newsletter that the writer had received — about the new “Elon Musk” biography, of all things — and misconstrued a quote within that newsletter to draft an entirely new email, which Bard claimed Roose had sent himself.
According to the columnist, Bard repeated the behavior, pulling an email in which Roose had allegedly complained that he was not “cut out to be a successful investor” out of thin air. The AI also repeatedly got airline information wrong and even made up a train that didn’t exist.
In response to Roose’s concerns, Jack Krawczyk, the director of Bard at Google, held the line that Bard Extensions is still experimental, and that this is the first iteration of the product.
Please click on this link to read the full article.
Image credit: Photo by Solen Feyissa on Unsplash