Google wants everyone to know that its Bard AI is definitely not trained on private Gmail data, despite Bard telling people that it is.
After teasing its AI chatbot earlier this year – to rival Microsoft’s ChatGPT-powered Bing Chat service – Google has recently unleashed Bard on the world. However, the beta program is still prone to making mistakes, including one recent gaff where Bard said it was trained using data from a variety of sources like Wikipedia, GitHub, and (worryingly) Gmail.
This information was shared by Kate Crawford (opens in new tab) (an AI expert) on Twitter via a post showing a Bard conversation in which the program was asked “where does Bard’s dataset come from.” In its response, Bard said it uses “Google’s internal data: This includes data from Google Search, Gmail, and other products.”
However, the official Google Twitter account (opens in new tab) corrected this error, saying in a reply to Crawford’s tweet that “Bard is an early experiment based on Large Language Models and will make mistakes. It is not trained on Gmail data.”
Umm, anyone a little concerned that Bard is saying its training dataset includes… Gmail? I’m assuming that’s flat out wrong, otherwise Google is crossing some serious legal boundaries. pic.twitter.com/0muhrFeZEAMarch 21, 2023
Don’t trust everything an AI says
Another day, another reminder that you can’t believe everything an AI tells you. Because as impressive as Bard, Bing Chat, and the best ChatGPT alternatives are, they’re also fallible.
Yes, they’re trained on a huge swathe of human data, but they don’t have a proper understanding of what any of that information really means. So, when you ask a question, it’s merely spitting out its best guess at an answer.
What’s more, Bard does seem particularly easy to trick. Michael Allison, one of our Phones team’s Staff Writers, managed to get Bard to say eggs didn’t exist (opens in new tab), and our friends at Tom’s Hardware found that Bard was a bit too quick to claim Tom’s Hardware’s CPU tests as its own (opens in new tab) – though it did then apologize.
I told Bard there were no such things as eggs and it agreed with me. Eggs don’t exist.March 21, 2023
As for why Bard would get information about itself wrong, most likely it hasn’t been told (or it’s been told this info should be kept private at all costs) to avoid it spilling its secret rules as Bing Chat did. If users find out how an AI works behind the scenes, it becomes much easier to manipulate it and use the bot for nefarious purposes. So, when asked about where its data set was gathered, Bard seemingly spat out an answer that combined what it knows about other AI and the kind of data Google could access.
If you want to try tricking Google Bard for yourself you’ll have to join the waitlist, but our guide on how to sign up for Google Bard should help you get through the steps super quickly.
email@example.com (Hamish Hector)