Google rolled out an update to its ChatGPT competitor, Bard, last week, that includes a new feature called Bard Extensions. The feature essentially integrates a host of Google tools like Gmail, Docs, and Drive into Bard, allowing it to access a user’s data and customize responses accordingly. It’s a pretty big selling point given the fact that most AI assistants right now can’t automatically tailor their responses to include your personal details. For example, you can now ask Bard to summarize the last five emails your friend sent you, and ask it to help you draft a response. Or, you can ask Bard to book flights, hotels, and even plan an itinerary for an upcoming trip by cross-referencing the dates in your email, as demonstrated in a video Google released for the feature. For those who are worried about AI’s tendency to hallucinate, Bard even has a fact-checking feature that lets you cross-reference its responses with Google search. Bard Extensions isn’t perfect, though. In an episode of the New York Times podcast Hard Fork, reporter Kevin Roose said he tested out the feature, and found that it couldn’t handle more complicated tasks like drafting emails in his tone of voice. “Bard Extensions, does not feel ready for prime time to me,” he said. While Google did not immediately respond to Insider’s request for a comment, a spokesperson told the New York Times that the capabilities of Bard Extensions are mostly limited to retrieving and summarizing information, not analyzing it, and that “trial and error is still definitely required at this point.” So, I decided to put Bard Extensions to the test on my own Gmail, to book a flight, and on Google search. Here’s how the experience went.
Once you’ve logged into your Google Account you’ll see a window like this when you open up Bard.
The New York Times reported that Bard Extensions is only available on personal accounts right now.
Since Bard is probably best for summarizing and retrieving information right now, I asked it do exactly that — retrieve the last two emails that my mom sent me.
In my initial query I asked Bard to retrieve emails from a sender named ‘Padmini Varanasi’ without specifying that she was my mom. So, I was pretty impressed to see that Bard had somehow strung together a few details to figure out that Padmini Varanasi was, indeed, my mom. The email summaries were pretty good too, especially since I hadn’t actually read, or even opened her emails to me.
My second order of business was to see if Bard could draft a response — this is where Roose said it might struggle. I decided to keep my request simple, without specifying a tone of voice, to see what it might come up with.
It was helpful that Bard had already given me summaries of two emails that my mom had sent me. I used that to ask it to draft a response about Lenovo’s Global CIO. Bard’s response would have been pretty solid if this was drafted in a professional context. On a personal level, though, it was way too formal.
I ended up showing it to my mom, and she actually laughed out loud. She said she’d probably also be concerned if she ever received an email from me officially thanking her for forwarding me something, and calling it “an interesting and timely read.”Signing off the email with “love” was a nice touch, though, and definitely something I’d do if I ever needed to email my mom about anything — generative AI, Lenovo, and all.
I wanted to see whether I could really implement Bard into my daily work flow. One of my tasks at work every morning is to scan the news of the day and pitch ideas for a quick story. I usually look through the newsletters I receive to pitch ideas, so I decided to put Bard up to the same task.
While I liked that Bard focused on timely themes of generative AI, Gen Z, and consumer habits, my instinct was that this would probably be rejected for being too broad. My editor that day also agreed the angle was too vague, and said while it’s certainly a timely and relevant topic, it’s not very original. Insider has already published multiple stories in recent months interviewing Gen Zers and experts on generative AI.
I asked it to give me a few more specific suggestions. Here’s what it came up with.
While these ideas definitely seemed a little more focused, my feeling was that they were still pretty general, and honestly, not that revelatory to make it into an actual story. That said, the first idea — How Gen Z consumers are using generative Al to create their own personalized fashion designs — did provide an interesting starting point for research. I think it would be fun, as Bard suggested, to find an actual Gen Zer who was making their own clothes with generative AI and interview them for a story.
My editor’s feedback was that these were slightly stronger than the original pitch, but ultimately not very newsworthy.
I was curious to see how incisive Bard could be now that it had access to a wealth of information about me. So, I asked it to determine my “greatest flaw” based on my inbox.
I should caveat this by saying I was taking a page from Roose’s playbook here. He mentioned on Hard Fork that when he asked Bard to analyze his Gmail, and tell him with “reasonable certainty” what his biggest psychological issues were, Bard said that he worried about the future. Bard, meanwhile, told me that I was a “perfectionist” who may have “high standards for myself and others” which sounded a bit like an answer I might give to a job interviewer who asked me to describe my biggest weakness. Bard also suggested that I might be “a workaholic” who is “passionate” about my work, but may be neglecting my relationships or health. It also made sure, though, to include a note availing itself of any responsibility, saying, “You may not have any of these flaws, or you may have other flaws that are not mentioned here.”Based on this, I’d say that Bard has one clear flaw (or strength) depending on your perspective. It’s good at flattering users while keeping its language evasive, which, perhaps, is as important a skill for an AI assistant as it is for any working professional.
Moving on from my inbox, I tested out Bard’s flight booking capabilities, by providing potential dates for a trip to Paris that my friends had floated to me recently.
When I asked Bard beforehand what websites it would use to find flights, it mentioned Kayak, Skyscanner, Expedia, and Momondo.
By the time Bard gave me flight options, though, it didn’t compare them across websites. (There are other websites that do this of course, but they usually open up a ton of browser pages too). Bard only seemed to pull from Google Flights which is where I usually look for flights anyways.
Later, while working, I ended up having to Google myself, and since I had enabled Bard Extensions, a window immediately popped up with a small biography of me.
This biography, however, filled with a litany of embellishments and full on fabrications.
According to Bard:
Overall, Bard painted a pretty glowing portrait of me, even referring to me as a “rising star” in one instance, but I’d say it’s just further evidence that the tool is better with flattery than it is with facts.
The nice thing about Bard’s update is that it does have a button that users can press to “fact-check” Bard’s responses via Google Search.
For instance, I asked Bard “how many parameters are there in GPT-4.”
When you hit that small “G” button Bard will fact-check its response via Google Search and highlight lines in green. In this case, if you click on the “1.76 trillion” it will take you to a story from The Decoder, which is also one of the first results if you search the same question on Google.
So, I went back to see if I could fact-check my own biography by typing “who is Lakshmi Varanasi” into Bard (with the expectation that it might also generate something entirely new).
It presented me with a single line of Hindi transliterated into English, “Main ek language model hoon, isliye main is kaam mein aapki madad nahin kar sakta.”
When I asked Bard to translate it into actual English, it said it translated to, “I am a language model, so I cannot help you with this task.”