Hands-on with Bard Extensions, which lets the chatbot use Gmail, Docs, and Drive data: the feature hallucinated emails, made unbookable travel plans, and more
Hands-on with #Bard Extensions from #Google, which lets the chatbot use data from Gmail, Docs, and Drive: the feature made up emails, made wrong travel plans, and more. … Shriram Krishnamurthi / @shriramk@mastodon.social : My lord, this sounds horrifying. Do you want this? Am I the only luddite who doesn't? None of the questions in this article [ https://www.nytimes.com/... strike me as things I'd want an language model to handle. — [image] Dare Obasanjo / @carnage4life@mas.to : The generative AI space moves at a rapid clip and assumptions about the limitations of LLMs keep changing. — While Google's Bard still doesn't seem to give as good answers as ChatGPT, it's leapt ahead in tools to let you know when search results can verify a statement made by their LLM or flagging where it's making things up. … X: Gary Marcus / @garymarcus : the more things change, the more they stay the same https://www.nytimes.com/... Kevin Roose / @kevinroose : It's a good bet that these Bard features (like most AI tools) will get better over time. But the fact that it was released now, even as an experiment, shows the kind of pressure Google is under to ship new AI tools fast. Lincoln Michel / @thelincoln : In the demo video, basically every example would have been quicker without Bard. Like why ask “What time did Bob say the meeting in the email titled X is?” to an AI when you can just search your email for X and read it in less time? https://twitter.com/... Kevin Roose / @kevinroose : I tested out Bard Extensions, which allows Bard to connect to your Gmail, Docs and Drive accounts and ask questions about your personal data. Cool idea! But it hallucinated fake emails, botched travel plans, and generally seemed not ready for prime time. https://www.nytimes.com/... Forums: BeauHD / Slashdot : Google Bard Extensions Brings More AI Power To Maps, YouTube, Gmail, and More
Context & Ripple Effects
Bard moved from an experimental chatbot to a broader product push: Google expanded access internationally and promised integrations with outside services in its May Bard rollout, then added languages and privacy features for Europe and Brazil in July.
Extensions take that arc into users’ most sensitive work and personal data. The reliability problems reported here matter more than Bard’s earlier, more constrained waitlist experience because the assistant is now being asked to retrieve and act on information from Gmail, Docs, and Drive.
First-order effects
- People using Bard Extensions cannot safely treat its summaries of Gmail, Docs, or Drive as authoritative when it invents emails or produces travel plans that cannot be booked; verification remains necessary.
- Google’s extension feature faces an immediate trust problem: failures are attached to concrete personal-data and task-completion workflows, rather than to generic chatbot conversation.
Second-order effects
- The reported failures raise the operational cost of connecting assistants to productivity suites: users and organizations may limit use to low-stakes retrieval until outputs can be checked reliably.
- Google will face pressure to make source grounding, citations, and handoffs clearer, especially given reporting that contractors already audit Bard answers across sensitive subjects. Competing assistant products pursuing similar integrations face the same expectation.
Third-order effects
- If assistants increasingly become the interface for searching and acting across workplace data, reliability will be judged by whether a task can be completed—not merely whether a response sounds plausible.
- The pattern points to a split between AI features that accelerate drafting or discovery and those trusted to execute consequential workflows; the latter will require stronger controls and auditable links back to source data.
The trend: This is one data point in the shift from standalone chatbots toward embedded, data-connected assistants whose usefulness depends on verifiable task execution.