Google opens early access to Bard

Google is finally letting people talk to its AI chatbot, but it is taking extreme precautions to make sure you do not confuse it with standard Google Search.

Google has officially opened public access to Bard, an early experiment powered by a lightweight, optimized version of its LaMDA research large language model. For now, access is rolling out via a waitlist at bard.google.com, exclusively for users in the U.S. and the UK.

Google explained that running this lighter model requires significantly less computing power, which allows them to scale the service to more simultaneous users while gathering crucial real-world feedback.

Multiple drafts and a built-in escape hatch

Rather than forcing a single, definitive answer down your throat, Bard introduces a subtle design twist: multiple drafts.

Whenever you ask Bard a question, it typically prepares a few different versions of the response. You can toggle between these drafts to pick the tone or structure that fits your needs best, ask follow-up questions, or prompt it to try again from scratch.

Then there is the prominent "Google it" button sitting right inside the interface. Tapping it generates suggested search queries and opens traditional Google Search in a new tab so you can actually verify web sources.

Google explicitly framed Bard as a companion experience rather than a replacement for Search. To keep interactions from spiraling out of control or going off the rails, Google also capped the number of exchanges allowed in a single conversation.

Calling out its own hallucinations

Instead of pretending the model is infallible, Google leaned heavily into transparency, openly highlighting Bard's capacity to hallucinate.

In one official test case, Bard was asked to recommend easy indoor plants. The bot provided solid tips with total confidence, but completely botched the scientific name of the ZZ plant, calling it Zamioculcas zamioculcas instead of Zamioculcas zamiifolia.

Google also revealed that Bard helped draft and edit the launch announcement itself. While its suggested conclusion did not make the cut, it did awkwardly remind the team to "stay hydrated and eat our veggies."

At launch, Bard is limited to English and cannot write code yet, though Google promised that coding support, additional languages, and multimodal capabilities are in the pipeline.

Why it matters

Google had to get an interactive product out the door to stop Microsoft and OpenAI from owning the generative AI narrative entirely. For everyday writers and brainstormers, the multi-draft feature is genuinely handy for kicking off blog outlines or explaining tricky concepts. But right now, we think nobody should rely on it for rigorous research without leaning heavily on that "Google it" button.

Strategically, Google is playing it as safe as humanly possible. Isolating Bard on a standalone domain and framing it as an "experiment" protects its multibillion-dollar search ad cash cow from the brand damage of hallucinated answers. It is a calculated, defensive maneuver compared to Microsoft's aggressive bet of stuffing ChatGPT straight into Bing.

However, stripping out popular capabilities like coding and throttling the model down to a lightweight tier means early adopters might find it underwhelming compared to GPT-4. Google is prioritizing safety and server costs over raw firepower, and while that may keep them out of hot water today, it remains to be seen whether it can win back the cultural momentum they lost.