Google has introduced Bard, a chatbot meant to help people brainstorm and answer questions while keeping Google Search close at hand. The design reflects both the promise of conversational AI and the company’s concern that its answers can be wrong.
A chatbot designed to work alongside search
Bard generates its responses from the model itself rather than looking up search results. Google presents it as a creative collaborator: in a demonstration, it suggested ideas for a child’s bunny-themed birthday party and offered houseplant care tips.
The company does not describe Bard as a Search replacement. A “Google It” button beneath the chat widget invites users to check an answer or explore a topic further. Google’s Jack Krawczyk said the search link helps offset the technology’s limitations.
That distinction matters because a fluent response can sound certain even when it needs checking. In the demonstration, Bard drafted a birthday invitation and supplied the street address for Gym World in San Rafael, California. Krawczyk used Google Search to verify it, and the address was correct.
Three drafts make room for uncertainty
Rather than show only one answer, Bard produces three versions of each response, which Google calls “drafts.” Users can switch among them or combine parts they prefer. The format is intended to signal that no single response should be treated as perfect.
Google says Bard is built on a new version of LaMDA, its large language model, and will be updated as the underlying technology improves. Like ChatGPT and GPT-4, Bard is fine-tuned with reinforcement learning from human feedback, a method intended to make responses more useful and less toxic. Google has not shared many technical details, describing this technology as valuable intellectual property.
Limits are part of the early release
At launch, access is free for people in the US and the UK who sign up for a waitlist. Google calls Bard an experiment and says early users will help test and improve it through feedback. Vice president of research Zoubin Ghahramani said the company is mindful of the ways large language models can go wrong.
Google is also limiting what users can ask and how long they can interact with Bard in a session. The chatbot will not provide sexually explicit, illegal, or harmful material as judged by Google, or personal information. In the demonstration, it also declined to give medical information, including how to spot signs of cancer. Krawczyk said Bard would not give medical advice.
Google has not confirmed the conversation limit for the release, but said it would start quite low and could change based on user feedback. The reason given is that long conversations can make large language models more likely to veer off track. Google pointed to strange responses that surfaced near the end of extended Bing Chat conversations.
Competition raises questions about readiness
The launch comes as Microsoft, working with OpenAI, has challenged Google’s position in search. Google’s first public teaser for Bard, released in February, showed a factual error. The company’s value fell by $100 billion overnight.
Google’s caution has drawn criticism. Margaret Mitchell, chief ethics scientist at Hugging Face and a former co-lead of Google’s AI ethics team, said the company had worked on LaMDA for years. She argued that presenting Bard as an experiment could let large companies reach millions of customers while stepping away from accountability if problems arise.
Search researcher Chirag Shah of the University of Washington sees pressure to move quickly as a central force. He and linguist Emily Bender had previously raised concerns about using large language models as search engines, but Shah said the speed of consumer product integration surprised them. As conversational AI enters products, Google faces pressure to respond even as Bard’s limits remain visible.
Google says it will integrate large language models into Search more deeply, with more details to come. For now, Bard’s drafts, session limits, content restrictions, and Search button show how the company is trying to make an AI assistant useful while reminding people to verify what it says.