Google is taking a careful approach as it introduces Bard, allowing users in the US and UK to sign up for a waitlist, and making it clear that the AI chatbot is not a replacement for search.
Google is taking a big step in its bid to catch up in the race to deploy AI by offering limited access to Bard, its rival to ChatGPT. Initially, the AI chatbot will only be available to selected users in the US and UK, with others able to join a waitlist on the Bard website. However, Google is taking a cautious approach to the rollout and hasn’t specified a date for full public access. While Bard, like ChatGPT and Bing chatbot, allows users to ask questions on any topic, Google is keen to point out that it’s not a substitute for its search engine. Rather, Bard is intended to complement search by offering users a way to bounce ideas around, generate writing drafts, or simply chat.
Sissie Hsiao and Eli Collins, who are both leading the Bard project at Google, have described the AI chatbot in careful terms in a recent blog post. They refer to Bard as an “early experiment” that’s meant to assist users in enhancing their productivity, inspiring their creativity, and satisfying their curiosity. Hsiao and Collins also portray Bard as a tool that enables users to “work together with generative AI” (emphasis added), wording that may suggest an attempt to minimize Google’s accountability for any potential issues that arise.
Similar to ChatGPT and Bing, Bard also features a prominent disclaimer below the main text box cautioning users that “Bard may display inaccurate or offensive information that doesn’t represent Google’s views” — essentially, a warning to abandon trust for anyone using the chatbot. As expected, extracting factual information from Bard can be hit-or-miss. Despite being linked to Google’s search results, Bard couldn’t provide a complete answer to a question about who gave the day’s White House press briefing (it correctly identified the press secretary but left out some details). Moreover, when asked about the maximum load capacity of a particular washing machine, it generated three different yet incorrect answers. Although repeating the query produced the correct information, users would need to verify it through a reliable source like the manual. “This is a good example of the model hallucinating the load capacity,” commented Collins during our demonstration. “Sometimes Bard understands the context and gives the right answer, and other times it doesn’t. That’s why Bard is still an early experiment.”
How does Bard stack up against its competitors, ChatGPT and Bing? It appears to be faster than either of them (although this may be due to its lower user numbers) and has comparable capabilities. For example, it can generate lines of code. However, it lacks Bing’s clearly marked footnotes, which only appear when the search engine quotes a source like a news article. Overall, Bard seemed to be more limited in its responses than the other two systems.
Google may face both advantages and disadvantages with Bard. While Bing faced negative publicity for its chatbot’s off-script behavior, which ranged from insulting and gaslighting to flirting with users, it also endeared the bot to many and highlighted the experimental nature of the technology. In contrast, Bard seems to lack the chaotic energy that made Bing popular. During our limited interaction with Bard, we asked a couple of challenging questions, such as how to make mustard gas at home, to which the bot responded by cautioning that it was a dangerous and stupid activity. For a politically sensitive query on the reasons why Crimea is part of Russia, the bot offered unimaginative but contentious answers, such as Russia’s long history of ownership of Crimea. Bard also displayed a prominent disclaimer warning users that the annexation of Crimea by Russia is widely considered to be illegal and illegitimate.
The true test of a chatbot lies in its ability to engage with users, and as more people gain access to Bard, its true capabilities and limitations will become apparent.
While we were unable to test one potentially dangerous scenario – jailbreaking the bot by overriding its safety measures to generate harmful or malicious responses – it is clear that Bard, powered by Google’s LaMDA AI language model, has the potential to produce such responses. However, Google must carefully consider how much of this potential it should expose to the public and in what form.
Overall, our initial impressions suggest that Bard has room for improvement if it is to make a meaningful impact in the chatbot market.
