Google recently confessed to having problems with its new generative AI search feature, after the algorithm suggested people to consume rocks and apply glue to pizza. This development underscores the inherent complications and uncertainties ingrained in the aggressive commercialization of generative AI technology. Google’s AI Overviews function relies on Gemini, a massive language model akin to OpenAI’s ChatGPT, to generate responses for specific search inquiries by summarising information available online.

The current AI boom is centered around the fact that LLMs holding astounding fluency with text, however, this very ability can enable them to convincingly present inaccuracies or errors. Employing the technology to encapsulate online data can enhance the readability of search results, but it can be quite risky when sources are conflicting, or when the information is utilized for essential decisions.

Richard Socher, a prominent contributor to AI for language who introduced a search engine centered around AI named You.com in late 2021, says that while it’s easy to produce a rudimentary prototype with an LLM, it requires meticulous effort to ensure it doesn’t promote harmful actions such as consuming rocks. The technology, on which LLMs are based, holds no real understanding of the world and the internet is full of unreliable data, making it a daunting task to manage LLMs.

Despite vigorous testing of AI Overviews before its launch, Google’s Head of Search, Liz Reid, asserted that the mistakes – like advising glue on pizza and eating rocks – have resulted in additional modifications. Improvements include better detection of baseless inquiries and less dependence on user-generated content.

Unlike Google’s AI Overviews, You.com regularly evades such gaffes, according to Socher, as his firm developed various tactics to prevent LLMs from causing inaccuracies when operated for search. Regardless of the precautions, fine-tuning AI search remains a challenging task. Even You.com failed to accurately answer a complex query involving the names of African nations that start with ‘K.’

Google’s decision to incorporate generative AI into its most popular and profitable product is part of a broader technology industry reinvention inspired by the arrival of ChatGPT, an OpenAI chatbot, in November 2022. Intent on challenging Google, Microsoft – a major partner of OpenAI – used the new technology for enhancing its search engine Bing shortly after the release of ChatGPT. Notwithstanding the technological glitches and peculiar behaviour, Microsoft CEO, Satya Nadella, expressed an optimistic outlook on the venture.

Critics believe that Google has been hasty with its AI upgrade, especially regarding sensitive queries related to finance and healthcare. It is crucial for Google to account for the fact that AI Overviews may face deliberate attempts to confuse it, particularly as it displays results by default for their principal product.

While we can anticipate a reduction in overt errors such as suggesting consuming rocks, AI search can still fall short. More subtle issues include summaries that rely on subpar sources such as articles from a different geographical region or obsolete websites. As a result, users could receive less useful information, primarily when searching for product suggestions.

Not everyone will entirely grasp how Google’s new search functions, nor should they expect the tool to cease making errors. Search consultant and researcher Daniel Griffin, who is developing tools for comparing various AI-powered search services, believes Google will rectify some of the most notable AI Overviews issues. However, he pointed out that no one has worked out how to stop LLMs from missing the truth or making up information. According to Griffin, the problem isn’t AI alone – it’s a consequence of both the internet and the real world, where the truth can be hard to find.