According to the company, the AI-driven answers in Google Search were not always found to be “useful.” Since last week, AI Overviews has been curtailed in its ability in several areas, and the company is now explaining the changes it made in more detail.
Glue on a pizza wouldn’t be recommended in a culinary blog. However, if we request recipe ideas from AI, this deadly combination might be suggested. A week after an X user shared screenshots of the glaring response, Google came up with an explanation.
Not enough relevant data on the Internet
Liz Reid, in charge of Google Search, puts some of the blame on the LLM. Specifically, it speaks of “data voids”. That plays a role in search queries for which there are only a small number of search results. Reid cites an example circulated online where the user asked, “How many rocks should I eat?
The statement is preceded by an explanation of how AI Overviews, Google’s name for AI-generated search results, work. It points out a difference from chatbots, which “simply” generate an output based on training data. “Although AI Overviews are powered by a custom language model, the model is integrated with our core web ranking systems and designed to perform traditional ‘Search’ tasks, such as identifying relevant, high-quality results from our index.”
On that basis, Reid concludes that AI Overviews doesn’t actually “hallucinate” or make things up to answer a question. One mistake the LLM makes is through misinterpretation or data gaps. The blog indicates that such issues have been dealt with by, among other things, limiting AI Overviews from producing humorous content. In addition, the AI may answer fewer user questions, as Google says the technology has not proven useful in all queries.