Google is putting more restrictions on AI Overviews after it told people to put glue on pizza

Liz Reid, head of Google search, admitted that the company’s search engine returned “weird, inaccurate or unhelpful AI previews” after they were rolled out to everyone in the US. The executive published an explanation of the more particular responses generated by Google’s AI in a blog post, in which he also announced that the company has implemented safeguards that will help the new feature return results more precise and less newsworthy.

Reid defended Google and pointed out that some of the more egregious responses in the AI ​​presentation, such as claims that it is safe to leave dogs in cars, are false. The viral screenshot showing the answer to the question “How many stones should I eat?” is real, but she said Google found an answer because a website posted satirical content addressing the topic. “Before these screenshots went viral, almost no one was asking Google this question,” she explained, which is why the company’s AI was linked to this website.

The Google VP also confirmed that AI Overview asked people to use glue to make cheese stick to pizza, based on content in a forum. She said the forums generally provide “authentic, first-hand information” but can also result in “unhelpful advice.” The executive did not mention the other viral AI Overview responses circulating, but as The Washington Post According to reports, the technology also told users that Barack Obama was Muslim and that people should drink lots of urine to help them flush out kidney stones.

Reid said the company tested the feature extensively before launch, but “there’s nothing like having millions of people using the feature with lots of new searches.” Google was apparently able to determine trends where its AI technology didn’t get things right by looking at examples of its responses over the past two weeks. It then implemented protections based on its observations, starting by tweaking its AI to be able to better detect humorous and satirical content. It also updated its systems to limit the addition of user-generated responses in previews, such as posts on social media and forums, which could give people misleading or even harmful advice. Additionally, it also “added trigger restrictions for queries where AI insights weren’t as helpful” and stopped showing AI-generated answers for certain health topics.