Put glue on pizza? Incorrect answers from Google with AI generate surprise

A recipe for homemade pizza that includes plasticola as an ingredient is the new viral fiasco of AI Overview, the new function with which Google He wants to give a twist to the search engine he created 25 years ago, and which debuted a few days ago in the United States (and can be tested in Argentina).

Some days ago Google activated AI Overview in the United Statesa function that uses Gemini (Google’s ChatGPT-like artificial intelligence-powered assistant) to generate a response to a search by compiling information from multiple sites and synthesizing the data into a single response which Gemini generates for that user.

The feature, which Google announced a year ago and is now available (and will reach more countries over time) aims to replace Google’s classic list of links with text generated based on information on those sites. The new tool is controversial, because it increases a practice that the company has had for a long time, which is to try to resolve all searches on its site, without referring users to another page; It is convenient for the user because they receive all the processed information; For the content creators who generated those sites where Google found the data, it means receiving even fewer visits. Google puts links to information sources at the end of the response generated by artificial intelligence, and allows you to deactivate this new way of searching; But the user already knows that the important thing is already in that response generated by Gemini, Google’s artificial intelligence, and that it is unlikely that they will find anything else in those links.

The problem is that Gemini – like other generative artificial intelligences, such as ChatGPT, Copilot, Claude, Grok and more – does not know if what it has read and is repeating is correct or not: it only relies on the reputation of the content it is reading. getting the information out. In the case of pizza with glue, everything comes from a agreement that Google and Reddit signed last February for which the search giant paid $60 million for the right to use content that Reddit’s 73 million users left in comments over the years, and which allowed Reddit to reach a stock market value of 9 billion dollars a month later, when it went public.

Reddit opened the door to Google’s artificial intelligence, which took that information as valid, including a comment from a Reddit user that was left eleven years ago, and that anyone with common sense can recognize as humorous. “To make the cheese stick [a la pizza] I recommend mixing 1/8 cup of Elmer’s glue [una marca de cola vinílica] with the sauce,” says the comment. “The cheese will not slide and it will give it a unique flavor,” he adds; Gemini probably valued it more because it received positive reviews from users celebrating the humor.

Other incorrect answers that are going viral on social networks have to do with Gemini’s statement that dogs have played professionally in American football, basketball and hockey leagues, or with considering that a viper is a mammal:

One more example: someone asked Google why a cat could be licking their hands… and to the classic answers (affection, anxiety) they added another, also posted on Reddit, in which someone suggested that it was because the feline was deciding whether to eat that person or not.

Google’s response

As Google told The Verge, these types of situations will only happen with unpopular searches, where there are few sources of information and – more importantly – Google’s rating algorithm is not sure which is the best source of information. AND AI Overview remains experimental. But in any case it is a problem for the company: many users are used to the ChatGPT hallucinations (and sometimes they get in trouble for it), but searching for data is one of ChatGPT’s many tasks; In the case of the Google search engine, offering reliable information is the main thing.