What happens when a search engine writes the answer itself?

When a search engine writes the answer itself, it first looks up pages, as it always did, and then asks a large language model to write a short summary of what it found. Google's version, called AI Overviews, displays an AI-generated response at the top of the results page, and it includes links to its sources. The technique behind it, retrieval-augmented generation, has the model refer to a set of documents first and respond second.

The idea is to give the model something to read before it speaks. A plain language model answers from memory, and its memory can be wrong: when Google first demonstrated its chatbot Bard, it gave incorrect information about the James Webb Space Telescope, an error that contributed to a $100 billion decline in Google's stock value. Reading real pages first is meant to keep the answer closer to the facts.

It does not make the answer safe. A model can still misread what it retrieves: in one example, it took a book's chapter title, a rhetorical question, and turned it into a false statement. After the launch in the United States in May 2024, Google temporarily restricted the tool when it suggested things seen as nonsensical and harmful, such as eating rocks or putting glue on pizza.

There is also a fight over who benefits. Publishers feared that people would read the summary instead of visiting their sites, and one publisher sued, while Google's spokesperson said the feature sends traffic to a greater diversity of sites. For you, it means the top of the page can now be written by a machine, so the links are worth checking.