Simon Willison has surfaced renewed cases where Google Search shows so-called “featured snippets” (🇩🇪 “hervorgehobene Snippets”) that answer the user search flat-out wrong. Simon writes on Bluesky:

This is a particularly bad case-study in how badly AI summarization can go when its exposed to the wilds of the internet

It’s not “AI Summarization” in the NLP sense of the word however (Bluesky comment): rather, I would call it “RAG poisoning” or perhaps “search result poising”. In essence, the problem to users is which search engines treat non-factual pages as sources of facts. In Simon’s case, Google used fan-fiction pages to make up Featured Snippets about allegedly confirmed cinema movie releases like this:

Both ChatGPT with Search enabled, as well as Perplexity, stayed clear of presenting counter-factuals as actual facts, with Perplexity acknowledging “fan theories”:

With another imaginary movie release, “Shawshank Redemption 2” as proposed in Simon’s thread on X, Perplexity did perhaps not do as well, confirming that “There is a sequel to The Shawshank Redemption in the works”:
Like Perplexity on Encanto 2, ChatGPT with Search enabled calls out the sequel as “imaginary scenarios”, “fictional” and unlikely to happen:

ChatGPT with Search coming up clean in these two tests does not mean it’s entirely safe to use, however. IT-news outlet The Verge has recently called “ChatGPT’s search results ‘unpredictable’ and frequently inaccurate” and called for “Stop using generative AI as a search engine”. I haven’t verified their claims, though.