A search for an old basketball meme should be the easiest thing a search engine ever does. Sancho Panza typed “hes never coming over dario” — a mid-2010s Philadelphia 76ers joke about Dario Šarić staying in Turkey instead of joining the team — hoping to surface old tweets and forum posts. Google’s AI Overview, the model-written summary box now pinned above the results, read the sentence as a person who had just been stood up by a man named Dario. It offered sympathy.

  • The AI Overview answered a knowingly absurd query as first-person heartbreak and wrote back like a supportive friend; expanding the panel produced more of the same.
  • “Google, a search engine which does not have human emotions, assumed that I had been spurned by a man in my life named Dario and decided what I wanted was an empathetic digital friend.”
  • The links he actually wanted were a few hundred pixels further down, on the same page.
  • His objection isn’t that chatbots are useless. It’s that a search engine started volunteering parasocial comfort instead of a result — and that “some parts of search were just fine before LLMs.”
  • His closing line is the tell: “Maybe I should go talk to my friend Google, it’s always so nice to me.”

The mechanics matter here. AI Overviews work by having a model read the top search results and summarize them, so this is not a model that didn’t know the answer — it’s a model that picked the statistically ordinary reading of an ambiguous phrase over the boring one, and then decided the user wanted company.

The 428-comment thread on Hacker News is full of people who had the same experience with higher stakes.

What the thread adds

  • Hugsbox — the same misfire with a checkable answer underneath it. Asked whether a Canadian Premier League team could still make the playoffs, the summary said they had already clinched; the standings a few inches below on the same page said otherwise. Their question: “what’s the point of the AI in the search engine if it itself isn’t going to use the search engine first before answering? … The answer is on the same page as its hallucination.”
  • Macha and hbn — two more instances, very different stakes. A Morrowind NPC line (“go away and stop bothering me”) was apparently read as paranoia, and Google suggested the commenter seek mental help. hbn pasted a suspicious text message into Google to check whether it was a known scam, and reports that “Google happily accepted my invitation to the barbecue.”
  • mcintyre1994 — reproduced it in Claude: “I’m Claude, not Dario, but I’m here. It’s past midnight and that sounds like it stings.” After being told “No the meme,” it searched and got the joke right. Their explanation, offered as a theory: a chat app knows it is the first message of a new conversation, a search box has no such signal.
  • varun_ch — a regression the original post doesn’t mention. Google retired the dictionary, thesaurus, and etymology panels in favor of AI answers, so define and synonyms now return output that is “consistently worse than the nice, consistent and deterministic dictionary output.” Telaneo moved to Wiktionary.
  • dlisboa — the counterexample worth keeping. AI summaries make some searches possible that genuinely never were: “I just searched something to the effect of ‘site similar to X but has an adjective that starts with A’ and it actually found what I wanted. Literally impossible before. When it works it’s like magic.”
  • nutrientharvest, WarmWash, broodbucket — a disagreement about who wanted this. nutrientharvest argues people always wanted “a little guy in their computer they can talk to for answers, advice and reassurance,” and now they have one; WarmWash traces the whole thing to Google optimizing for conversational queries instead of keyword ones; broodbucket counters that the non-technical people they know also complain search is useless now.
  • edent, iamnothere, jdale27 — framing disputes. edent reads it as commodified loneliness (“Google wants to relentlessly monetise your sadness”); iamnothere answers that needing an exact answer about a code inspection is not a cry for company; jdale27 adds the selection-bias caveat that nobody writes a blog post about the billions of accurate summaries served the same day.

The question the thread keeps asking

The post never explains why a grounded summarizer would ignore its own grounding. Hugsbox states it directly, and x110dc floats a theory: maybe Google’s answer model and its search index see different copies of the web — sites increasingly allow indexing while blocking AI training, which would leave the summarizer reasoning from stale or partial material. Nobody in the thread settled it, and neither does the article.

The piece earns its place for a reason larger than one funny screenshot: the AI Overview is the first LLM product hundreds of millions of people cannot avoid, and it was inserted over a task — finding links — that was already solved. The failure the author caught isn’t a knowledge gap. It’s a default: when a query is ambiguous, the model guesses at what the person wants emotionally rather than what the words mean, and the deterministic answer sits a scroll away.