When did Google get so weird?

Original Article ↗ Hacker News Discussion ↗

Google’s new AI “overviews” in search are drawing criticism for misreading user intent, offering cloying pseudo-empathy, and sometimes hallucinating confidently wrong answers even when correct results are on the same page. Commenters see this as part of a broader decline in Google search quality driven by ad incentives and a pivot from “find documents” toward “talk to an AI,” with real risks when users treat these outputs as authoritative. Many are experimenting with alternatives, lamenting the loss of precise text search, and warning that anthropomorphic chat responses in core utilities like search can blur trust, privacy, and responsibility for bad advice.

AI Overviews Misreading Intent & Playing Therapist

  • Many reports of Google’s AI overview treating plain-text queries (quotes, lyrics, memes, venting-style text) as personal confessions.
  • It often responds with empathy, safety talk, or relationship advice instead of information retrieval.
  • Some find this creepy and intrusive, especially when they only wanted to locate the source text.
  • Others see this “goofy” behavior as old-school “googley” humor, or as aligning with how non-technical users already talk to search.

Declining Search Quality and Exact Match Issues

  • Strong consensus that classic search quality has worsened: more ads, SEO slop, AI-generated pages, and fewer high-signal results.
  • Exact text / quote search that used to work (lyrics, book passages, plagiarism checks) is now unreliable and often intercepted by AI.
  • Boolean operators, quotes, date filters, and site: are reported as less effective or inconsistent.

Hallucinations, Reliability, and LLM Limitations

  • Numerous examples of Gemini confidently giving factually wrong answers (sports standings, video contents, product availability, weather, etc.) even when correct info is visible in the same SERP.
  • Users complain it fabricates sources, misreads URLs, and only admits “I made that up” or “I can’t access X” after being challenged.
  • Big sub-thread debates whether LLMs “understand” or merely predict tokens; some emphasize inherent limits (math, counting), others argue modern models can reason when used with tools/reasoning modes.

User Expectations: Search Tool vs Conversational Agent

  • Tension between “search as precise tool” users and “ask it like a person” users.
  • Some argue Google is correctly optimizing for “normies” who want a Star Trek-style assistant.
  • Others say they never wanted therapy or banter from a search box, only ranked links and snippets.

Social & Psychological Concerns

  • Extended discussion about loneliness, parasocial relationships, and people using LLMs as emotional support.
  • Worries that empathetic personas plus hallucinations could mislead vulnerable users on serious topics (health, law, relationships).
  • Some fear societal effects: dehumanization of information, replacement of human interaction, and manipulation via AI slop.

Business Incentives and Product Strategy

  • Many see AI overviews as ad- and KPI-driven: keeping users on Google, increasing token usage, signaling “AI-first” to investors.
  • Criticism that Google prioritizes cheap, fast models and A/B metrics over accuracy or UX; internal process and promotion dynamics are blamed.

Alternatives and Workarounds

  • Suggested mitigations: browser extensions to hide AI, using udm=14/udm=web on Google, or switching to DDG, Kagi, Brave Search, Yandex, SearXNG.
  • Mixed experiences: some find alternatives vastly better; others note their own spam/AI-slop issues or ethical concerns about certain backends.