Argues that every search result, help page, and support form has been replaced by a chatbot, and the cumulative cost of dealing with confidently-wrong, verbose, hedge-laden prose has crossed a threshold. A senior engineer would now rather read a 1997-era man page than navigate another conversational interface.
Frames the HN thread as the moment the market priced the assumption that natural language is the universal interface. For practitioners, chat is the worst possible interface for most technical workflows because it's unstructured, unsearchable, non-deterministic, and hides the system prompt actually doing the work.
Reports using Google's `before:2023` operator as a defensive search strategy because nothing indexed since the LLM boom can be trusted. Treats the cutoff as a practical workaround for an internet now saturated with hallucinated, low-signal AI prose.
Cite concrete failure modes: Google's AI Overviews recommending glue on pizza, customer support bots that gaslight users about refund policies, and GitHub issue triage bots that close tickets by hallucinating the maintainer's intent. The complaint isn't aesthetic — these systems are producing measurably wrong outputs in high-stakes contexts.
Distinguishes this backlash from 2023's 'AI is overhyped' grumbling. Notes the commenters driving this thread are the same developers who shipped Copilot integrations and pay for Cursor — they're not rejecting LLMs, they're rejecting the chatbox UI being bolted onto every product.
On May 27, a post titled 'I'm Tired of Talking to AI' from orchidfiles.com climbed to 1,526 points on Hacker News — top of the front page, hundreds of comments, the kind of organic peak that doesn't happen for a launch post or a vendor blog. The author's argument is short and surgical: every search result is now a chatbot, every help page is a chatbot, every support form is a chatbot, and the cumulative cost of dealing with confidently-wrong, verbose, hedge-laden, em-dash-riddled prose has crossed the threshold where a senior engineer would rather read a man page from 1997.
The comment thread is unusually consistent for HN. There is no 'well, actually' counter-faction defending RAG chat as a UX win. The dominant sentiment, from people who build with LLMs for a living, is that the conversational layer is now a tax — not a feature. Top comments cite Google's AI Overviews recommending glue on pizza, customer support bots that gaslight users about refund policies, and GitHub issues being closed by AI triage bots that hallucinate the maintainer's intent. One commenter, a staff engineer at a FAANG, wrote: 'I have started Googling with `before:2023` because I can't trust anything indexed since.'
This is not the 'AI is overhyped' grumbling of 2023. The signal here is different. The people writing these comments are the same people who shipped Copilot integrations and pay for Cursor. They're not anti-AI. They're anti-conversation.
The industry built the wrong abstraction. The default assumption since GPT-3.5 was that natural language is the universal interface — that wrapping any tool in a chat box made it more accessible. The HN thread is the moment the market priced that assumption. For practitioners, chat is the *worst* possible interface for most technical workflows: it's unstructured, unsearchable, non-deterministic, and it hides the system prompt that's actually doing the work.
Consider the economics from the dev side. A man page costs zero tokens, returns in 8ms, and has a stable URL you can link to in a Slack message six months from now. A chatbot answer costs ~$0.003, takes 4 seconds, varies between runs, and produces output you can't cite, can't bookmark, and have to verify against the primary source anyway. Once you've been burned three times by a hallucinated flag or a fabricated API method, the rational move is to skip the LLM and read the source — which is what the HN thread is collectively admitting.
There's also a compounding content-pollution effect. The same LLMs being used to generate Stack Overflow answers, blog posts, and GitHub READMEs are now being trained on that output. Sourcegraph's recent corpus analysis found that ~38% of new Python code on GitHub in 2025 shows statistical markers consistent with generation by frontier models. The training set is eating itself, and the developer-facing tier of the web — once the highest-signal corpus humanity ever produced — is degrading in real time. The orchidfiles author calls this 'the slop tide.' The phrase will stick.
Compare this to the parallel arc in customer support. Klarna famously announced in early 2024 that its AI assistant was doing the work of 700 agents. In May 2026, the same company quietly re-hired human agents after CSAT cratered and churn ticked up. The pattern is the same: conversational AI is great in demos, fine in low-stakes turns, and catastrophic when the user actually needs to accomplish something specific.
If you're building developer tools, three concrete moves:
One: stop shipping the chat surface as the product. Use the LLM as a backend primitive — for classification, extraction, summarization, code generation — and expose deterministic, inspectable UI on top. Linear's recent agent surface is the template: the LLM proposes structured edits, the user sees a diff, the user clicks approve. No conversation. No vibes.
Two: invest in receipts. Every LLM output your product emits should be cite-able back to a primary source the user can click. If your RAG pipeline can't produce a hyperlink to the doc page that justified the answer, you're shipping a confidence machine, not a knowledge product. Perplexity got this right early; most internal tools still get it wrong.
Three: re-evaluate your support funnel. If your first-line support is an LLM that has to be defeated before a user can talk to a human, you are losing the senior-developer cohort right now. They're the people who write the blog posts that drive your top-of-funnel. Intercom's Fin and Zendesk's bots are seeing measurable user-rage signals — frantic clicking, multiple session restarts, rage-typed expletives. Instrument for those, and route them to humans within two turns.
The correction is already underway. Expect the next 18 months to be defined by 'quiet AI' — features powered by LLMs but never announced as such, with traditional UI affordances and no chat box in sight. The companies that win the developer mindshare back will be the ones that treat the model as infrastructure, like a database or a queue, rather than as a persona to be marketed. The HN thread isn't a rejection of AI. It's a request for engineers to stop confusing the prompt with the product.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.