People Are Tired of Talking to AI. The Quality Backlash Just Went Mainstream.

The backlash against AI that hit Hacker News last week was not against the models. It was against three specific patterns: forwarding AI answers you didn't read, shipping AI output faster than you edited it, and using the model as an answer machine rather than a thinking partner. Each failure mode got named publicly in the same week. Each one has a skill that addresses it.

What converged this week

On May 27, a post titled "I'm Tired of Talking to AI" hit 1,994 points on Hacker News — the highest-voted AI post of the week. The author described GitHub issues filled with AI-generated answers nobody had verified before posting, a business client sending ChatGPT screenshots instead of reading the content, and several Reddit exchanges that turned out to be with an AI agent. The frustration was precise: not that the AI was bad, but that the people using it had stopped doing their part.

The same week, software engineer Nolan Lawson's post "Using AI to write better code more slowly" reached 1,248 points (May 25). His method: run multiple models as independent reviewers rather than maximising output speed, clear context between reviews, and understand why something is wrong before accepting a fix. Separately, researcher Ethan Mollick published "Choosing to Stay Human" (May 26), synthesising three studies including one in which Turkish high-school students who used ChatGPT for homework scored better on AI-assisted assignments but worse on independent tests. Paul Graham went on record the same week penalising startup founders who sent AI-composed emails, describing it as deceptive. YouTube announced it would auto-detect and prominently label AI-generated content (1,309 HN points, May 27).

Three failure modes, named simultaneously, from separate directions.

The three skills — one for each failure

1. Use AI as reviewer, not just generator. Before you ask the model to "write this," ask it "what's wrong with this?" Run your own draft, your brief, or your thinking past it first. Lawson's insight isn't about being slower — it's about staying in the driver's seat. The model surfaces problems you missed. You decide what to do with them.

2. Edit every output until you'd send it under your own name. The Paul Graham test is practical: if you'd cut a sentence, tighten the opening, or remove a phrase you didn't intend, the output isn't ready. Editing AI text is not a sign the model failed — it's the downstream work that makes the output yours. Five minutes of editing removes the version of you that forwards things you haven't read.

3. Use AI to explain its reasoning, not just produce a result. Mollick's tutor-versus-answer-machine distinction: asking "why is this the right answer?" keeps you cognitively engaged in a way that accepting the first output does not. The Turkish student data is specific. The cognitive shortcut accumulates.

Olivia: the forwarding problem

Olivia is the content lead at a 55-person HR technology company. She runs a weekly AI industry digest for her CEO — she'd prompt a summary of the week's news, paste the output into Slack, and send it. Three weeks in, her CEO forwarded one back with a correction: the summary had described a company acquisition that hadn't happened. The model had synthesised a rumour from a speculative source. The output was confident. It was wrong. She hadn't read it carefully enough to catch it.

Her fix was applying skill one before skill two. She now asks "what might this summary have gotten wrong?" as a second prompt before publishing. The digest is still AI-assisted. It's now her digest.

Marcus: the speed problem

Marcus is a strategy consultant at a 12-person boutique. He was drafting slide talking points for a client pitch using Claude — the model produced a five-bullet framework that looked complete, and he shipped it. In the meeting, a client immediately identified a missing dimension: the framework didn't account for regulatory constraints specific to their sector. Marcus had read the bullets. He hadn't stress-tested them.

Lawson's method transfers directly to non-technical work. Before using an AI-generated framework, ask the model "what's wrong with this?" or "what's missing?" You don't have to accept its answer — but running the critique loop means you've decided, rather than forwarded.

The signal behind the signal

The "I'm Tired" post resonated at nearly 2,000 points because everyone on the receiving end of unread AI output recognised the pattern. YouTube's labelling move (May 27) is the platform version of the same signal: audiences will increasingly know when the author let the model do the work. The professionals who build the habit of editing, critiquing, and staying cognitively engaged now are building something the models can't replace — judgment that comes through clearly in what they produce.