Cymerys catalogs the tells with forensic precision — em-dashes as garnish, tricolon rhythms, 'not just X but Y' scaffolding, confident hedges, mandatory restating conclusions. He argues that none of these are errors individually, but their aggregate ubiquity across every blog post, PR, and Slack summary has trained his brain to route around them the way it once learned to route around banner ads.
The editorial frames the essay's second-order implication: anyone shipping LLM-generated content faces a distribution problem because the training set is enormous, exposure is continuous across email/docs/PRs, and reader immunity is forming far faster than it did for display ads. What took a decade to erode ad-industry trust is happening in months for AI text.
An essay titled *I'm Becoming AI-Blind* by developer Cymerys landed on Hacker News on August 21 and pulled 333 points and hundreds of comments within a day. The thesis is simple and uncomfortable: the author has developed a Pavlovian reflex to skip past any block of text that smells like it came out of an LLM. Not because it's wrong. Because reading it feels like chewing packing peanuts.
The piece names the tells with forensic precision: the em-dashes deployed like garnish, the tricolon rhythm ('fast, cheap, and reliable'), the 'not just X but Y' scaffolding, the confident hedge ('it's worth noting that…'), the mandatory closing paragraph that restates what you just read. None of these are errors. They are, individually, the moves of a competent writer. In aggregate, at scale, on every blog post, PR description, Slack summary, and support email you now encounter, they've become a fingerprint. And the author's brain has learned to route around the fingerprint the way it learned to route around banner ads.
The HN thread is where the piece stops being one person's confession and becomes evidence. Comment after comment: *I do this too. I didn't realize I was doing it until I read this.* Developers reporting that they skim past coworkers' Jira tickets differently now, that they close tabs on documentation that opens with 'In today's fast-moving landscape…', that they've started treating a rogue em-dash as a bozo bit.
The first-order story is aesthetic complaint. The second-order story is a distribution problem for anyone shipping AI-generated content, which — if you've touched a Claude or GPT API in the last eighteen months — is you.
Banner blindness took roughly a decade to develop and cost the display-ad industry the trust of an entire generation of users. AI-blindness is compressing that arc into months. The training set is enormous (every ChatGPT output since November 2022), the exposure is continuous (email, docs, PRs, comments, search results, LinkedIn), and the pattern is more legible than any banner ad because it's structural rather than visual. You can hide a banner behind a scroll. You can't hide the shape of a paragraph.
What's interesting is *which* signals readers have latched onto. It isn't factual accuracy — LLMs have gotten meaningfully better at that. It's stylistic. The confident-but-hedged register. The compulsive symmetry. The way every explanation resolves into three parallel clauses. These are precisely the features that RLHF optimized *for*, because human raters mistook them for competence. The model learned to sound smart to a rater skimming a hundred completions an hour. The reader, encountering that same voice on every surface, has now learned it means: *no human bothered.*
The HN comments surface a sharper version of the argument: the offense isn't the machine writing, it's the *absence of a decision*. When a human writes badly, you can tell what they were trying to do. When a model writes fluently, there's no author behind the prose — no one chose these words over those words, no one cut the clause that didn't earn its keep. Readers can feel that absence, even when they can't articulate it. The em-dash isn't the problem. The em-dash is the *evidence* that no one edited it out.
Compare this to code. Nobody complains about AI-generated code the same way, because code is executed, not read for pleasure. It either compiles or it doesn't, either passes tests or it doesn't. Prose has no compiler. The only test is whether a human keeps reading, and readers are increasingly voting no.
If your product surfaces LLM output to end users — summaries, explanations, generated docs, RAG answers, email drafts, PR descriptions, ticket triage — you have a UX problem you probably haven't measured. Your users are learning to skip your output. Not click through. Not complain. Just… glaze. The engagement metric that will detect this is not 'thumbs down' — it's dwell time on generated blocks, and it's already trending in a direction you don't want.
A few concrete moves worth considering:
Strip the stylistic tells at the output layer. Post-process to kill the tricolons, cap em-dashes at one per 500 words, delete any sentence beginning with 'It's worth noting' or ending with 'making it a powerful tool for…'. This is a 40-line regex pass and it recovers a surprising amount of readability. The Anthropic and OpenAI style guides both quietly recommend it now.
Prompt for structural variety, not just tone. 'Be concise' does nothing. 'Vary sentence length; no more than one em-dash per paragraph; do not restate the question; do not close with a summary' does a lot. The models can follow structural constraints far better than vibes-based ones. Treat prose style as a spec, not a preference — write it down like you'd write down an API contract.
Show your work. When the output is machine-generated, say so, and show the seams — the source it pulled from, the confidence, the alternative it rejected. Readers forgive machine writing when they understand what the machine did. They resent it when it's smuggled in wearing a human voice.
Log the reflex. Add a 'skip' or 'not useful' signal that fires on short dwell time, not just explicit rejection. The AI-blindness reflex is fast and silent — it won't show up in your explicit-feedback funnel until it's already cost you the user.
The deeper bet here is that the median LLM output has become a commodity, and commodities get skimmed. The next competitive edge in AI-generated content isn't a better model — it's the editorial layer on top of the model that makes the output worth reading in the first place. That's a product problem and a taste problem, not a benchmark problem, which means it won't be solved by waiting for GPT-6. Somebody is going to build the equivalent of an ad-blocker for LLM prose — a browser extension that greys out anything with the fingerprint — and when they do, half the internet is going to disappear. The teams that survive are the ones already treating their generated output like something a human would want to read on purpose.
My son is currently learning Romanian and I was trying to help him with verbs. I don’t know Romanian but recalled when learning a foreign language for the first time it really helped me to break down how a verb form or tense worked in English, then learn the equivalent in the new language. So I want
I've been stuck in loop all week where a coworker asked me to review some methodology documents that are clearly generated by Claude (which is fine in this case) but every time I sit down my brain can't get through 2 sentences without feeling like I'm reading nothing. It actually feel
I also find it impossible to parse half the comments that Claude tries to sneak into our pull requests. I’ve never had an issue understanding code comments written by humans like this before. The structure of the information is like a waterfall that leaves me unable to swim to the surface and grab t
That last image is bizarre. The quiche, cream and even the salad look like they've been given the trypophobia treatment.Which might even make sense, because there were always (still are?) those horrible ads in the chumbox area of news sites that used trypophobia and other creepy body-horror stu
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
There's some psychological mechanism by which my brain immediately recognizes AI generated text and just short-circuits to "there is no information here".And when I force myself to read AI-generated text I realize I'm making my brain do creative work to impart meaning to the word