Bedor argues that asking for human attention has always carried an implicit price — the writer's effort in framing, editing, and cutting served as proof-of-work that earned the right to a reader's time. LLMs have unilaterally collapsed that producer-side cost to zero while leaving the reader's cost untouched, breaking the equilibrium that kept written channels from being spammed.
Submitted the essay to HN, where it rocketed to 1,306 points in 18 hours — a level of engagement normally reserved for major outages or Carmack posts. The submission framing endorses Bedor's thesis that the asymmetry between cheap production and expensive consumption is now the dominant pain point in professional communication.
Senior engineers, hiring managers, OSS maintainers, and tech leads in the 540+ comment thread describe a 5-10x increase over 18 months in the volume of plausible-looking artifacts they're asked to review — bloated PR descriptions wrapped around 40-line diffs, templated recruiter emails with name-substituted tokens, Claude-cleaned bug reports masking 'it crashed once, idk' observations, and confidently-wrong Stack Overflow answers. The collective exhale of the thread is that almost none of it des
Bedor draws an explicit parallel to bot-flooded early Usenet: the producer's cost of writing was the mechanism that kept communication channels viable. He frames LLMs as doing the same thing bots did, but at higher resolution — generating artifacts that pass a surface plausibility check but carry none of the underlying signal that made the original convention work.
On June 11, Tom Bedor published a short essay titled *If you are asking for human attention, demonstrate human effort*. Within 18 hours it hit 1,306 points on Hacker News with 540+ comments — numbers normally reserved for outages at AWS or a new Carmack post. The piece is barely 800 words. The thesis fits in one sentence: the act of asking another human to read something has always carried an implicit price, and LLMs have made that price collapse to zero on the producer side while leaving it untouched on the reader side.
Bedor's examples are mundane and exactly the ones working engineers recognize. The PR description that runs 3,000 words but the diff is 40 lines of boilerplate. The cold email from a recruiter that's clearly a templated GPT-4 paragraph with your name token-substituted in. The bug report that's been 'cleaned up by Claude' into a structured incident write-up — which would be helpful, except the underlying observation was 'it crashed once, idk.' The Stack Overflow answer that confidently restates the question.
The HN thread reads like a collective exhale. Comment after comment from senior engineers, hiring managers, OSS maintainers, and tech leads describing the same experience: the volume of *plausible-looking* artifacts they're being asked to review has 5-10x'd in the last 18 months, and almost none of it deserves the time it asks for.
The pre-LLM economy of written communication ran on a quiet equilibrium. Writing was expensive: framing the problem, getting the prose to flow, deciding what to cut. Reading was also expensive, but the producer's cost served as a filter — you had to *want* the reader's attention enough to spend an hour earning the right to ask for thirty minutes of it. The cost of writing was the proof-of-work that kept the channel from being spammed.
LLMs did to written communication what bots did to early Usenet, but at a much higher resolution. A prompt-and-paste now generates the same surface artifact — a structured PR description, a polished email, a Markdown-formatted issue — that previously required real cognitive effort. The reader can't easily distinguish 'this person spent two hours thinking about the problem and writing it up' from 'this person spent twelve seconds prompting Claude.' Both arrive looking equally legitimate. Both ask for the same expensive minutes of human attention.
The rational response is the one we're now seeing in practice. Maintainers at projects like Curl and Homebrew have started publicly closing AI-padded contributions on sight. Daniel Stenberg called LLM-generated security reports 'noise' months ago and his rule has spread. Recruiters who paste GPT cold emails are being filtered automatically. Several commenters on the HN thread admitted they now downrank any PR description longer than the diff itself — a heuristic that would have been insane in 2022.
The interesting question isn't whether AI-generated content is *good*. Much of it is. The question is whether the asymmetric collapse of producer cost destroys the social contract that made attention markets work at all. Bedor's answer is yes, and the comment volume suggests he's not alone. Patrick McKenzie summarized it well in a reply: 'The bezzle of treating LLM output as a costly signal is being priced in.'
There's a deeper engineering parallel here. Spam filters didn't work by detecting *bad content*; they worked by detecting *cheap-to-produce* content. Proof-of-work systems — from Hashcash to Bitcoin to CAPTCHAs — exist precisely because asymmetric cost is what makes signals trustworthy. When you remove the cost asymmetry, the signal collapses, regardless of what the content says. We are watching the same dynamic play out in human communication channels, in real time, with no equivalent of SpamAssassin to absorb the blast.
If you ship code, write reviews, or run an open-source project, you're already paying the tax. Three concrete adjustments are showing up in practice and worth stealing.
First, invert the length heuristic when triaging. The 2019 default — long, detailed write-ups deserve more attention — is now actively wrong. A four-sentence PR description with specific reproduction steps and one weird observation is dramatically more likely to come from a human who actually debugged something than a polished 'Problem / Approach / Testing Plan / Risks' template. Curl's maintainers now use length-to-diff ratio as a first-pass filter. Consider doing the same.
Second, make the cost-asymmetry signal explicit in your own outbound work. Senior engineers are increasingly leading PRs and emails with the one specific, weird, off-script sentence that no LLM would have produced: *'I tried three things before this; the second one segfaulted in libcurl 8.4.0 specifically.'* That sentence is cheap for a human who did the work and expensive (read: hallucination-prone and detectable) for an LLM that didn't. Specificity is the new proof-of-work. Use it.
Third, stop letting your tools paste padding into communications that ask for attention. The polished GitHub Copilot–style PR description is, in 2026, a negative signal. Reviewers have learned. If you used an LLM to organize your thoughts, fine — strip the scaffolding and ship the substance. The two-line PR with a link to a failing test passes the smell test. The eight-section formal write-up does not.
The meta-pattern here is that every channel where producer cost was the trust mechanism is going to repeat this cycle, and we're maybe 20% through it. Email next, then job applications, then incident reports, then RFCs, then conference talk submissions. The defensive moves will look ad-hoc and rude for a while — closed PRs, ignored emails, summary rejections — until the equilibrium re-stabilizes around new costly signals: real-time pairing sessions, video walkthroughs, in-person attendance, reputational stake. Bedor's essay is going to be cited a lot in the next year, not because the insight is novel, but because someone finally wrote down the obvious thing that everyone reviewing PRs at scale has been thinking since GPT-4 shipped.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.