Half of Hacker News Is AI-Written? First Ask What They Counted

Hacker News in 2026 can feel like one sentence repeating: a model launch, an agent demo, a founder recap. A statistic is making another lap of X and Reddit to explain the déjà vu: half of the daily top is AI-written. Treat that as fact and half of the tech public square is not a person talking; the real argument is what the score was pointed at.

Why “half” keeps surviving the argument

Pangram sells itself as a detector for machine-generated text. Security researcher lcamtuf — Michał Zalewski, of AFL fame and a longtime Hacker News regular — ran that scoring against HN posts and revived the uncomfortable question: how much of the front page still looks like a person wrote it.

The latest month of discussion has been thinner than the original fight. The number came back. The caveats did not. The headline still travels because the board already looks like an AI storefront. You do not need a classifier to see model cards, GPU threads, coding agents, and benchmark screenshots in the daily top.

When the subject is already AI, “the prose is AI too” sounds like pattern recognition. It is often just proximity. Intuition is cheap. Sampling is not.

A post about AI is not a post written by AI

Those are two different phenomena, and Hacker News is very good at mixing them.

One day on the front page will usually include a model announcement, a chip-supply thread, an agent demo, and a productivity memoir. That is demand. Readers upvote it. Labs and YC companies submit it. AI as a topic has dominated this site since late 2022 and has not really left. That is a cultural fact about the 2023–2026 tech cycle, not a detector result.

Pangram-style tools score something else: a statistical fingerprint. Smooth conjunctions. Even sentence length. Endings that decline to take a side. Silicon Valley already likes that voice. The funded-startup blog, the changelog that sounds like Stripe, the “what we learned shipping” essay — HN has paid that register in upvotes for a decade. Some of what a tool marks as machine-written is a human doing the local accent.

The inverse exists too. A founder can take a model draft, put a name on it, and publish a “personal” blog that would pass a casual scroll. Collapse topic is AI and author is AI into one share and the narrative leans over immediately.

What Pangram flags, and why lcamtuf flinches

AI detectors are not evidence. They false-positive, especially on English revised by non-native speakers, on release notes that were templates before LLMs existed, and on startup posts that went through an editor.

That is native terrain for a security researcher. People who hunt false positives do not bless a round 50 percent headline. “Half” is a great tweet. It is a poor measurement until you say whether the sample is the daily top 30, the comment section, Show HN, or the body of the blog the title points to.

Fifteen of thirty top posts can be about models while far fewer were drafted by one. A first-person recap under a human byline can still light up the detector. The figure can be internally consistent and still be labeling the wrong column.

The industry problem is older than this thread. Generators change, fingerprints change. Writers who sand the prose to dodge the tool change the score again. Hacker News is a hostile environment for this kind of classifier: short sentences, code, and outbound links in the same post. That was never a clean input.

A verdict on Pangram is a dead end. The live question is how many of the sentences you read as human a person would actually sign.

Echo chambers show up as loops, not as 48 versus 52

The part that should bother you is not whether the share is 48 or 52. It is whether the set list keeps coming back.

Hacker News has been called a Bay Area echo chamber for as long as it has had a front page: startups, Unix taste, a performative distrust of marketing that still boosts a well-timed launch. Model evals and agent demos now sit in those chairs. No model has to type a word. If people keep handing each other the same AI story, the feed still looks generated.

Replies split the usual way. One side says Show HN has become near-identical wrappers on the same API. The other says detectors cannot tell technical writing from machine cadence. Both have a point. One is a narrowing of topics. The other is a bad instrument. Tape them together and you get “half of it is AI.”

Models are not the only way to get an echo. Ranking and upvote taste lift the same links. Readers call that consensus. Models then rewrite the consensus more smoothly. Same on-ramp, similar last sentence.

The cheap check, before you trust the headline, is the unit. Daily top or comments. Title or body. Topic or author. Say “half” without those four boxes and the number is doing the echoing.

The detector fight will not settle what you are reading. The more honest move is to treat the daily tech feed as a blend of lived experience, model drafts, and marketing copy, and to keep those layers visible. Who do you think typed the first sentence of the HN post you actually clicked today?