Bryan Cantrill says detectable LLM prose drives readers away, points to Pangram as a spam-filter analogue
- Bryan Cantrill argues that substantial LLM-authored writing breaks the writer-reader social contract because readers cannot tell which claims and reasoning the named author actually produced.
- In Cynthia Dunlop's survey of 668 developers, 78% said they stop reading immediately when they detect LLM writing, 71% said they avoid the author later, and 98% preferred an author's imperfect original over LLM-polished prose.
- Cantrill compares LLM prose to email spam: once filtering made spam identifiable at scale, its economics worsened and being labeled spam became costly for legitimate senders.
- After testing multiple detectors, Cantrill says Pangram 3 had very few false positives in samples whose origins he knew, though it missed more AI-written text than he wanted.
- Cantrill says Pangram 4, released a little over a month before the post, substantially improved on Pangram 3 in his use, with low false-positive and false-negative rates.
Hacker News opinions
I think this misses that posts can be progress notes, not polished final products. AI can draft and critique, but the real problem is phoning in the entire process.
I want a browser extension that runs Pangram on the top 100 HN posts so I can skip AI prose. It feels like roughly 20% of front-page links lately are AI slop.
There is already a site that shows Pangram results for HN, but it has no browser extension or API. I would build one if Pangram usage were not so expensive.
I do not buy the claim that readers can reliably identify AI text. If someone reads generated text without noticing, they never register it, and humans also falsely label human writing as AI.
I think the issue is mostly raw copy-paste output. People recognize the default LLM cadence after enough exposure, but someone who meaningfully edits the text has done the work that readers are asking for.
AI writing has become shorthand for writing people dislike, much like other political labels that get applied too broadly. The term risks losing any precise meaning.
At work, people paste bot-written specs and design proposals, and the excess text is obvious. I am fine with bots for research, but I will not read a document somebody slopped together.
I work with someone who uses Claude for nearly every communication, including reading Claude's response aloud on a one-on-one Zoom call. I have told him I want to work with him, not Claude, but he cannot stop.