In my opinion folding posts into an expandable container once they reach a certain length would already go a long way
I just read the Guidelines (for the first time!) and I see nothing there about English or any other language having an official status (although the fact the that Guidelines are written in English might suggest that). Given that, why are we implicitly treating non-English as “foreign”? It would never occur to me to regard a non-English post here as automatically “rude”. Less convenient for me personally, probably, but…so?
Here’s your paragraph as one cogent sentence:
“AI prose is dense and meandering at once, so word count no longer signals information content and the reader does the work the writer should have.”
Did I lose anything? If not, the technique isn’t diagnostic of anything except that most writing is longer than it needs to be, including yours. Oh by the way, an LLM did that summarization.
Why not both?
I like the flavour text to be human written. But if it’s a list of features or a migration guide, i much prefer the LLM output. It’s much more thorough and precise. For telling a story around the features and the release definitely a human.
I’ve been combining both in my announcement posts and i much prefer the results this combination produces. When i used to fully write the release notes i’d always forget to add something, i would make small mistakes everywhere (some of my earlier announcement posts have 10+ edits), there’d never be a migration guide as i’d forgotten what the old version looked like. But now i can get the LLM to use the commit messages and git diffs to do all the mechanical work and free me to recount the tail of how the new release came to be and what it offers.
We actually agree 100%! Flagging posts as “LLM-generated” is just a proxy for posts made by people that have failed to exercise judgment.
I don’t think anybody minds if someone here uses Claude to iterate on a problem, even help them structure an argument, even draft a post. But at the end of their own iterations with Claude, they should have an understanding of what they’re posting, and how other people on the forum will understand it. They’ll have gotten to a point where the final text matches their own understanding. I don’t think we really care if there’s still some Claudish phrase in that text, except that people have gotten really fed up with the firehose of LLM output in their lives, and react allergically to these phrases. And the poster should probably be aware of that, and make it part of their judgment to edit these things lightly.
Posting unfiltered Claude output is what shows a lack of judgment. I think it’s probably entirely possible to draft a perfectly fine post without ever manually editing any of the text, if you iterate and guide and correct Claude through a sufficient number of iterations. That’s the same as agent-assisted coding (like your packages), which we’ve long agreed are not problematic in any way.
The core of what we’re trying to get to is human understanding, and human effort. Whether that effort is spent writing by hand or micro-managing an agent is somewhat secondary.
Nobody has ever objected to machine translation. If you write a post in your native language, and just run it through an LLM for a strict translation, I can promise you that that’s not something that will be banned as “LLM slop”. This is a complete red herring.
Edit: the same applies to “poorly written” English from a non-native speaker, that they’ve then run through an LLM with “Fix typos, grammar, and expression”.
It’s a whole lot more efficient for the author to translate their post to the common language of the forum once, than every reader having to go through this translations by themselves. Original text getting “sloppified” in translation is just not a thing.
Except that your LLM generated summary is actually not really faithful at all to what I said. My point wasn’t that AI writing is too long, it’s that the information in it is poorly organised and lacking a hierarchy, it overemphasises details and underemphasises the central ideas. It almost feels like it doesn’t have a notion of what is actually important in an argument. Which makes it very very exhausting to read.
I also struggle with this when I use AI agents for coding, no matter how I prompt them I always have to skim over fluff and ask them to expand on the actual substance which they’ve somehow managed to summarise into one obtuse word. It would have been so much easier if they had just presented things the right way from the get go. For coding this will always be a necessary evil but I don’t really think we need to have that in documentation or on a forum.
I held off on commenting, and I won’t bother catching up on 80+ comments, but I’m glad this topic got debate and controversy. LLM detection has little empirical validation and is inundated by its own biases and myths (one accuracy score is ridiculous for a situation with false positives and negatives), and even incomplete measures of efficacy degrade quickly as people adopt workarounds (e.g. watermark manipulation https://doi.org/10.48550/arXiv.2509.23019) or switch to different models and versions. I’m not claiming LLM detection cannot possibly be a step in a filter, but giving it the final say in moderation is as foolish as unfettered AI agents. Coming from a guy who thinks people should communicate directly to people or not bother at all, there is just no simple solution to moderating social media anymore.
LLM detection by people isn’t reliable either. Sometimes it’s “obvious”, sure, but we are dangerously overconfident by the time we’re proving it with minute details like counting words, em-dashes, and emoji bullet points. I’ve written “hundreds of words” here before, and nobody has ever confused it with AI slop in the broader context. I’m one of the luckier ones who doesn’t regularly use chatbots (or need to). Just as LLMs are designed to mimic human language, users learn LLM biases in turn (https://doi.org/10.48550/arXiv.2409.01754). In combination with LLM spotters trailing behind frontier models, the AI-fication of human language is a source of false positives, and it can be very detrimental if we don’t curb that in our own scrutiny. In a rather different community, artists are taking increasingly expensive measures to prove their works are not AI-generated to an increasingly skeptical and even hostile customer base, and new models progressively threaten to make those measures obsolete.
That might be what you intended, but IMO you failed to put it into words in your initial post. The LLM consequently couldn’t read your mind when writing the abstract.
That said - my “bulleted” post was surely not written, as the German say, “mit dem vollen tierischen Ernst”, and I’d say it’s not kind of literature really worth that much discussion.
I agree with this 100%. To prove this point, if you really want to torture yourself, try getting Claude to come up with a concise, to the point issue report for something technical and non-trivial. Even after several rounds of back and forth it’s still full of fluff. It’s just incapable of being concise. Same problem for docstrings, documentation, any text intended to be read by a human really.
Maybe it’s interesting to think about why I had been doing this. I was asked on Slack a few weeks ago and I didn’t really have a solid answer… but this comment made me realize a bit more:
I think it’s that I had already been thinking of them as non-narrative data dumps, AI-driven or not! But, really, the best announcements are the ones that don’t just dump a README… they describe the why! I’d surely read more of them if we dissuade changelog/readme dumps entirely… and instead get folks to talk to us more.
And it shows. It fails to capture an important point, and misuses the idiom “at once”. And there are other defects.
You have not shown that most writing is longer than it has to be. You have shown that LLMs have trouble producing even a single normal-sounding sentence that could pass as something written by a human.
I would argue that the idiom “at once” is used correctly according to Webster.
You’re right that I was less than precise, I wasn’t completely sure where I was going with my argument in the original comment. But you could tell that it wasn’t strictly about verbosity from the fact that I mentioned “incredibly dense” as one of the issues. Hope my second comment makes it more clear.
I think the borderline antagonistic back-and-forth on LLM summary capabilities and our opinions of such is missing the point. We can assume LLMs are perfect, and the disconnect between people is still the fundamental problem. To put it politely, I have observed a rise in the LLM era of infodump ANNs that fail to justify a massive list of sometimes complicated features for an underdeveloped motivation. Developers need to convince potential users to care in a much shorter time, and if they can tell an LLM the raison d’etre and eyecatching features in a compact prompt, they can also say that directly to people. An LLM could try to make some up, but users aren’t going to get far into anything when that falls apart. If I had to give one concrete bit of advice, “look what I can do” is the wrong attitude for an ANN. Please tell me why I should try this instead of the existing alternatives, and make digestible tutorials and documentation I can read after I’m convinced.