I don’t want to read LLM output

Meta Discussion
1

Today I came across an announcement of a new package that turned out to be hundreds of words of LLM slop. It was unbearable.

Other fora have a hard rule against LLM-generated postings or comments, and they benefit from it. There are automated tools that detect LLM garbage with over 90% accuracy. This forum should implement automatic LLM deletion as a matter of urgency. I’m here to read questions and announcements written by human beings. If the trend continues I’ll stay away. That may be no great loss, but I’m sure there are others who feel similarly.

2

The Julia Discourse also has such a rule (source):

  • Don’t post generative AI outputs (but direct human language translation and minor editing is ok).

When you see such posts, please use Discourse’s “flag” functionality (as described here) to bring them to the attention of the moderators.

3

Could you tell us which announcement it was?

4

Can I ask that we please not do so?

Instead of discussing any specific announcements, can I ask that people flag posts that violate the Julia Discourse guidelines?

As the guidelines say:

Just flag it. If enough flags accrue, action will be taken, either automatically or by moderator intervention.

5

I don’t want to pick on someone who may not understand that he’s doing something offensive. I’m sure he had no intention to harm the discourse here. The flagging solution has not worked and clearly is not working. Instead of trying to sweep up the garbage after we’ve had to wade through it, why not prevent it from littering the landscape in the first place?

6

I’m aware that moderators have (deliberately) given packages announcement (as opposed to “normal” posts) a bit more leeway when it comes to being LLM generated.

Personally, I think we should stop giving such leeway. In particular, because as far as the General registry is concerned, an LLM-generated README is mostly against the guidelines. So, not letting people copy-paste that README here either would go some way of enforcing that guideline.

I understand that LLM-assisted coding is becoming more and more prevalent as the latest generation of models has gotten exponentially more capable. That mode of development can produce amazing results, as long as the maintainer guiding it puts in sufficient care and effort, and at this point it’s not something we can or should limit. But user-facing communication, which includes the README, non-reference parts of the documentation, and announcements here should still be written authored by humans and for humans, since LLMs still lack the judgement to effectively express the human author’s intents, or understand the background of the target audience. At the very least, these should be heavily edited from whatever an LLM drafted.

7

I am also against applying this leeway to announcement threads. LLM generated text is basically just data, so such a thread is a data dump. It’s not communication with other humans. If you want to say something about your package to others, just say it. No need for 500 lines of features listed automatically.

8

I’ve looked at all the recent package announcements and I didn’t find any which seemed unbearable to me.

Have you considered simply not reading things you don’t enjoy? It seems like different people have different tastes in things. Some people like myself love jazz music. Others dislike it. But I think it would be the wrong call if haters of jazz music asked people not to record it, or tried to prevent it from being shared and heard by others.

I’m pretty sure that Discourse has all the tools needed to curate one’s own experience and choose what to read or engage with.

Do you have a monopoly on what is or is not offensive? I find your position offensive, but I’m not going to ‘flag’ you to the moderators. Other people might have a different opinion than I do, so why should I try to silence your speech?

Maybe feeling offended from time to time isn’t the worst thing?

Personally, I haven’t seen a lot of garbage, there seem to be some cool projects being worked on lately, shouldn’t we encourage that kind of creativity?

Those numbers don’t stand up to scrutiny. A peer-reviewed evaluation published this year put leading detectors at 69% and 61% overall accuracy, found performance on hybrid human–AI text dropping to nearly zero, and measured accuracy on scientific writing as 28–38 percentage points below humanities writing [1]. A package announcement is technical prose, so I’d expect it to sit on the lower end of performance. Hybrid is also a very normal case, and these systems fail there. Meanwhile anyone actually trying to evade detection can do so by paraphrasing. In one study, a single follow-up prompt asking the model to rewrite in more literary language dropped detection from 100% to 13% [2].

There’s another issue that’s also important to me. The people these tools tend to misclassify most often are non-native English speakers, flagged at several times the rate of native speakers, because writing in a second language can produce a similar statistical signature to what detectors key on [2]. They are also the people who benefit most from having a model help them write. Work published this year across four scientific domains found generative AI narrowing the publication gap between authors from English-speaking and non-English-speaking countries, with the largest effect where linguistic and institutional resources were most constrained [3]. Julia has a substantial international community. An automatic-deletion rule might have an outsized impact on members with the least ability to communicate or argue about it.

References

[1] Hadra, M., Cambridge, K. & Mesbah, M. (2026). Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity 22, 4. doi:10.1007/s40979-026-00213-1

[2] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns 4(7), 100779. doi:10.1016/j.patter.2023.100779

[3] Does Scientific Writing Converge to U.S. English? Evidence from Generative AI-Assisted Publications. arXiv:2511.11687

9

Is it only me? I am a bit uncomfortable with rules that judge contributions according to who or what helped to write them, rather than by the contribution itself. Is a contribution defective because a probabilistic detection tool marks it as “AI-written with 90% likelihood” (what priors, what likelihood, what false-positives)?

Note, that the rule is not neutral economically: A company may employ specialists and someone from that circle may sign content as “theirs.” A solo developer may instead us AI assistance. I see no good reason here for treating the latter worse than the former?

Why not simply judge contributions by clear, observable properties, and hold the person submitting them responsible? If a post or code is verbose, repetitive, incorrect, irrelevant, not clearly understood by its author, then moderate for those reasons/defects.

I certainly see the argument on the other side: Sinking marginal costs of content production induce a rising cost of moderation. But I still believe that this needs to be dealt with according to “adequacy” or “proportionality” and not by a general, unspecific AI-ban.

10

No, you’re definitely not alone here. I’d say your discomfort is well placed.

11

To give an analogy: A bunch of kids show you cool drawings they made. When looking at that, it’s not about technical judgement on the drawing, it is a lot about the little person who did this, the process, the relationship you have with the kid, and the shared joy in the thing.

To give another analogy: When I do a code review at work, it’s not just about reviewing the code and making sure that crap doesn’t get merged. It is also about teaching and learning and building a shared understanding of the codebase, as well as a shared view on coding-patterns and visions for the codebase. Thoughtful code-reviews by senior devs are how junior devs become senior devs.

The julia discourse is not just about naked information, it is about community-building and shared joy in doing stuff with julia.

If you post LLM-generated stuff here without attribution, then this is abusive: People engage with you as a person, writing thoughtful replies… but you are not a person, you’re a slop-spewing clanker. You lack the biological parts to engage with a thoughtful reply and learn from it [*]. You elicited an emotional and effortful response under false premises, and cheated us out of the opportunity to build a community of like-minded people.

This is evil for exactly the same reasons catfishing is evil.

[*] LLMs do learn stuff. But they don’t do in-context learning. In other words: You cannot have a meaningful conversation as a person with an LLM. A meaningful conversation that does happen is between all of humanity and all of LLMs: Our replies to an LLM-generated post become part of the training sets after the next scrape.

12

Personally, I am not completely convinced that your analogy holds for each and every post in a technical discussion forum. Certainly, there are topics and threads where community and personal engagement are in the foreground, but is that so in general? I am quite often interested in good content and do not care how it was produced.

So what about

Junior uses AI-assisted code.
Senior questions code.
Junior explains the code, notices flaws, revises the code and learns from this?

Learning comes from engaging with something. I would hold that AI-assistance does not necessarily lead to humans not learning or understanding something. That’s a non sequitur, it depends on how AI is used and how you engage with something.

That presupposes that an LLM has completely taken over my account and posts in my name. That need not be the case, I may have discussed something back and forth with an AI agent and then endorsed a final answer. Why is that making the answer (from a hybrid process) non-human and despicable? Note, that you are also mixing up content-quality issues (slop) with content-source again, where you simply assume that bad quality content is simply what AI+humans by some necessity produce.

If a human endorses content that was produced in responsible interaction with AI-help, then that is not cheating in my rulebook.

13

please, there are enough threads broadly debating the merits of AI usage. can this one focus on the original motivating point which was package announcements consisting of several screens of unabridged fully AI-generated content?

14

I haven’t seen any, did I miss something?

15

I don’t know. maybe? but that is what the OP is discussing.

16

That’s what I addressed in my response to them. I’m not supportive of the suggestion, which was:

I think that’s definitely not something which is in service of @foobar_lv2’s ideals around “community-building and shared joy in doing stuff with Julia.” In fact I would say it would likely lead to community-destruction and woe.

17

On the contrary, I find recent ANNs are much better than last year and before. I like the detailed feature list and the motivation behind the package in the ANN.

18

I don’t think anyone here will post LLM generated texts without review. If I post LLM generated texts, that already means the texts fully express my idea. Actually, as a non-native speaker, I find LLM help me a lot in communication, especially in writing formal materials.

19

I think this is a fantastic point, thank you for speaking up about your experience.

20

hey, just to be clear, this is exactly what the OP called out and exactly what the suggestion is to prevent.

it sounds like we all agree that LLM generated texts without review shouldn’t be welcome on this forum?

the discussion is quickly spiraling into pointing at strawman arguments. nobody (yet) in this thread has proposed banning all usage of LLMs to help users express themselves, so there’s no need to start defending that opinion.