I’ve just updated the site FAQ Guidelines. In addition to more consistently using the language of Guidelines for the page itself, there are two major changes:
The first section is re-ordered to more fully anchor to the Julia community standards. Unlike the rest of the guidelines, those are (and have always been) the hard and fast rules of our community. The section now reads:
This is a Respectful Place for Public Discussion
The Julia community is committed to maintaining a welcoming, respectful, and constructive environment. We expect the Julia Community Standards to be observed and upheld by all participants.
The additional guidelines here are not hard and fast rules; they are merely aids to the human judgment of our community.
Please treat this discussion forum with the same respect you would a public park. We, too, are building a shared community resource together — a place to share skills, knowledge, and interests through ongoing conversation. Use these guidelines to keep this a friendly, well-lighted place for respectful public discourse.
The previous intro section is here, for reference.
Please treat this discussion forum with the same respect you would a public park. We, too, are a shared community resource — a place to share skills, knowledge and interests through ongoing conversation.
These are not hard and fast rules, merely aids to the human judgment of our community. Use these guidelines to keep this a clean, well-lighted place for civilized public discourse.
An additional bullet point was added to the “Keep It Tidy” section that provides the explicit guidance:
Don’t post generative AI outputs (but direct human language translation and minor editing is ok).
This really isn’t a significant new rule change. Posting generative AI output often goes against two existing guidelines, and we’ve been moderating it as such. (1) The community standards has long urged all participants to be concise. It’s important enough to be a top-level heading! Generative AI is typically the exact opposite of concise. (2) The guidelines here have always asked folks — people! — to improve the conversation by adding to the conversation and thinking about it.
And, yes, to be clear, if your usage of the generative tool passes the Turing test and adds to the conversation in a relevant and concise way, that’s just fine! Minor editing and human language translation are good examples of such usages and that’s why they’re explicitly okayed. It’s also important to note that — as is now hopefully more fully emphasized by the first change — this is a rough guidelines and not a hard and fast rule. This also just applies to the text you write on this forum, discourse.julialang.org, and not code or content elsewhere.
To help facilitate this, I’ve updated some of the language around our flags and flag messages. Specifically, verbose and low-value generative AI may now be directly flagged for moderator intervention. The flag reason that was previously named “Spam” is now “Spam or Generative AI.”
Speaking personally, I’d almost always rather read the prompt itself. Anyone who wants generated output can go to their favorite model and ask for it.
This is something I’ve talked with folks about doing for a few weeks now. The goal here is to make the guidelines a bit more clear. Remember, unlike the standards, these are just guidelines, and flagged posts are more of an organizational tool and directed feedback than a reprimand.
Could I suggest adding a small note in the guideline detailing why there’s this guideline regarding gen-AI?
I think one who is enthusiastic about modern chatbots might read that guideline and think “ah whoever wrote that is some Luddite who hates the future!”
This post itself provides a somewhat concise reasoning for the rule. So perhaps a small sentance or a link to a footnote saying something like (edited from your post Matt)
This rule is really a subset of two older, more general guidelines: (1) The community standards has long urged all participants to be concise. Generative AI is typically the exact opposite of concise. (2) The guidelines here have always asked folks — people! — to improve the conversation by adding to the conversation and thinking about it.
Coming here from the recent discussion I don’t want to read LLM output - #141 by mbauman it still strikes me as a strange “preference” for an open minded community that fosters participation to impose stricter rules on written contributions than major scientific publishers do these days of fast co-evolution with GenAI?
At Nature Methods, part of the broader Nature Portfolio, we realize that we can’t hold back the oncoming tide that is generative AI, but we can set some guidelines for its use in the publishing process. The key to responsible use, in our view, is transparency.
Nature Portfolio’s policies do allow authors to use generative AI tools to help write sections of their manuscript or edit their manuscripts for readability, but they must declare this use in the Methods section. We advise our authors to not take AI-generated text at face value, but carefully check and edit such text for accuracy. Note that a generative AI tool cannot be listed as an author, as, being non-human, it cannot take accountability for the content of a manuscript.
Elsevier has a similar policy allowing for carefully reviewed and disclosed use of GenAI.
Essentially, both publishers acknowledge the positive influence GenAI can have and weigh the pros and cons to come up with what I find a pragmatic rule to allow for resonsible and disclosed use of GenAI in preparing communication.
(The other thread you mention is still best for describing the rationale behind the guidelines and discussing the merits of generative AI posts, so I will try to stay focused.)
From Elsevier’s page
Editing and adapting all material thoroughly to ensure the manuscript represents the author’s authentic and original contribution and reflects their own analysis, interpretation, insights and ideas.
From Julia Discourse
Don’t post generative AI outputs (but direct human language translation and minor editing is ok).
In both cases use of LLMs is allowed, but just dumping the output of an chatbot is not. It is doesn’t seem to me that the policy here is that much more strict.
Also, I do not think it is possible to cite two journals and assume they represent all “major scientific publishers” either. For example, the esteemed Science magazine would seem to have a slightly stricter policy than either of the ones you mentioned. The magazine goes so as far as to say that
“Editors may decline to move forward with manuscripts if AI is used inappropriately.”
Again, this is very in-line with the established policy for this forum where posts may be removed at the discretion of the moderators if they are flagged by users as not following the guidelines.
Lastly, I do not think that the standards of science journals can or should be applied here. The two communication styles serve difference purposes. The point of a journal article is to tell the world about some area of research in enough detail that others can understand it, judge it, and build off of it. The entire point of a forum like this one is to discuss a specific topic with other humans in a direct back and forth. Because the way the communications happens is so different, different rules are needed to apply to each.
EDIT: The well-regarded JAMA medical journal has an AI policy that allows for AI/LLM use in its article, but is prohibited in the peer review process where they expect humans to respond to each other directly, like in this forum.
That’s some leap there beyond scientific evidence against it. It is quite unclear whether AI-use even for direct communication is not benefitial, see the literature I cited here. Furthermore, the OP started the thread addressing announcements for packages and that to me is closer to scientific papers (e.g., announcing new findings) than to human interaction?
While “all” was never explicitly said, it was definitely implied that the two sources cited are the majority view.
As I said, Science magazine also reserves the right to deny manuscripts for reasons owning to AI/LLM usage, again similar to the moderators discretion.
I would say that the Announcements category is more akin to a press release. When big findings are announced, a carefully crafted short, non-technical blurb is put out, which then points to other sources for the technical information. In this analogy, the docs and/or github readme is the scientific paper: full of details and implementation specifics.
What are you talking about? Of course you can refine your thoughts using AI! What gets blocked are posts of unfiltered LLM output, attribution or not.
I’m quite sure that the policy here is pretty similar in spirit to that of those publishers. If anything, less strict: we don’t have an attribution policy at this time. And I would say, “attribution” probably works against you: if you’ve refined your thoughts sufficiently using an AI, no attribution should be necessary. You are standing behind these thoughts, even if parts of the text may still have been generated by an LLM (to an unrecognizable level). If you need attribution, you’re pretty much doing it wrong.
I understand why in some context, LLM attribution can be useful. But informal communication between users in a forum really isn’t one of them.
So I write some text/thoughts T and have that be refined by the AI to A(T) and then I correct the AI and sharpen this by prompting gwr(A(T)) and then the AI returns A(gwr(A(T)), which I use to produce a revised text gwr(A(gwr(A(T))). Some iterations later we have something that is a say polished thought. Is that “vibe texting”? In my opinion: no.
What’s happening here is that an unreliable process (see the cited literature on how reliable AI detection is) is used to label text. Absurdly then, bad grammar, repetition, incorrect use of words, typical human sloppiness, in short human fallibility is taken as a quality indicator, i.e., this is written by a human being (who is honest and worthy to be a member).
I completely doubt that you can clearly say what is unfiltered or not. If you could, you should write a paper and submit it to the journals, that published those papers I cited.
EDIT: And if you think about it, you might notice that indeed disclosure of use is punished, while simply being silent is not. BTW I claimed that my posted entry was written “using AI”, but I never said, what kind of use that was. Can you prove that I did not simply use AI for translation and minor edits? Or is the burden of proof now on me, which is making this impossible: How to prove the absence of use or the minority of editorial scope?
You are way overthinking this. This isn’t about “can we detect whatever nested function of human and LLM output something is”. If you look at a post and can easily tell “this was copy-pasted straight out of Claude”, you flag it, and then a moderator blocks it. This really isn’t complicated. Don’t worry about these hypothetical edge cases so much.
There’s no automated tooling “detecting AI” here. We aren’t using Pangram or GPTZero or whatever. But, yeah, if you post in big bold letters THIS WAS WRITTEN USING AI then what are moderators to do?
You explicitly stated that “use of AI” is not prohibited per se, it depends on the kind of use and the quality of the contribution (e.g., is it valuable for what is being discussed and does it clearly express what the human wants to convey).
I wrote that I “used AI”.
So, since my text is blocked you must by logically necessity claim that I did not use AI properly. And effectively, while it sometimes sounds as if there is “a lot of slack”, what you are saying is a very strict rule: Only translation is allowed and minor editing. Period.
How can I prove that I used it properly? Show you a German text on a piece of paper so that somehow we can rule out it coming from another ChatBot?
The whole process is ridiculous and will only work reliably for slob, which indeed nobody wants including myself. You cannot prove against me reliably and I cannot prove the absence of something.
Come to terms with the tides: Look for quality and some line of thought, and maybe old-fashioned character. If it is indeed my tone and what you know me to say regularly, well then that is the human Guido talking. I hold that that is exactly what my text does convey.
Where are you getting this? There is no such rule. Not even remotely. “Translation and minor editing” are listed as examples of things that are “guaranteed safe”. Use LLMs to improve posts any way you see fit. Don’t post things that anyone will immediately recognize as straight LLM output. Don’t say “This was written by an AI”.
You are not in a court of law. You don’t need to “prove” anything. If it looks like AI-slop, it may get flagged. If you think a post was flagged unfairly, reach out to the moderator.
Crypes and you guys give me the vibes of Fahrenheit 451. I am all for free speech and free expression, and yes unfortunately as German given recent history, I am kind of allergic since there are some countrymen of yours there in high office who accuse Germany of censorship and violation of freedom of speech rights.
If in doubt, a text should be allowed because blocking speech should be something penultimate in my rulebook. And I believe to have made decent arguments as well.
All is well. Those are the rules. Still, your rules as I argued are stricter than what is allowed in scientific journals and I don’t see where we get to vote on those rules, or do we?
You argued it, but it was pointed out that your argument has no connection to reality. It doesn’t seem like that is getting through to you. So I’m not sure what else can be said.
Voting is the same as for free speech – this is not a state. This is a privately run forum.
If you book a hotel room, you also can not have a vote whether you are allowed to smoke.
I also do not agree, we are stricter here than journals or for example arXiv, where you might get banned for AI slop. Journals just want to make money, that is a complicated comparison.
Besides that I think you are overcomplicating things – and I am German so I am aware germans like to do that. Have a rule for everything.
But not everything can be put into rules and it is also good to not have rules everywhere but really rely on common sense, humble and nice people, and that in general everyone wants to contribute constructively (in general/the broad sense).
I think the rules here are fine as they are. We as a community have to come to an agreement what is acceptable and what not, but there is also probably not a strict line to draw as the many discussion indicate.
That is a crude example because you can still participate in the core service offered: sleep a night in a bed. You will also run into trouble if you say “only for bipods”, people with one leg may not sleep here or other discriminations. So you argument hinges a bit, because that is exactly what is at stake here: You cannot participate in a forum called “Discourse” not because you provided unhelpful or false arguments in a discussion, but because you used a prohibited tool and your text was too polished and a tad too long.
Thanks for trying to put me into a box. The fact that everything I have addressed so far in my posts is being substantiated by the papers I cited in my last post in the other thread. It is instead my experience that some people from the natural sciences and engineering oversimplify softer topics dealt with in the social sciences like sociology (e.g., power distribution).
Do I have to take your word for it or what is the exact governance procedure here? Currently, there is a strict rule (e.g., only minor editing and pure translation) with loose administration, i.e., the moderators have a lot of discretion to decide what is acceptable and what is not.
While you deny this being a democracy or even a state, the principles from democracy are hard to waive: Majority rules but minority rights, clear rules are what give us a free space to meet and engage to the best possible outcome. My freedom to say or write statements should entirely be judged by its content and responsibility, not its background and not depend on mere “discretion.”
Certainly. But Discourse is paid for. And it is paid for by Julia Hub to my knowledge, which receives funding from organizations like NumFOCUS. Such a sponsor is committed to “promote open practices in research, data, and scientific computing” and envisions “an inclusive scientific and research community”.
I certainly wonder whether they really understand that the main discussion and announcement platform for Julia is less clear and substantially more restrictive than major science journals (e.g., disclosure and responsible use will get you published) given that the evidence for GenAI-moderation harming discussions is more than inconclusive; in fact, there seems to be even evidence on the contrary as it may help to avoid misunderstanding?
