The scsynth.org Terms of Service already say “the Content is not spam, is not machine- or randomly-generated…” so I guess what I would suggest is to use the guidelines/code of conduct to clarify this with respect to LLMs. Probably the “citizenship” section is the right place for this (subsection on LLMs?).
I could edit the guidelines, but the code of conduct in the SC repository should go through a PR review process. So I think it would be better to start with a PR, hammer out consensus there, and then move it over to the forum.
I’ve made a PR using @luka’s text (I don’t know your github name — if you have one — otherwise I’d of credited you better!).
If anyone wants to make an edit, please do so over on github or this is going to get very confusing! You can do this directly in the browser by going to ‘files changed’ then on the far right hand side of the text, the drop down arrow offers a ‘suggest change’ feature.
From my understanding, the way it is already written covers the original problem. It is all there.
edit: I read the PR now. It is okay, but it is also contradictory in this discussion. The description mirrors what we discussed here regarding ethics and genuine effort, but the patch seems to talk about punishing the use of something impossible to identify.
I agree with that, but writing it like that also opens a door for other kinds of problems. And we had problems with harassment using “official” means here. I will not talk about it anymore; it is more than clear.
We ain’t gotta get personal, new people in the future should keep it that way. It feels like fellas are tripping a little over this. – “Them superpowers got neutralized, I can only watch in silence. But don’t tell no lie about me and I won’t tell truths 'bout you”. Who been to hell and back, can show you vouchers. And I tell this thinking about the future of this project.
Yes, you may be talking about one case, maybe someone who is not neurotypical. Then, you use this unique case to write something with huge (hopefully unintended) consequences. With a description incompatible with the injected words. It feels malicious just like (malicious) code. It deserved a soft tap; it was not more than a tap.
Most people want the forum to be a place for human exchange and would not like being tricked into communicating with a machine. Many communities (forums, reddit subs, FB groups) already have rules about AI content. We are actually quite late to the game.
These are community guidelines. Of course, we cannot detect every single violation, but that’s beside the point. The guidelines tell users what kind of behavior is acceptable and which is not. If there are obvious violations, users and moderators can point to the guidelines as a reference.
There was a big discussion about moderation and potential abuse (triggered by your experiences) and I think we are now all aware of these issues. Since then I think moderation has been very fair, and if it keeps that way I don’t see why someone had to worry about the new AI slop rules. Every rule can be abused by bad players; this doesn’t mean there should be no rules.
Yes, this is not a moderation guideline, but a public declaration of intent.
There are obviously no ways to “enforce” such an intent, and that is perfectly fine. Most participants will simply refuse to engage with what they consider automatically generated pull requests. It is good to know in advance for those who are unsure, and indeed the guidelines are explicitly welcoming imperfection. It is very important to say this clearly, and the proposal does this.
I’m still not able to access github, so I’ll start here, summarize there, later.
The characteristics of LLM posts that bother people enough to ban them are (IMO) 1/ the ease with which LLMs slip into (in the technical, Harry Frankfurt sense that distinguishes from lies in that, unlike a lie, the author of is unconcerned with the truth or falsehood of the content), and 2/ the ease of generating large amounts of with minimal input.
If those characteristics are tolerated without any resistance, it would degrade the quality of discourse here. I think it’s right to be concerned about that.
But it’s also possible to have LLM-assisted posts that are not significantly .
When I read the guidelines as drafted in this thread, it did occur to me that they might, only on suspicion of LLM involvement, target posts that don’t deserve to be removed.
So I wonder if it should be reworked to focus on the characteristics of the material rather than its provenance. (I could hand write a couple screens full of and that would warrant criticism, even without an LLM.)
It’s a fair concern, about potentially silencing people based on assertions about someone’s text that would be difficult to prove. (At the same time, if an LLM spins a prompt into a large digressive “plan” of sorts, one may doubt whether anyone is being silenced as the thoughts are unlikely to belong to the person who copy/pasted them. But this may not be true of every use of LLMs. Btw I don’t like LLMs at all – there are already enough cases of AI psychosis and enough teen suicides attributable to heavy LLM usage to convince me that this technology in its current form is not good for humans. I’ve also seen it empower, in at least one case. So I think focus on the material, not on guessing how it was written.)
I used an anti-AI tool, and the guidelines are more than 50% AI probably. Also, many moderators posts are more than 80%, especially some from last year.
James’s post returns as 100% human. ahah You are good with style.
Summarizing the output from James’s post: it was the messiness, AI does not do that! 100% human.
@scztt March posts are very high, especially the checklist style ones, with rhetorical questions, picking more than 2 topics at once. Almost 100% chance AI.
@muellmusik with a hat, official post from October 2024 is really well written, documentation style post, (which may be misleading), but it is 100% human-written.
It is all over the place, with and without (among moderates). Grammar checker is possible with linguistic analysis over time, and checking consistency. Can do that too. For the Anglo gringos only. I will not do that with non-natives.
I will avoid releasing about the priests, nuns and professionals writing our new Linguistic Constitution. Better yet, I will keep it confidential, saved as a holy relic, after a complete analysis of all the input.
I’m out
[This post is a joke. There is no tool that can do that.]
The issue with these plan/proposal posts is that someone has to go through and read all the text in great detail to figure out it is incoherent, overlooks a lot of basic ideas, and does not reflect any of the ongoing conversations. This takes a huge amount of time.
If I see a plan or proposal longer than a page on GitHub I just ignore them now. That isn’t because I think the post is ai, but because I don’t know that it isn’t, and am not willing to spend the time understanding it if there is even the slightest chance the author didn’t do the same.
This isn’t about whether the author used ai, it is whether they used their brain.
I’d be a little cautious of this, the wording could easily accidentally suggest ‘we will ban you if your posts aren’t good’. I think while maybe a little harsh, it is currently clear and actionable.
I don’t know why we are discussing auto ai detection, no one has suggested this and is completely useless imo.
There seems to be a (hidden? or not-so-hidden?) assumption that any power granted to moderators is highly likely to be abused dictatorially.
I’ll speak only for myself, but I don’t have time to inspect every post under a microscope looking either for telltale signs of LLM usage, or for indicators of “low quality,” whatever that is. And I would certainly gain nothing from removing posts or banning users for arbitrary reasons… so… why would I do that?
I mean, sure, somebody could be that kind of a jerk, and volunteer as moderator just to mess with them. But surely such behavior would draw complaints? Surely the majority of moderators would recognize that this is Not Ok and remove that person as a moderator? (Isn’t that one reason to have a team of moderators?)
I do recognize that, for instance, my own country is an extreme example of what can happen when a bad-faith actor acquires power. Perhaps that’s coloring the discussion. I find myself bristling, though, at the implication that a moderator will be unable to resist abusing what morsels of power they have in this tiny tiny corner of the internet, and that to prevent this, it’s essential to argue over how to say “try to stay relevant and avoid wasting others’ time.” Really? Haven’t the current moderators established a good-faith reputation over years (in some of our cases, a couple of decades) of involvement?
Before the emergence of online proofreading tools, I wrote all my posts manually. Once such tools became available, I began using Grammarly, LanguageTool and other paraphrasing tools to proofread and correct them. Occasionally, even native English speakers have provided incorrect or misleading instructions, despite my cross‑checking.
In the last few months (if I remember correctly), free AI services such as Microsoft Copilot have been able to produce a good draft for me when given a task such as: “Write a forum issue regarding the following points: xxx, yyy, zzz, in polite and logical British English.” Of course, such drafts still need to be revised through a series of my own corrections and AI adjustments. Sometimes this process remarkably shortens the working time, while at other times it can take longer than expected for a relatively simple task. A quick check of AI‑generated writing may appear satisfactory, but rereading it after posting often leads me to manually correct a number of expressions.
Nevertheless, I feel that using modern tools makes communication with other users smoother, regardless of their mother tongue.
––
As regards AI‑generated code, I am not entirely sure how reliable it really is. At times it proves helpful, and AI can assist with formatting changes, though sometimes it struggles even with a simple conversion. I used to obtain correct regular‑expression code for text processing quite often. Occasionally the results were satisfactory, but at other times the process became time‑consuming, as I spent more time without gaining much benefit compared with working unaided. Producing correct SuperCollider (sclang) code frequently fails, although on occasion it does resolve mistakes I have made myself. Whether using AI is advantageous or not seems very much to depend on the individual case.
So, I would add the following instruction to the guide:
- Please carefully read any text or code produced with AI‑generated or AI‑assisted tools. Do not rely on code produced solely by AI without your own proofreading and correction.
- Do not post AI‑generated text or code that you have not carefully reviewed and edited yourself.
This text was originally written by me with some intentional grammatical and spelling mistakes, then cross‑checked by Copilot and myself, and finally proofread again (including translation into Korean using DeepL) by me.
The PR description talks about welcoming imperfect contributions and understanding genuine effort, but the actual policy language is about detecting and removing AI content, which, as I noted many times, creates a paradox.
Just adopt the Linux Foundational practical (and scalable model), and double check if it is something that doesn’t put non-native speakers—or anyone with an atypical writing style—in a defensive position. And it was not Linus, it was their legal department that decided that!
(Yes, it is impossible to know that. My previous post was a joke)
US Language tools have been developing since the 1970s, and their goal was to make English easier and simpler for the Navy personel understand written English. After that, it spread to other fields and were software model. English-speaking universities are the only ones that also follow this doctrine.
At the same time, in the real cultural lyricism, the master wordsmiths of this generation barely finish basic education, and are capable of doing this: https://youtu.be/-BjuA1_4Dqk
The funny thing is, originality seems to bring too much suspicion! People would, first of all, be suspicious of anything original. And the wordsmiths of this generation are far away from universities. Very, very far. And Eminem is accused of what? Exactly!!! Using AI. Kendrick Lamar was accused of using ghostwriters.
Where do you read that? Nobody has proposed the use of AI detection tools (which, of course, do not work reliably). The case of AI slop that are worth flagging should be so obvious that you would not need a tool to detect them.
I do not see any references to specific tools in the GitHub PR. Also, please note again that the proposal does not fully ban AI. This is about low-effort generative AI content (“AI slop”). Maybe we need further clarification what we mean by “AI slop”, but the overall intention seems very clear to me.
Some people would calm down if this text is removed from github, because it is inpractice the ‘TRUTH’ of SC project, the source, the official place.
Man, just figure this out before putting things there
(I’m sorry, but the disconnection between description and patch was shocking)
One about ethical effort etc, the patch was about tools.
EDIT: Just quick: for now we will be fine. Do whatever you think it works. Let’s lower the bar from “a framework for the future world and people here” to something to here asd now.
Sorry, I have a hard time following you. If we want to have an actual discussion, you need to refrain from vague allusions that nobody understands.
(I’m sorry, but the disconnection between description and patch was shocking)
I still don’t see the disconnect. If you think that the PR description does not match the content, then propose a better description. (Maybe something “clarify AI usage”?)
I have you have concerns, then please be specific and ideally propose an alternative. Vague allusions and personal attacks are not helpful.
I actually found one part of the PR that I also find a bit problematic:
If AI tools are used, community expects complete transparency about it.
Maybe that is what you meant? I have no idea because your posts are just so vague.
(in the technical, Harry Frankfurt sense that distinguishes