The answer here: What should be our default policy in ChatGPT-provided answers? suggests that we've had a lot of posts that were either written by AI or with the help of AI.

A set of secret instructions were given to diamond moderators, telling them what they're not allowed to do when they suspect that a post is written by AI or with the help of AI, and even though we might never know what those instructions were, it seems that some diamond moderators don't like them: Moderation Strike: Stack Overflow, Inc. cannot consistently ignore, mistreat, and malign its volunteers

None of the diamond moderators in this site seem to have signed the petition (although one of them is extremely active in the strike discussions on the Discord server, and works for a competitor to SE for which he's listed as a co-founder since 2020, and has barely moderated this site since 2020).

I am happy to see that the only "active" diamond moderator (also the only elected one) doesn't seem to be noticeably bothered by the instructions given by the SE staff. I have also not signed the petition or gone on strike, partly but not entirely because I don't know what the secret instructions were, nor have I seen AI have a negative impact on any of the sites on which I spend time. I'll also say that I haven't seen the strike have a negative impact either, but it's too early to say that it won't have an impact on the SE experience.

Whether or not I sign that petition, or ignore it, or do something to support it or to act against it, all depends on the answer to the question in the title of this question. If this set of secret instructions is really a bad thing, please let us know why you feel that way, and if it's something that doesn't impact us much, please let us know why you feel that way!


2 Answers 2


TL;DR: The policy has had no particular impact on this site, although it has had a large impact on a number of the larger sites on the SE network. I support the various mods and sites around the network affected by this but am undecided as to whether I will strike or simply ignore SE's instructions or e.g. try to get clarity on whether they even care about smaller sites following said instructions.

To fully answer and explain all this requires a lot (several years worth) of context of the SE network and company but I'll have a go at addressing the individual parts:

AI-written/aided answers

we've had a lot of posts that were either written by AI or with the help of AI.

We (QCSE) haven't had that much in the way of AI answers or AI-aided answers, as far as I'm aware - at least, as far as both answers I've read and flags I've seen (or rather, not seen) go, things like ChatGPT don't seem to have affected us all that much on QCSE. I might be missing something because I don't particularly use the various languages like Qiskit much, so am less capable of spotting any AI used in writing such code.

However, we (the SE Network) have indeed had a number of posts that are either AI-generated or AI-aided. This has had an effect on some of the bigger sites, in particular Stack Overflow itself. However, the company and the SO moderators strongly disagree on the amount of AI-content on the site, which led to a top-down mandate from the company to effectively stop moderating the content which SO mods believe to be AI-generated/aided, because the company believes it's not AI-generated. This gives us (the SE network) several issues:

(Some) Reasons for mods being on strike

  1. The company didn't undergo previously agreed methods of discussing such changes with the mods prior to implementing them e.g. there are several parts of their data analysis that missed details of how LLMs are used as an aid in answering questions and enough mods have enough examples that clearly indicate their (SE's) final estimated number of AI-posts is wrong. They could and should have picked up on this by simply talking to mods first, which is a reason for such processes existing in the first place.
  2. This edict from the top goes against the community consensus of the majority of (if not all) sites, which is not the principle that SE was built on. Here on QCSE, we've decided to not allow ChatGPT-generated posts, however this edict effectly disallows us from implementing this (unless it violates some other rule). As this is perhaps the point most relevant to QCSE, I'll discuss it further below.
  3. Some of the language used by SE puts mods in a bad light by e.g. implying the mods don't know how to mod, what's good for the community, they're making lots of incorrect decisions etc.
  4. They're not discussing potential moderation mistakes with mods in private.
  5. Similar things have happened before and there's evidently nothing stopping them from happening again.
  6. They're 'x/y-ing' the problem - as in, they're telling mods and users one problem (suspending people for using ChatGPT) while the actual problem they seem to be facing is that fewer people are using SO, leading to a decrease in ad revenue.
  7. Mods had asked for a banner indicating that there was a rule against using LLMs in answers, which would hopefully reduce AI-aided answers and the number of people surprised at getting suspended for using LLMs. The company didn't listen to this.

It may or may not be obvious from what I'm saying here but this also gives another point about how SE (the company) treats SO and the various sites:

  1. They spotted a (potential) problem occuring on big sites and as a result, incorrectly assumed that the same issue applies to all sites (or simply ignored the other sites).

In addition, after the strike had started, it was found out that the company had disabled their data dumps. While people here may find this more or less important, this has long been considered to be 'the insurance against the company becoming evil'.

My views on the letter/petition and moderation options

In general, I haven't been particularly active for the past couple of years. I have handled a few flags and regularly looked around to check things haven't been exploding. Part of that was due to finishing a PhD and starting a job and part was down to needing to reduce how much time I spend on here to avoid burnout caused by the stress of moderating (and the real world) at that time. I don't believe any more details on that are relevant for this post but I would like to become (at least a bit) more active again in the near future.

I'll confirm I did help found a not-for-profit social enterprise 'knowledge sharing' site alongside other current and former SE mods - whether or not it's a 'competitor' to SE, I'll leave up to readers. Any work I do for it is strictly unpaid - my full time job is doing quantum photonics theory. SE (the company) is fully aware of this.

For completely personal reasons that have nothing to do with the strike etc., I'm currently 'inactive' (i.e. officially not performing moderation duties) for the next week(-ish). After this time, if there's no real progress, I do intend to sign the letter, at least to support other mods around the network.

However, there are alternatives to striking, aside from doing what they say (which might not make any real difference anyway, due to the few problematic posts we get): I/other QCSE mods can follow the path of physics SE and simply ignore what the company told us to do; or we can limit the amount/type of moderation we do to e.g. only moderate 'red flags' (spam, rude/abusive). While I'm currently undecided which of these options I'll take, it won't make much (if any) difference as I'm not presently the most active mod and we don't get much AI activity anyway.

  • 1
    $\begingroup$ physics.SE worded it more nicely, but that was basically what I had in mind. For all practical purposes, now instead of closing "because it's LLM-generated" we can close because it's gibberish/useless/etc. For our volume this is perfectly acceptable (though I recognise that things are different for SO). Admittedly, this leaves out the cases of LLM-generated content that is actually well-written, concise, and on-point. There might be such cases that are also recognisably LLM-generated (though there haven't been yet). I'm not sure what we should do with those tbh $\endgroup$
    – glS Mod
    Jun 13, 2023 at 15:02
  • 2
    $\begingroup$ I feel similarly. I've taken a step back from Meta Stack Exchange/the broader network because of the frankly exhausting drama from ~2020, and I mostly stick around QCSE/PSE/MSE now. (I'm also less active in general because of college and so forth, though I do regularly check if there are flags - glS just beats me to them :) The thing is that QCSE is so low volume that issues which are a huge deal on sites like SO, and somewhat of an issue on sites like PSE/MSE, just don't even come up here, or can be handled on a case by case basis. $\endgroup$
    – auden
    Jun 21, 2023 at 20:48

If you're asking why I haven't signed the petition, the honest truth is that I didn't really find the time to read too much about or interact with the whole situation.

I'm aware of the new policy, but I personally don't really think it makes that much of a difference, at least not in the context of this site. Instead of saying a post was closed because it's "chatGPT generated", now we just say it's closed because it's gibberish. As far as I'm concerned, that was the real reason we closed these posts anyway, GPT was just a recognised common source of this particular kind of gibberish. If the post is GPT-generated but actually well-written and useful, well, we probably wouldn't recognise it as such, and I agree it becomes harder to argue that they shouldn't be allowed. A little bit of care while using the model is sufficient to make the post not recognisable as LLM-generated anyway (modulo tools that reliably recognise hidden "watermarks" of a text being LLM-generated, which currently do not exist anyway), as far as the writing style is concerned.

I recognise part of the protests and debates are probably about the way CMs handled things, rather than the policy itself. But, I don't think I have anything useful to say about this topic here.

On the practical side, we haven't had any case of (recognised) LLM-generated post here after this new policy was enacted, so we were never faced with the issue of deciding how to handle them. But as I pointed out above, I recon the only difference is that we'll justify closing these posts because they're bad/misleading/whatever, rather than just saying it's because they LLM-generated, which is arguably the real reason we'd close them anyway (again, at least on this site; there might be different issues on SO due to volume differences, I don't know).

Also, the "secret set of instructions" is not really that secret. The "instructions" are pretty much what was outlined in this meta post, and can be losslessly compressed to "don't use GPT-generated as a close-reason". It's true that, as pointed out by a bunch of people in the comments to that post, the policy as outlined there was worded differently than the "secret instructions" we received. But I'll admit I honestly don't see that much difference between the two. They both just say "don't use 'GPT-generated' as reason to close posts/suspend users", modulo semantic differences which I don't think will matter that much in practice (at least, again, not here). Well, one says "apply a strict standard" and the other says "don't". But really, I'm not sure in what practical scenario this difference would actually materialise.

Some additional details about the data analysis that CMs used to make this decision was recently made available to mods, but that has no relation to actual "instructions" we were given. I don't know anything about "Discord discussions" (nor that people were discussing this on Discord). Also, I don't understand the comment about "elected mods". All current mods are "elected mods". You might argue we are technically all pro-tempore (elected) mods, but again, that's true for all three of us.

A final note, and an attempt to connect this with what was said in the previous post about ChatGPT on this meta. While ChatGPT might not be direct cause for closing/suspending, the fact remains that LLMs can be used to more easily create gibberish posts that are more tiresome to recognise as such. I'd recon that if this becomes more common, a possible counter would be to increase our quality standards accordingly. For example, deciding that very long-form posts that spend a lot of time in apparently unnecessary background information are off-topic, as they'd unnecessarily hurt the conciseness of the post (and are also hallmarks of the standard output of currently widely available language models). But again, I'd suggest spending time to think about these solutions when the problem actually arises.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .