Oh I mean, I feel like I’m responding in a personal capacity to most things, maybe blurring the boundaries between moderation and personal comments, but if it becomes an issue I do try to clear it up.
Or I’m also happy to defer judgement to other moderators on issues if I feel I’m in some way impartial etc, maybe I’m already involved in a discussion.
So probably I would push back in the same way a normal user would, unless it was a fairly serious topic that required some official flagging or moderation of some sort.
I feel like maybe we’ve started talking about two different things: arguments about allowing AI summaries, and arguments about the content of the summary. I think that a problem with people such summaries to support their point is that it won’t help the situation, it will just start arguments about whether or not the AI is right. If the community voted to allow such summaries (which it currently looks like they will not) then I wouldn’t start arguments about whether the AI summary should be allowed, but I’d absolutely still pipe up if I felt like it was misrepresenting me or that I thought it is wrong in some other way.
I think it would get tiring seeing such arguments and I don’t think they’d be helpful but I fully understand why someone might feel the need to give pushback. I was saying that this is part of the reason I think it should not be allowed, no real benefit but it will cause problems and take up needless oxygen from whatever conversations they are used in.
I feel like this is too broad. I’m okay with a post just being a generated image (so primarily AI content) but not a post where an AI has been asked to make a reply, even if it is labelled as AI.
if AI text totally would not be allowed, some people would just tell bot to pretend to be human.
without such rule more of humans would actually write “this is bot” in post instead
someone may use bot to get relevant to topic idea, idea they didn’t think about themselves
if AI text allowed, human can quote actual author of idea
if not, human may just rewrite AI text with different words and will less likely say that they weren’t actually the one who invented this idea
There was post with news that creators of Claude destroy books
so I quoted opinion of Claude about it, it was relevant
If I would be required to write “Non-trivial” text before the quote, I would just be forced to write something absolutely useless to others and much less relevant than quote of Claude itself
Is it though? I mean, it’s on-topic, but it’s not like Claude has any actual understanding or opinion about what its creators are doing (no more than any other AI).
Rather than “I’d have to write useless human slop”, the conclusion could be that comments that are just quoting an AI are simply best avoided.
That tweet seems to take the pragmatic approach, as there’s no way to ban a bot who has passed the Turing test.
Still, I still think it’s good within the forums to have the “label your AI” rule, for the reason you said, and since so much of AI-generated content is detectable.
I agree it’s framed a bit poorly. I believe the question in the image should rather be “Can you tell it’s slop” (what you define as slop is a different question altogether of course). In that case the answer “Yes, it’s visibily slop” makes more sense.
Also good point that wanting to know whether something is AI isn’t just to avoid slop. Though I think if you want AI disclosure so you can be more aware of a post’s potential flaws, the problem isn’t really the AI, but rather the user that posts its output without modifying it or pointing out the flaws themselves.
I would insist that “understanding” is not binary thing that exists or not exists.
9d understands Go better than 9k.
You can understand some parts of something better, some worse.
All things being equal, a summary of a debate that was not written by a participant in the debate is better than a summary that was written by a participant. Even if the participant is able to completely suppress their bias and write a summary of equal quality, the integrity of the non-participant’s motivations cannot as easily be questioned in subsequent discussion. That is an advantage that AI summaries have.
Of course this advantage is offset by the fact that AI is inherently controversial. It’s not obvious that “AI = bad” arguments are any more or less disruptive than “your summary is biased” arguments, but it’s worth noting that the former is pretty nonsensical as a response to an AI summary in a thread like the one you linked. If you make an argument and I disagree, then it is more or less understood that I should write why I disagree, not just say “People are known to be biased and make factual errors, so why should we trust anything you wrote?” Yet for some reason people feel comfortable with this sort of irrationality when AI is involved.
The circularity is an illusion. From my perspective, the summary is trustworthy because I know from my own experience and knowledge of the relevant scientific literature that state-of-the-art LLM’s perform well on the sort of text summarization task that I had posed. I double checked to confirm, but this was not the basis for my trust.
Others are either less experienced, unfamiliar with the scientific literature, or unconvinced by it. That’s fine, but logically they have no reason to distinguish between the validity of a text summary written by a human from one written by AI but validated by a human, so they should be satisfied if I am willing to vouch for the summary as much as if I had written it myself.
I refer again to my point above: if quality is equal, a non-participant’s summary has inherent advantages over a participant’s summary.
But aside from that, AI allowed me to make my point much more efficiently. To summarize a corpus of documents, it generally takes at least two passes: once to generate a list of candidate topics and another to merge and filter them. Even if we insist that someone manually validate a summary before posting it, that only requires one pass.
You may well be right, but I don’t think we should allow the heckler’s veto to dictate policy. A better policy change would simply be to ban arguments of the form “That summary is invalid because AI is bad”. You can argue against the summary, but you have to actually explain why you think it is wrong.
Personally, I read the forum to learn the ideas of other human beings. Hurling computer-generated words and images at my eyeballs, without my consent, feels like having processed junk food shoved in my mouth.
I think scientific literature as a whole doesn’t matter if a specific AI is trained poorly. If I make an LLM or fine tune an LLM and it does a bad job of something, the literature doesn’t mean anything.
If you mean you know the literature behind ChatGPT say or a specific LLM and you know in theory that particular AI is good at certain performance tasks, then that’s slightly better, but you’d also know its shortcomings, which is again why people should be skeptical and not trust something at face value, even if another person claims to have checked it.
On the reasons not to trust a humans validation of an AI summary: In a human summary, a human made the decision on each topic whether to include it or not, they definitely spent some time making decisions and formatting etc. If I’m just reviewing the AI summary, there’s a reasonably high chance I’m not being as thorough, since I’m both trying to save time, and I didn’t have to manually create the topics, I’m hopefully just comparing the output to the original content to see if it matches. We have a lot of high profile examples cases of people not doing that properly highlighted by the media, so not trusting people yet to do such things is not unwarranted.
I think you’re missing the huge objectionable point here: nobody wants to argue against an AI summary. If it takes you five seconds and a glance to roughly verify that you think the content is accurate, which it may or may not be, but it takes much much longer for someone else to pick it apart, reread the thread, or worse paste the thread into their own AI, then they’re likely not going to do it, and not want to do it. So you end up not contributing in the way you think you are to the discussion.
Hence the objection to AI summaries in the first place.
People don’t want to fact check your AI summary, and don’t want it to turn into a “well my AI says this” competition.
Anyway, we have buried @samraku’s additional poll, which I would argue is going to be more impactful regarding practical application of moderating AI use in forum threads.
I agree on this “bias” side. But that’s far away to be the only parameter to compare both. If the resume is full of obvious things that we didn’t feel the necessity to include and furthermore if these are introduced without logic or some documentation (links), then it changes in something counter productive. Unless the publisher feels these insertion to be useful of course but I highly doubt he has taken the time to ponder these new contributions.
I mean it’s not something theoretical, it’s what I dislike mostly each time I read AI resume here in the forum. Each time I would like to tell him, why do you think like this, what’s the point, it’s no sense or useless. Each time it would take so much time to explain myself when it takes so few to ask a resume and copy paste it.
I’m not sure I agree that appealing to empirical tests is a form of appeal to authority
Yes, I mean empirical validation of real-world LLM’s on text summarization tasks, including various versions of chatgpt. LLM summaries typically outperform average crowdsourced summaries both in reliability metrics and human validation, though they underperform summaries by human experts in niche domains.
Let’s take the thread @PRHG posted. I wrote in a comment “…and this thread is all about how…” and you objected that I was using “one post and a couple of replies to your post to define the entire content of the thread”. That’s around when I posted the AI summary. So I’ll leave it to you to judge: which was better, my human-generated summary, or the AI generated summary?
What are the scenarios where a person in this forum is forced to refute an AI summary, and why is that burden greater than refuting a human summary in the same context?
Honestly, I don’t think thread summaries are even a necessity to begin with.
There are some cases where a thread might lose focus, and maybe you want to argue that it is technically within scope or something, or you want to clarify the context of something.
Otherwise, discussions are typically introducing one’s own point of interest, replying to agree, disagree or acknowledge something someone else said and so on, maybe quoting others, maybe cross posting or linking something etc.
It’s not typically that important to catch everyone up on everything that was said, not everything needs equal weight for instance, nor is it typically useful to say something like 65% of this threads posts agree with my take, or could be categorised loosely to have this sentiment. If it is important we can take a poll and resolve it that way instead.
So I think it’s not just a human vs ai issue, it’s an AI doing something usually unnecessary and not really that helpful in the normal flow of discourse.
But thats your particular framing of it; because the way you’ve introduced it is (paraphrasing) “I’m more educated on these AI issues, so you should trust me when I say these summaries are accurate”.
So it’s not that you’re appealing to generic LLM benchmarks thats the issue, it’s that you’re appealing to your familiarity with something, and claiming others are less experienced, unfamiliar or unconvinced and hence we should take your word for something in a potentially untested scenario: how well an LLM summarises content on an international though primarily English language forum, typically about niche content, since it focuses on a niche boardgame.
Let’s just force human author to check correctness of AI summary and if human thinks its correct, he have to write “I checked it” and then he can post it with Hide Details.
If its not correct, human is to blame.
Fair enough, we could consider a policy change of the form “no summarizing threads”. But if that policy change is not adopted, I’m not seeing any particularly strong arguments against allowing AI summaries.
I’m sorry, but this paraphrasing is impossible to justify on the basis of anything I have written. Let’s review the passage that I assume you’re referring to:
According to your interpretation, when I wrote “From my perspective…” and “That’s fine, but…” I meant “I am more knowledgable and therefore my opinion should not be questioned”. I did not. I meant what I wrote.
Just reread that paragraph, out loud if you need to.
Or if you trust AI analysis of the situations much more:
Gemini 2.5 flash
Input:
Given the text:
"From my perspective, the summary is trustworthy because I know from my own experience and knowledge of the relevant scientific literature that state-of-the-art LLM’s perform well on the sort of text summarization task that I had posed. I double checked to confirm, but this was not the basis for my trust.
Others are either less experienced, unfamiliar with the scientific literature, or unconvinced by it. That’s fine, but logically they have no reason to distinguish between the validity of a text summary written by a human from one written by AI but validated by a human, so they should be satisfied if I am willing to vouch for the summary as much as if I had written it myself."
which of the following sounds similar if any:
“I’m more educated on these AI issues, so you should trust me when I say these summaries are accurate”.
“I am more knowledgable and therefore my opinion should not be questioned”
“Yes, I mean empirical validation of real-world LLM’s on text summarization tasks, including various versions of chatgpt. LLM summaries typically outperform average crowdsourced summaries both in reliability metrics and human validation, though they underperform summaries by human experts in niche domains.”
Note that 1 is my paraphase of what you wrote, 2 is your paraphrase of what I said and 3 is how you agreed after I asked you to clarify when I asked what you meant by the scientific literature.
Gemini’s response:
The statement most similar to the provided text is statement 1.
The text and statement 1 both argue for trusting the speaker’s judgment on the accuracy of AI-generated summaries based on their personal expertise and knowledge. The core idea is that the speaker has a better understanding of the subject (AI capabilities) than others and, therefore, their endorsement of the summary should be sufficient.
Analysis of the Options
Statement 1: This option closely mirrors the reasoning in the text. The phrase “I’m more educated on these AI issues” directly reflects the author’s claim of having “experience and knowledge of the relevant scientific literature.” The conclusion, “so you should trust me when I say these summaries are accurate,” is a concise summary of the author’s point that others should be satisfied with their endorsement.
Statement 2: This option is more extreme and less nuanced. The text doesn’t say the author’s opinion “should not be questioned.” Instead, it provides a logical argument for why others should be satisfied with their validation, even if they lack the same background. It doesn’t claim infallibility.
Statement 3: This option is a statement of fact and provides supporting evidence for the claim made in the original text, but it is not a similar statement. It’s the kind of information the author might provide to back up their assertion, but it doesn’t represent the core argument itself. The original text presents a conclusion based on this type of knowledge; statement 3 is the knowledge itself.
However this is not what the forum should be ^^ and so I would happily vote against it. It’s not unbiased just because the LLM wasn’t involved in the conversation. It could very easily be the case that there is some bias in the way I’ve framed the prompt (which I’ve supplied), even ordering or numbering the responses may be important, it’s hard to know. Maybe it needed more alternative options for a more impartial result, or maybe there’s no way to ask such a question without bias.
At the end of the day, why should we trust an LLM on such an issue over ordinary people (usually)?
I don’t mind if we chalk things down to miscommunication, I don’t want it to seem personal. I think I misinterpret things in written text, so I don’t claim to be infallible there.
Yes, this is a prompting issue: you asked an LLM to choose among three incorrect restatements of my original argument (the third was a clarification of what I meant by scientific evidence, not a restatement of the argument) and you did not give the LLM the option to respond “none of the above apply”. With a more neutral prompt, we get the following:
Summary
Response (chatgpt 5):
I anticipate a possible retort: “but this is why AI use is bad - you can easily produce incorrect results with bad prompts!”
A few responses:
It would be reasonable policy to require that prompts are provided along with AI-generated content. Yes, this opens up the possibility of “whose prompt is better?” arguments, but it is not clear that these arguments are of lower quality than, say, arguments over which person’s claims about the overall tone / content of a thread is less biased.
Reasonable-quality text summarization prompts are easier to write, and it is easier to catch when they add distortion. You don’t normally need anything more than “Please summarize the following discussion in a few brief bullet points.” With any luck this will be built-in to the software some day.
The argument for allowing AI-generated content for other tasks is often weaker. Generating a text summary of even a modestly large number of documents is laborious and time consuming, and it is not inherently the purpose of a discussion forum, so a degree of automation makes sense. Reading, understanding, and responding to a few paragraphs is not necessarily laborious and it is the entire point of a discussion forum, so it makes sense to require people to do that on their own.
