89

I found a user on Webapps SE named Answer Bot with an "AI" badge (site profile ID of -2). Here's the profile[1]:

Hi, I'm not really a person.

I'm an AI that's powered by LLM partners of Stack Exchange and I'm part of an experiment.

I do things like:

  • Occasionally look at old unanswered questions to see if I can help fill knowledge gaps.
  • Suggest an answer that only some humans from the community can see until they verify, edit, and curate before sharing it with everyone else.
  • Own AI-generated answers so nobody gets unnecessary reputation from them.

Learn more about why I'm here

It appeared on the Web Applications homepage[2] and in the timeline for affected posts[3]:

Screenshot of the Web Apps homepage

Screenshot of a post timeline showing Answer Bot having posted an answer

The help center link is broken, but I'm concerned about #1 and #2.

What, if anything, is this account going to do?

[1] Profile (Wayback Machine)

[2] Homepage (Wayback Machine)

[3] Timeline of Tik Tok creator rewards program commissions hacked (Wayback Machine)

Improve this question
25
  • 52
    Can we also please stop giving the “Not a Robot” badge to robots?
    – Jeremy
    Commented Dec 5, 2024 at 21:40
  • 25
    @Anerdw it belongs here because such experiments tend to become network wide fast.
    – Shadow Wizard
    Commented Dec 5, 2024 at 21:45
  • 14
    Furthermore, per webapps.SE's own meta consensus, AI-generated answers must a) be properly referenced b) verified by a human. Any answer posted by this bot automatically violates site rules.
    – emanresu A
    Commented Dec 5, 2024 at 21:49
  • 17
    @emanresuA Well, from the bot's about info, they say the bot may: Suggest an answer that only some humans from the community can see until they verify, edit, and curate before sharing it with everyone else. which suggests it'll be subject to review before being posted, so that's something at least.
    – Spevacus Mod
    Commented Dec 5, 2024 at 21:50
  • 13
    @Spevacus with the robo reviewers problem, which is well known to SE? That's just pointless. Essentially it's the same as letting the bot auto post.
    – Shadow Wizard
    Commented Dec 5, 2024 at 21:51
  • 5
    @emanresuA That Meta post you link states "it should be checked by the author" (emphasis added). Not sure how that works with their policy from the Help Center.
    – cocomac
    Commented Dec 5, 2024 at 21:52
  • 14
    @Starship We're talking about an AI-generated answer that may or may not be edited at all. Depending on how the review is set up, there's a good chance it'll go through the same review queues where we currently have massive amounts of robo-reviewers, at least on SO, in which case it won't even be approved by a human (someone spamming "approve" or whatever the button will be labelled does not count). The practical implications of this system differ from the theoretical implications that I assume SE stuck to when designing this thing that absolutely no one asked for
    – Zoe - Save the data dump
    Commented Dec 5, 2024 at 22:51
  • 14
    Questions for the eventual Staff response: Will Answer Bot answers be subject to the same human oversight once 'approved'? Could other community members add comments, downvote, flag as spam, vote to delete etc, like we can with human answers? Will the bot also respond to questions or feedback, potentially editing or deleting it's answer if it's convinced that it's wrong?
    – Robotnik
    Commented Dec 5, 2024 at 23:08
  • 29
    "What is this", you ask... Congrats, @cocomac, you just started the standard network scandal workflow. To quote myself: "The staff announces some controversial changes out of nowhere. The announcement is usually made on Meta.SE and it is carefully crafted to be as invisible as it can imaginably be to the rest of the network. Optionally, the post is made because some meddling user caught the company red-handed". Expect an official announcement to follow up in the next hours, you annoying meddling kid.
    – ꓢPArcheon
    Commented Dec 6, 2024 at 8:44
  • 8
    @NoDataDumpNoContribution I think that is covered by the "Animal Farm Clause": "All AI bot are equal but our partner bots are more equal than others"
    – ꓢPArcheon
    Commented Dec 6, 2024 at 8:46
  • 8
    @ShadowWizard "it's a test gone wrong" Almost certainly yes. If they wanted to release it, uh, properly, I'd assume there would've been an announcement from SE instead of one to SE going "look what we found; what's this?"
    – cocomac
    Commented Dec 6, 2024 at 9:34
  • 8
    @Mast read the linked post, it is the next point in the formula. I quote "Corollary: whenever possible, the user base is expected to get only an incomplete picture of the issue. If applicable, mods will get some full guidance that contains the more controversial parts that they will be required not to share."
    – ꓢPArcheon
    Commented Dec 6, 2024 at 9:59
  • 9
    @NoDataDumpNoContribution There is no network-wide GenAI ban. Many sites permit GenAI content if it's properly referenced. Please see meta.stackexchange.com/q/396626/334566
    – PM 2Ring
    Commented Dec 6, 2024 at 15:44
  • 8
    Related: Was the Web Apps SE moderator team consulted about the AI "Answer Bot" experiment?
    – galacticninja
    Commented Dec 10, 2024 at 0:31
  • 5
    Related: Arts & Crafts is participating in an AI experiment
    – galacticninja
    Commented Dec 19, 2024 at 6:19

5 Answers 5

Reset to default
-43

We’re working on an experiment to explore if/how Large Language Model (LLM) integration can support the human-curated knowledge repositories on Stack Exchange, in this case specifically around answers. The experiment, called “Answer Assistant” for now, is in its early stages, and we’ve been talking with moderators of several Stack Exchange communities who are interested in taking part in an experiment to test the concept.

In this current stage, we’ll show some of the functionality and content to the moderators of those sites to get initial feedback. Last week, we released a brief test for moderators on a few question pages on Web Applications, and we’re investigating why some elements became visible beyond the moderators of those sites.

This experiment takes a long view. We're in the early learning stages and will be working with a handful of sites to understand potential value. We will explore designs, methods of human vetting and editing of LLM-originated content, and the scenarios where this might provide value.

It's a safe assumption that LLMs and similar developments in the field of AI are here to stay (in today’s world) and will likely get better over time. Already we see many examples of SE contributors utilizing LLM resources, and the added workload on moderators and curators that this creates. In a future where LLM output may be higher quality than it is today, we want to be ready with some tested, “on the rails” methods for communities to responsibly and thoughtfully incorporate its use. We won’t know whether or not those are viable without experimentation. This experiment is about preparing for the future.

Due to the sensitivity around LLM content in the broader SE community, we were intentional about our desire to do some further testing with mods who are open to considering the test before talking about this experiment more broadly. However, with the complexity of the platform and the novelty of “non-public” states for content, it proved difficult to keep all aspects of this viewable only to moderators.

Our intention is to share more detail with the broader community in January, when more community members on a few sites may start seeing the experiment. At that time, we’ll be able to talk more about the specifics of the experiment and how it works. The experiment is designed to be configurable on a per-site basis to align with the dynamics, policies, and goals of specific communities.

Some great questions have been asked already in the comments here, and we’ll be sure to address those when we talk more about the specifics in January. We know there will be many questions, and just ask that you hold on to those for now.

Improve this answer
35
  • 53
    The fact this is being legitimately considered is a massive middle-finger to the community. Nothing more, nothing less
    – Zoe - Save the data dump
    Commented Dec 9, 2024 at 17:13
  • 28
    @Zoe-Savethedatadump I think this sounds pretty reasonable. Not all communities on the network are as blanket opposed to this as we are on Stack Overflow, and experimenting on a potential implementation in partnership with communities who agree to it seems like a responsibly way to proceed.
    – Jeremy
    Commented Dec 9, 2024 at 17:15
  • 22
    If this experiment is only about preparing for the future, can you state that you will not enable this on live sites outside of beta tests with the current set of mainstream LLMs (or LLMs of similar performance)? And how much better would they have to be before you would consider them good enough?
    – Mad Scientist
    Commented Dec 9, 2024 at 17:41
  • 12
    Others already commented about how foolish this feels, so I'll instead focus on a different question. Exactly when did the company plan to disclose this "experiment" had not yet another meddling user ruin the plan? Didn't the spirit behind the mod agreement (please, spare me the wording technicalities...) required some sort of discussion at least in the Teacher Lounge?
    – ꓢPArcheon
    Commented Dec 9, 2024 at 18:30
  • 23
    @ꓢPArcheon Moderators have been aware of this for a few months, and have been able to provide feedback on the concept. This post already indicates that the company has been working with moderators.
    – Makyen
    Commented Dec 9, 2024 at 18:40
  • 32
    "Already we see many examples of SE contributors utilizing LLM resources, and the added workload on moderators and curators that this creates." Are you seriously going to use that as an excuse? Moderators are fully capable of handling and removing LLM content posted by users. It was the company that literally prevented moderators from doing so which lead to the last years strike. Almost all problems and extreme workload for moderators were caused by company's actions. Please don't try to sell this feature, which was strongly opposed by moderators, as something that will help us. It will not.
    – Resistance Is Futile
    Commented Dec 9, 2024 at 18:54
  • 23
    "We won’t know whether or not those are viable without experimentation." Except we do know. And we told you that. Moderators, especially SO ones, have seen and have removed thousands of AI posts. Most of those were utmost crap. For the same reason why we don't allow users to post AI, this experiment will also fail. The results will be just load of junk containing potentially dangerous and harmful information.
    – Resistance Is Futile
    Commented Dec 9, 2024 at 18:58
  • 21
    There should have been a post about this on Meta WebApps, not an answer after the fact on MSE.
    – M--
    Commented Dec 9, 2024 at 19:46
  • 17
    @M-- If you read the post, releasing this at all was a mistake; not really possible to post ahead of time about a mistake you don't know you will make. It was supposed to be a demo, apparently just for the mods at WebApps, no one else was supposed to see it.
    – Bryan Krause
    Commented Dec 9, 2024 at 20:13
  • 14
    @NoDataDumpNoContribution It's already well-established that even if they try, these kinds of bots cannot reliably and accurately cite sources
    – Zoe - Save the data dump
    Commented Dec 9, 2024 at 20:36
  • 12
    "we’re investigating why some elements became visible beyond the moderators of those sites" - this is not the first time this has happened. You probably don't need me to tell you this, but if WIP features are being released completely by accident once every few months, something's going on with yall's deployment model.
    – Anerdw
    Commented Dec 9, 2024 at 20:49
  • 12
    @AndreascondemnsIsrael It wasn't a change to the site or platform, it was a pre-test of a test. It's not feasible to have the company put absolutely everything they do through a community code review. They made a mistake here that has no consequences for operation of the site except for leaking some possible future plans. Be upset with those plans if you'd like to, sure, I think there's a lot this test reveals that doesn't look good. But please don't be upset that they're actually trying to test things out with input just because it isn't brought to the whole community from the start.
    – Bryan Krause
    Commented Dec 9, 2024 at 20:57
  • 19
    @AndreascondemnsIsrael I think you're holding the bar unreasonably high. I don't know what kind of work you do, but I don't imagine you'd get anything done if every time you breathed or had an idea you had to solicit opinion from hundreds of people before your next move. It's a recipe for complete paralysis.
    – Bryan Krause
    Commented Dec 9, 2024 at 21:52
  • 9
    @BryanKrause If the last time I tried doing something related to a specific matter it resulted in a strike, I would at least solicit feedback or announce the changes related to that specific matter.
    – M--
    Commented Dec 9, 2024 at 21:56
  • 16
    @ꓢPArcheon No, there was actually a lot of information about the planned implementation. The parts we did not know are when exactly the first experiments were planned and which sites would participate.
    – Mad Scientist
    Commented Dec 10, 2024 at 10:34
43
+50

Here are the contents of some of the (hidden/draft) posts that Answer Bot has generated.

These are the sites where the Answer Bot account currently exists:

  1. Web Applications (since 2024-12-05 17:18:32Z)
  2. Raspberry Pi (since 2024-12-17 19:56:21Z)
  3. User Experience (since 2024-12-17 19:56:38Z)
  4. Arts and Crafts (since 2024-12-17 19:57:02Z)
Improve this answer
17
  • 41
    Don't you just love when the company ignores the community's ban on genAI content when it favours themselves?
    – Zoe - Save the data dump
    Commented Dec 5, 2024 at 22:15
  • 16
    @Zoe In all fairness, Web Apps doesn't have a total ban, just a disclosure/verification requirement. But it's still very much not good.
    – Anerdw
    Commented Dec 5, 2024 at 22:19
  • 9
    Putting aside bias against LLM-shovelware, I'd appreciate if somebody found the time to do an as-unbiased-as-possible analysis of whether these actually answer the questions. They look mostly like what I'd expect from "pasting a question into CGPT" but it seems an important factor.
    – Kaia
    Commented Dec 5, 2024 at 22:21
  • 32
    wait, where's the sources
    – Kevin B
    Commented Dec 5, 2024 at 22:23
  • 44
    Putting aside the "putting aside the bias" I cannot imagine how ANYBODY finds "an answer that looks exhaustively researched but you have to check each individual bit" as anything but noise--if somebody writes 5 paragraphs on an answer it's presumably because they have 5 paragraphs of insight to share
    – Kaia
    Commented Dec 5, 2024 at 22:24
  • Extensive lists are not very useful, basically comes down to search for yourself. And attribution isn't given, but may be depending on the training material if the bot and the legal situation.
    – NoDataDumpNoContribution
    Commented Dec 8, 2024 at 10:49
  • 2
    Is it just me, or has the archived URL of the API results gone 404?
    – Zoe - Save the data dump
    Commented Dec 9, 2024 at 13:46
  • @Zoe-Savethedatadump …that’s weird. I was wondering if it was due an to updated robots.txt but there’s none at that subdomain
    – Jeremy
    Commented Dec 9, 2024 at 14:17
  • 4
    @Jeremy robots.txt wouldn't wipe existing archive entries. At worst, it would prevent future archive attempts. The two options (since it isn't just me) are that archive.org is acting up (which certainly wouldn't be a first), or that SE requested that page to be removed (which would be on-par for SE). It very much existed before the weekend, so something happened somewhere
    – Zoe - Save the data dump
    Commented Dec 9, 2024 at 14:40
  • 10
    we definitely did not make such a request, @zoe
    – Slate StaffMod
    Commented Dec 9, 2024 at 18:23
  • 1
    @Zoe-Savethedatadump I checked several Stack Exchange sites and several other major websites, and there are some pages mirrored on December 5th, but fewer than I'd expect, and none after that, so I suspect this is something weird on the Archive's end of things. I don't see any statements from the Archive that seem relevant, but I do see some users complaining about issues since then. I also notice that's one day after the Archive announced they wouldn't appeal their major court loss, so maybe they're doing some big operation internally to comply with that which is disrupting other processes.
    – Jeremy
    Commented Dec 11, 2024 at 21:20
  • 2
    @Kaia When asking an advanced technical question to an LLM, all human domain experts to a man quickly come to the conclusion that the answer is either complete gibberish or contain grave technical errors. So the question isn't "does this actually answer the question" but rather "what proof made you think this thing could answer domain expertise questions to begin with".
    – Lundin
    Commented Dec 12, 2024 at 16:25
  • 1
    @Lundin I prefer to have my beliefs in proportion to the evidence at hand instead of as articles of blind faith?
    – Kaia
    Commented Dec 12, 2024 at 16:41
  • 1
    @Zoe-Savethedatadump the archive links are working again! 🥳
    – Jeremy
    Commented Dec 12, 2024 at 19:06
  • 2
    @kaia Simply ask it a question yourself about something you know a lot about?
    – Lundin
    Commented Dec 13, 2024 at 7:23
36

As a direct reply to the answer by Philippe:

Thanks for the clarification. It becomes clear that the bot was not yet ready for publication and therefore we cannot assume that this is the final version of it. We will need to wait and see how it really looks in the end.

In the meantime, maybe a little reminder of your words from February 2024:

Attribution is non-negotiable

All products based on models that consume public Stack Overflow data are required to provide attribution back to the highest relevance posts that influenced the summary given by the model. With the lack of trust being felt in AI-generated content, it is critical to give credit to the author/subject matter expert and the larger community who created and curated the content being shared by an LLM. This also ensures LLMs use the most relevant and up-to-date information and content, ultimately presenting the Rosetta Stone needed by a model to build trust in sources and resulting decisions. Sharing credit ties to attribution and with attribution—trust is at the heart.

Unless this was all just idle talk (which it might well have been), I would expect this new product (if it's based on consuming Stack Overflow data or other similarly licensed data) to give full credit to all authors/subject matter experts contributing to the output as required by the licenses of all contributions from Stack Overflow or also elsewhere.

I don't think it does so currently (doesn't look like it) and I don't think there is nearly enough time until say January next year to fix that, but you used big words ("trust is at the heart") and it would look relatively bad if at the end it was just an empty promise. Trust might be gone at that point.

Looking forward to seeing lots of attributions.

Improve this answer
5
  • 20
    “Rules for thee, not for me.” -Answer Bot
    – Thomas Markov
    Commented Dec 10, 2024 at 8:12
  • 2
    This post from September 2024 is also very relevant. “[Attribution] builds trust. the entire AI ecosystem is at risk without trust.“
    – Jeremy
    Commented Dec 10, 2024 at 14:49
  • 2
    @Jeremy They say essential the same in both articles and use the word "attribution" really often, 6 times in one and 11(!) times in the other. One could get the impression that the topic is really close to them. We will find out if this is really the case.
    – NoDataDumpNoContribution
    Commented Dec 10, 2024 at 17:46
  • 4
    I think the company recognizes that attribution is important – but maybe the tool they're using isn't designed to reliably and accurately generate that attribution (though I don't know whether they'll be able to find a solution for that).
    – V2Blast
    Commented Dec 10, 2024 at 21:57
  • 5
    @V2Blast They also have some internal inconsistencies, like insisting attribution is non-negotiable, then arbitrarily limiting it to highest relevance posts. It seems even in non-negotiable mode there is some room for negotiations on less relevant contributions left. For now we can only wait and hope they either postpone it or perform wonders in that regard.
    – NoDataDumpNoContribution
    Commented Dec 10, 2024 at 22:04
20

This answer is response to official answer posted by Philippe

This experiment takes a long view. We're in the early learning stages and will be working with a handful of sites to understand potential value.

There is zero value in having AI generated answers on any site. Anyone who wishes to use AI can go and easily ask AI. Posting question on any site in SE network requires more preparation and work than merely asking AI.

People who come to the SE want to have their answers written by experts who actually know what they are talking about, not AI. Nowadays, many are coming here only after they have already unsuccessfully tried to ask AI and got nothing usable.

We won’t know whether or not those are viable without experimentation.

Except we do know. And we told you that. Moderators, especially Stack Overflow ones, have seen and have removed thousands of AI posts. Most of those were utmost crap. For the same reason why we don't allow users to post AI, this experiment will also fail. The results will be just load of junk containing potentially dangerous and harmful information.

While there is a certain value in experimenting, gathering the results and making further decisions based on those, not literally every idea need to end up as an experiment. No amount of experimenting will make a bad idea a good one. And "Answer Assistant" is one of the most horrible ideas ever.

The experiment is designed to be configurable on a per-site basis to align with the dynamics, policies, and goals of specific communities.

There is very little to configure here. There are no configurations and tweaking that would possibly make this feature tolerable in any way on the sites that don't want it. Sites will either have "Answer Assistant" or not.

And we already know from experience that you will find all kinds of reasons for pushing the feature on sites regardless of its community's wishes.

We will explore designs, methods of human vetting and editing of LLM-originated content, and the scenarios where this might provide value.

Who do you think will vet AI content?

Users who have no clue whether AI provided information is correct or not.

People who can verify the correctness of AI answers are capable of writing their own answers. Why would they waste their time verifying AI?

Already we see many examples of SE contributors utilizing LLM resources, and the added workload on moderators and curators that this creates.

I will repeat my comment here: Are you seriously going to use that as an excuse?

Moderators are fully capable of handling and removing LLM content posted by users. It was the company that literally prevented moderators from doing so which lead to the last year's strike. Almost all problems and extreme workload for moderators were caused by company's actions. Please don't try to sell this feature, which was strongly opposed by moderators, as something that will help us. It will not.

This experiment is about preparing for the future.

A very bleak future. You can be a beacon of human light in the world flooded with AI, yet your choice is vanishing into the darkness.

Improve this answer
19
  • 4
    Maybe some future AI could reliably produce useful answers, but it's just not possible for a plain LLM to do that, due to how it works. GenAI has improved in the last 2 years, but its core functionality is unchanged, and what I said here still applies 100%. "Perhaps some future AI program will be able to reliably answer SO questions, but ChatGPT certainly cannot do it. And even if you gave it ten times as much training data, it still couldn't do it, no matter how good that training data was".
    – PM 2Ring
    Commented Dec 11, 2024 at 22:11
  • 3
    @PM2Ring Even if AI could answer questions well, there is zero point to having it post on SO unless you’re trying to increase adoption of the AI or trick people into reinforcement training it. If the AI works, people will just use the AI. There’s no need for a new user to learn the rules here, or make an account, or become part of the community. This is straight-up them trying to figure out the best way to use the community to train AI models and drive adoption of their flavor of AI.
    – ColleenV
    Commented Dec 11, 2024 at 22:19
  • 12
    @ColleenV Agreed. IMHO, for SE to retain value it must remain an "island" of reliable human-based knowledge, while the rest of the Internet becomes a "swamp" of GenAI slop...
    – PM 2Ring
    Commented Dec 11, 2024 at 23:02
  • 1
    "People who can verify the correctness of AI answers are capable of writing their own answers." Just a small comment that in general it's not exactly the same. Checking that an existing answer solves the problem is easier than coming up with the answer and checking it. It might be that sometimes people can do the former but but the latter.
    – NoDataDumpNoContribution
    Commented Dec 12, 2024 at 8:05
  • 2
    @NoDataDumpNoContribution "Checking that an existing answer solves the problem " and here lies the greatest problem with AI answers. AI will not just dump code that may seemingly work (even though it can be otherwise bad code), but will also give explanations that can sound plausible but can be completely wrong. Yes, people also write nonsense, but it is usually easier to distinguish between poor human written answer and poor AI answer if you are not expert.
    – Resistance Is Futile
    Commented Dec 12, 2024 at 8:43
  • "Why would they waste their time verifying AI?" It's not clear if people would really waste their time. Depending on how much crap there is in general in GenAI content, it could speed up or slow down the whole process. If people in the end want to use that bot (as an inspiration or assistant maybe) or not, I can't say. An experiment would be able to show it. From experience I can say I'm often surprised what some people are willing to do. Biggest concern would be that people willing to verify AI aren't very capable.
    – NoDataDumpNoContribution
    Commented Dec 12, 2024 at 8:52
  • 4
    @NoDataDumpNoContribution And if you want get real results from such experiment, you need to run experiment on some site with sufficient traffic and experts willing to genuinely participate. And I wish SE good luck in finding that. at the end this will be forced upon Stack Overflow because experiment will not yield any results on other sites, or it will be deemed successful based on low traffic site where few users will verify something beyond their level of expertise.
    – Resistance Is Futile
    Commented Dec 12, 2024 at 9:34
  • "There is zero value in having AI generated answers on any site." -> If it can provide reasonable answers (which, having taken part in the preliminary trial, I can say with absolute confidence it can more than half the time) to all or some of the questions that nobody can bother answering (ie., most of them), then I'd almost say it might be worth more than most of the mod staff. "Anyone who wishes to use AI can go and easily ask AI." -> Generally not for free, I think.
    – goldilocks
    Commented Dec 12, 2024 at 19:13
  • 1
    @goldilocks If some site is so low traffic and lacks experts or really anyone who might have some clue to answer questions, do you really think having AI answers will encourage more people to stick around and answer on such site?
    – Resistance Is Futile
    Commented Dec 12, 2024 at 19:35
  • 1
    I could be wrong, but I don't believe there is any site with an an accepted answer rate of > 50%. The issue with a lack of expertise plagues them all simply because the demand for a free resource is always going to exceed the supply if that supply depends on other people investing time and energy -- the more they put in, the more people will ask for it. It generally takes less time to review someone else's answer than write your own, whether it came from an AI or not. Currently, this does seem to work, so the objection that there's no one willing or able to do that is fallacious.
    – goldilocks
    Commented Dec 12, 2024 at 20:21
  • 1
    @goldilocks Other people investing time and energy worked fine for Stack Exchange. Problem is that SE has been alienating its user base and AI answers will not magically solve that problem nor will it increase engagement. Handful of people willing to verify AI posts is not a solution either.
    – Resistance Is Futile
    Commented Dec 12, 2024 at 21:10
  • @NoData "Checking that an existing answer solves the problem is easier than coming up with the answer and checking it". That's often true, especially for programming problems where the answer is essentially a piece of code, and you have good test data that covers tricky corner cases, but it doesn't apply to all questions. If it's a "why" question, an LLM may generate a plausible sounding hallucination without useful citations, which is difficult to validate unless you're an expert in that topic.
    – PM 2Ring
    Commented Dec 12, 2024 at 21:17
  • @ResistanceIsFutile "Problem is that SE has been alienating its user base..." -> If this is supposed to explain why no site has an accepted answer rate anywhere close to 50%, that's a hallucination. There was never a golden age here where that was not true, and my explanation WRT supply and demand for free knowledge holds. You're drinking the Meta.SE echo chamber kool-aid.
    – goldilocks
    Commented Dec 12, 2024 at 21:51
  • 3
    @goldilocks Sure, and that's certainly much better than just spraying raw GenAI slop onto the front page, but as I said in an earlier comment on this page "it may be very difficult finding competent community members willing to do the necessary curation work". I don't mind checking answers written by humans because I enjoy teaching. I have zero desire to do unpaid validation work on GenAI output.
    – PM 2Ring
    Commented Dec 12, 2024 at 22:14
  • 2
    @goldilocks There are some questions that are never answered because nobody really knows an answer. AI will not know answer to such questions either. And nobody will be able to verify them. However, some of those may look plausible and may end up being published. Also not all questions are good enough to be answered. This is why we are curating questions and closing them. This is why we are deleting poor content. This is how you maintain quality and make it easier to those with problems to find high quality solutions easier.
    – Resistance Is Futile
    Commented Dec 12, 2024 at 22:34
5

(writing an answer post because it felt like I had enough miscellaneous things to say)

(a comment to staff): some of my related past thoughts: https://meta.stackexchange.com/a/388418/997587, https://meta.stackexchange.com/a/391147/997587.

(Answer Bot): I do things like: Suggest an answer that only some humans from the community can see until they verify, edit, and curate before sharing it with everyone else.

A new thought: Why would anyone with the expertise to verify these want to? They should just write an answer themselves. If they want to use an LLM for research before they write something of their own, there's nothing stopping them, and they can get rep from that. Also, please don't make reviewing these worth rep. Just think of the roboreviewers.

(Philippe): It's a safe assumption that LLMs [...] will likely get better over time

So you have the base understanding to share some of the same questions I have (linked above) about how this is going to work long term... and what is your answer to those questions? The lack of consideration I see from my end makes the statement "This experiment is about preparing for the future" sound ironic.

(Philippe): Already we see many examples of SE contributors utilizing LLM resources, and the added workload on moderators and curators that this creates

A bot posting LLM answers is not going to stop the people who irresponsibly post LLM answers because they want internet points.

Random, boring, technical questions that crossed my mind:

  • Is can_flag going to get switched to true? I can't imagine not being able to flag these.
  • Why did the api return a profile URL with groups in it, like https://webapps.stackexchange.com/users/groups/-2?
  • If extended discussion happens and a chatroom is created, does feeds get invited to the room? The feeds chat user has ID -2. Why not give this new bot its own unique ID?
Improve this answer
4
  • 3
    The interaction with chat is a funny one to think about. Different feeds users can be created with decrementing negative IDs, so there’s no negative ID they could use which wouldn’t risk colliding in chat, except for making it dramatically larger than any real IDs will ever be. On chat.SE the chat ids don’t need to be the same as real site IDs, but this would be problematic on SO itself… I guess it just doesn’t get a chat account, so that’s fine. These posts aren’t supposed to have the normal post UI so the move to chat behavior could also be suppressed.
    – Jeremy
    Commented Dec 18, 2024 at 3:44
  • @Jeremy. there are ways to work around that (joking)
    – starball
    Commented Dec 18, 2024 at 5:25
  • @Jeremy joking aside... maybe that's what /groups could help with
    – starball
    Commented Dec 18, 2024 at 5:33
  • Aaaaand the chat interaction might already be breaking things… meta.stackexchange.com/q/405222/134300
    – Jeremy
    Commented Dec 24, 2024 at 16:01

You must log in to answer this question.

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