AI-generated Answers experiment on Stack Exchange sites | Hacker News Reader

AI-generated Answers experiment on Stack Exchange sites

meta.stackexchange.com

74 pointsnsoonhui61 comments

Any site like Stack Exchange is sure to rely on a handful of really dedicated users who live for their subject of special interest. It is an ingenious way of extracting value from the free work of these users.

But answering question and teaching strangers about your favourite subject is nothing like rating computer generated sentences on their correctness, no matter what the subject matter is. The whales will leave, and the community will be a shoal of drone fish. There may be a business model there too, but it's nothing like the current one.

These people know their community. So I wonder what thinking was behind this decision. In my current place of work, there have been many discussions like "how can we integrate AI in our business model". Maybe it's not more complicated than that. Maybe it's a coincidence that they were recently bought by a company that needs return on investment. Either way it's sad that we so rarely can have nice things on the Internet anymore.

In other words, the web 2.0 business model is to monetize the free work of other people. If there is a web 4.0 (let’s skip 3 for clarity) its business model is to monetize the content of web 2.0 companies.

Legality aside, it’s hard to escape the irony.

Quora has been doing it for a while.

But honestly everything about Quora is... ass (from interface to answer quality to moderation). The starting point is so low that I feel AI answer bot is a net positive for them.

I won't say the same for SE.

The Quora AI answers I've seen are often just outright wrong. I almost wish Quora would use Stack Overflows's AI approach, and that Stack Overflow would just stick with human answers only

The fact that quora hasn't been marked as spam years ago and downgraded in search results is what tells me google never had impartial ranking. It's got to be the worst website on the internet. Its login walls alone would've downranked any other website. The wrong answers by pretend experts could also have done it. But not even the AI slop. I sometimes wonder if quora's leadership is someone from google's nephew or something.

Note that the founder and CEO of Quora is on the OpenAI board, which may explain why they’ve been doing it for a while. Interestingly, he is the only one of the board members that voted to remove Sam Altman and didn’t resign, and is still there.

https://en.wikipedia.org/wiki/Adam_D%27Angelo

He has his own Poe AI or some shit. It’s probably garbage and he tried to whack the competition.

Quora was the first website I blocked when I started using Kagi.

Life is much better now without it dominating my search results.

Haha, right, I was like “wow that site still exists?” when I read that and then I remembered I had it blocked on Kagi.

I've never met anyone who's used Quora, nor have I seen it mentioned or linked to anywhere except inserting itself here and there into search results. Who is using Quora?

Quora had some really good content when I first started using it (maybe 10 years ago). Some really interesting people at the top of their field would post thoughtful things there.

Then after a couple of years I think it changed hands. The dark patterns started (hassling to install the app, hiding content if you weren't logged in, etc) and it all went downhill pretty quickly. I would guess that a lot of those posters migrated here, or to their respective branch on the SE network, or who knows where else

Quora is so bad when someone first pointed me at something in Quora I genuinely thought it was some sort of joke website that I just wasn't understanding properly. Believe it or not, within some parts of the SV VC bubble they take Quora really seriously and everyone acts as though it's actually good.

Quora is worse than Yahoo Answers. Yahoo Answers was at least funny some of the time (e.g. "am I gregnant?")

What's the point of having AI-generated answers on Stackoverflow when I can just ask the AI to begin with ? There is no added value here, especially given that the AI is trained on SO data to begin with.

It's a weird attempt to keep users on the site. Plus those AI generated pages will show up on google, when otherwise there might be nothing or the pages are being downranked due to being old.

Google promotes AI slop above everything else, so may as well get in on the grift before it's too late.

I think it might also be an attempt to extract some final value from the users by having them train AI. Quality RLHF data is expensive, and if you can get experts to do it for free...

The added value would be humans upvoting helpful answers.

So the highly upvoted ai answer on SO would not be a hallucination.

Maybe not... You have a lot of faith in SO upvotes. I think they're very likely to upvote good sounding answers which don't actually work.

Not sure about others, but if I notice it's an AI answer, I'll likely stop reading, as such an answer has no added value to me - as others have pointed out, I can simply get that answer myself, and I likely have before venturing to SO.

And if I don't read through an answer, I'll also not upvote it.

The added value is that it is seen by all and is "free". The broadcast all and voting is quite valuable.

The deliberations / clarifications / back-and-forth that used to happen on stackoverflow are now happening with LLMs. That data is now lost forever and is now proprietary, siloed.

Isn’t that already the case since the rise of closed/hidden discord channels(trying to replicate IRC era community)?

Two wrongs don't make a right

Take it the other way. What is the point of a human q/a site when you can get good enough answers from an AI? It's their business model that is threatened, and they need to pivot in order to survive.

The thing is, you currently can't. Sure an AI can answer relatively complex technical questions, but sometimes doesn't do it properly. And when it fails, you don't know. I had cases where AI was clearly inventing calls to APIs that didn't actually exist although they were legitimately looking; such content would just be downvoted enough on SE that you'd be very careful before even considering it. If I ask an AI, I might either get a good answer or a bad one, but without a vote count nearby telling me how good it might actually be.

> And when it fails, you don't know.

This can be addressed in some cases with actual code verification. It’s how DeepSeek was trained. But it is hard in the general case.

What if you're not in the "good enough" business?

I've tried many times to get an answer from these AI engines. They have always been just "good enough", to help me find a direction for more research to get a good answer, or flat out wrong.

The paid version of ChatGPT has had built in web search for over two years. It’s easy enough to just say “verify your answer” and it will search the web.

Even with code - albeit most of my experience is with having it write AWS automation scripts in Python using the AWS SDK (Boto3), I’ll either give it the link to the API in question or tell it to “verify Boto3 calls on the web”.

The worse thing it usually does is have code where it wants you to include your credentials in code (don’t do that). But even most people on SO do the same

It gets harder though. Even before AI, a lot of the low hanging fruit “how do I sleep for 2 seconds in JavaScript” are already answered. Beginner Qs are often marked as duplicate. So you’re left with long tail difficult or obscure questions that take a lot more effort to answer.

I'm not a user of Stack Overflow, but I am contributor to (and avid readerof ) the network in general.

What's amazing is that you can ask a question at any level of difficulty (high school through research) and get an answer from one of the top researchers in the world in their field.

That's going to go away if this goes through.

Maria del Rio-Chanona, Nadzeya Laurentsyeva and Johannes Wachs reported an interesting trend last year and found a noticeable impact on user contributions after the introduction of some now widely-known systems[0]. I wonder what the continuing trend will be if the response to a decrease in user participation is introducing additional measures that will likely take away incentives to contribute. A 'redusage' trend, compared to the early years of these knowledge production communities, whose collaborative works now provide much of the training data.

[0]: https://academic.oup.com/pnasnexus/article/3/9/pgae400/77548...

Am I alone in pretty much having completely stopped using stackoverflow in the past year?

Even before AI came out I was having problems with using it due to their community having turned into an insular group that was against the Outsider (e.x. having my questions closed as a duplicate of another question despite my question have quoted the original and pointing out the difference in my situation, not being able to get perms to do anything useful without having earned karma you cant get because the established players keep everyone else out, etc)

Now though I just end up reading docs on the tool in question or asking ChatGPT. In the times where ChatGPT was wrong or hallucinating, I never found the right answer on stackoverflow after.

Oh and thats beyond the seemingly endless amount of copycats who look like they are just scraping stackoverflow and then reposting with better SEO? Not entirely sure what it is but for a while I kept finding websites with duplicate down to the typo answers, as stackoverflow

Insular communities seem to be the fate of sites that put a strong emphasis on voting as a way of moderating content. SO gets scraped quite a bit. It's easy to do and even if the knockoff site only gets 0.1% of SO's traffic, it's worthwhile. Being text heavy means it's easy to host and gets good SEO.

I found it increasingly annoying that almost every time I searched for some variation of "what is a good tool/library/device for X" there was a SO thread high in the search results, but of course the question was closed (often with some great, but old, answers) because they do not accept questions like that. They could have had a lot of more happy users if they figured out a way to support that kind of questions somehow.

Someone actually took the time to close my 16 years old question of that type. There wasn't any activity on it, and it had good answers. Some people are just bored.

I can't quite remember, but don't you get mod points for actions like that. Someone looking to increase their mod level, or boost their points (is it useful on a CV [Resumé]?), might go chasing older questions.

Also, they have a really annoying "community bot" that revives v.old questions into the new question queue... it tricked me a few times, then I started to check it wasn't a community bot question before looking further (pretty moronic and anti-answerer behaviour imo).

You don't. Assuming that by "mod points" you mean badges, no badge listed on https://stackoverflow.com/help/badges can be earned by closing questions.

I've just had the same experience: last month someone decided it was a good idea to close a question of mine from 2010 because it was "opinion-based", asking me to "Update the question so it can be answered with facts and citations. This will help others answer the question". The question, other than being 15 years old, had more than one thousand votes, 15 answers with positive score and a bunch of low quality answers. The question is still up on the site, just nobody can answer on it, but it feels hostile to me (and to the people who took the time to contribute answers) and made to accommodate the wishes of someone who cares about tidying up the place more than anything else.

I purposely deleted all of my Stack Overflow accounts for this reason. Tired of that site.

> figured out a way to support that

Seriously, sites need to stop fighting their users. If people are constantly asking for that content, there is a need for it.

The data shows that you are not alone. AI is decimating StackOverflow.

The draconian approach to keeping the site clean made some sense before AI, because they used to have tons of really, really, really low-effort questions. ("do my homework", basically.) But now those go to AI....

I have a Kiwix archive of all of StackOverflow from shortly before ChatGPT launched. I suspect that's all I'll ever need, and even that will become less useful as time passes.

Agreed from a software engineering perspective. I can't imagine who still uses stack overflow.

However I'm still in the phase where I think of opening a browser before realising my IDE with it's LLM is a better first choice.

My usage is more in line with how it was originally conceived, I search engines as an interface to Stack Overflow. If I have a problem, I do a search, if it lands me on Stack Overflow, then that's where I get my information.

The really good answers are getting harder to find, so I see Stack Overflow less and less, but I disagree with those who comments that LLMs have taken the role of primary source of information. I don't use ChatGPT, CoPilot, or any other LLM and I also seen a sharp reduction in the amount of times I find myself on Stack Overflow. Depending on what I do, my go to is either the official documentation or source code. Only exception is probably C, where I will frequently go to the OpenBSD man pages first, as they are really well written, and then adapt if something doesn't line up with glibc.

I still use stackoverflow through searching for very obscure things that got answered a while ago and a question kept alive and relevant through people doing good Samaritan work and updating answers to the current year.

Each time I tried to get an answer from any LLM, the thing was always deeply wrong or obviously flawed. Mind you I am always searching for specific, niche, really crappy things that are often an overlap of SW and HW...

But again this is very cherry picked opinion from my personal use.

Generally SO sites are my fallback after asking AI, and my way to know that the working solution I got is not using outdated or suboptimal features.

An actual helper would be good imo: it could

* Add markup * Suggest other answers/questions to consider * Suggest web sources that seem most relevant * Say if the answer appears to cover the whole question

But, it should not answer the question. It should be an adjunct.

I hope this goes better than the PhysicsForum admin flooding the forum with AI generated answers post hoc to unanswered posts.

I'm quickly developing a large dissatisfaction with "AI" and it has escalated violently in the past weeks. I'm tired of it. LLMs are a great innovation with so many amazing applications and the idiots sets of to use them to generate empty content, scams, remove artists from their livelihood, anything that will maximize profit, without even a second thought to the potential damage.

The straw that broke me was after a call with my dad the other day, he wanted help to identify a scam (an AI assisted scam obviously). After that call I've come to realise that all trust is quickly being removed from society. The internet is pointless, because most trust has been removed. We wheren't doing to great before, but LLMs have just accelerated everything. Why would I ever recommend anyone use the internet for anything that requires a minimum of trust? Customer service is screwed, because more and more people aren't going trust anything that's not in person anymore.

You can't order online, because anything that not a small local shop is likely a scam, Amazon included. You can't get customer service, because that's either a scam or a bot which can't help you beyond the most trivial answers that could be found if the company had actually bothered to write documentation. You can't trust that whatever is written on a page isn't actually just an AI inventing things that didn't happen. Videos are AI voiceovers of AI written Reddit posts, which AI generated images playing.

It's not that there isn't good stuff. Some people are creating the most amazing works of art and engineering, but the AI, scams and corrupt capitalism just drowns it out.

The future sucks and I'd very much like to off this clown bus.

Y'know what might be good: instead of showing the AI answer as a regular one, like the cesspool that is Quora, show the answer only to the poster with a CTA to accept or reject the answer.

If it's accepted, it's posted.

Not only would this be more user-friendly preventing random slop but would probably allow them to collect a non-negligible amount of feedback for training.

Guess now is a good time to hoard some recent dumps. What would be some good local tools to browse those?

https://archive.org/details/stackexchange

This nightmare seems to have hit MathOverflow as well:

https://meta.mathoverflow.net/questions/6039/has-mo-been-red...

I cannot emphasize enough what a blow this is to the online discourse on research mathematics.

With search engine results polluted with "AI slop" (often long-winded and useless articles, not answering the questions; a relatively advanced SEO spam), I notice that Stack Exchange websites are among the good ones now: if there are answers, they tend to be decent. It is disappointing to see it heading that way, though commercial projects tend to fall into something like that (if not slop, then ads, feature creep, UI bloat/annoyances/captchas, "enterprise" and any hyped stuff, paywalls). SE had its issues before as well, but even if it will become unusable, there is a potential for better Q&A places in the future, maybe federated ones. A quick search did not reveal any projects like that though.

At -256 ....

-257, I just added my [down]vote.

Flagged as rude, too.

Not just that, but +20/-282 is the up/down split

I'm afraid now that SE is using AI to generate answers and relying on human to upvote/downvote, in the future AI bots will also be used to upvote/downvote answers. Going a step further, it's not inconceivable that AI will be used to seed a Q&A site with questions ( who needs to manually seed a site ALA Reddit early style?).

And then we reach a point where a Q&A site that is seeded with AI questions, answered by AI answerers, and reacted by AI bots. Even comments aren't spared.

The AI interacted Q&A sites will then be indexed and searchable via Google, to be consumed by human and bots alike, or maybe even presented as authoritative information. This is truly an Internet dystopian.

Next, paid answer by real human.

How would you prove it's not AI?

...aaand Stack Exchange has just lost a lot of its value. I mean, who is going to curate those answers? Even wrong answers were eventually ranked down.

Reply on news.ycombinator.com