The consequences of generative AI for online knowledge communities
nature.com
nature.com
This equates declining activity with degraded community, which I don’t think is a valid assumption to make.
Taking Stack Overflow as an example, LLMs should be decimating question volumes there. That’s the whole point of LLMs! To answer questions! But which questions are being eliminated? The boring, rote ones that can be answered by a glance at the documentation? Or the interesting ones that need expert knowledge and insight? My guess is that LLMs displace the majority of the former but less of the latter. So what’s left once you take the lazy spam questions away from Stack Overflow? I would say that if LLMs can act as first-line support for those questions instead of expecting humans to answer them manually, the result will be a healthier community that more experts are willing to participate in. And sure enough, the paper reports:
> upon ChatGPT’s release, a systematic rise began to take place, such that users were increasingly likely to be more established, older accounts. The implication of this result is that newer user accounts became systematically less likely to participate in the Stack Overflow community after ChatGPT became available. Figure 6 depicts the effects, indicating that questions exhibited a systematic rise in complexity following the release of ChatGPT.
That looks like an improvement to me, not degradation.
The other extreme - drowning in "noob" questions - is also harmful, which is why you end up with FAQs and rules. But there are human dynamics that underpin all these communities, and LLMs are undoubtedly destroying the incentives to participate.
> The other extreme - drowning in "noob" questions - is also harmful, which is why you end up with FAQs and rules.
This isn’t reflective of what actually happened with Stack Overflow though. These are not equal dangers. There were plenty of reasonable questions, then it became overwhelmed by people who didn’t want to read tutorials or documentation but instead wanted other people to do their work for them or to have the documentation looked up and recited to them by a person not a search engine. Stack Overflow didn’t much care because it’s increased engagement, but the expert users absolutely did care and started to leave.
“There aren’t enough interesting questions” is not a danger. “Experts don’t want to be a flesh interface to a search engine” is not only a danger but has already been harming the community for a long time. LLMs have the chance to reverse this damage, and they can do this by selectively reducing the volume of the harmful engagement. So I am rejecting the notion that declining activity necessarily means degraded community.
As such, they're certainly still useful for answering questions for which StackOverflow has a dozen duplicates, and that may very well be a net positive for StackOverflow.
Here's an example of something that came up in my work the other day - fiddling with prompts and checking the responses is ultimately more work than just writing the code from scratch. GPT4 and Sonnet 3.5 never came close to a correct response, even though the code always looks sort of "directionally correct".
https://chatgpt.com/share/ba85ced9-63d6-4fb3-8dc1-b8efbaa58f...
And that's despite this almost certainly being somewhere on Github in some form.
I think this kind of opinionated community will still thrive because LLMs are great for quick answers but don't yet provide strong opinions, and likely won't for a while due to regulations. In that way, having a close-knit human community with strong opinions on a subject will become even more valuable.
In either case LLM are really a steam shovel technological moment for the shill marketing world. What used to take active posting and constant maintenance of your bank of canned responses can now be done with one operator running a script that simulates all these active posters with enough variance to not have the responses seem to be from the same bank of material. You could dominate an entire subreddit, have 1000 of your accounts to 1 user putting out enough AI derived signal that even the few real users who post start regurgitating it.
Here's one I asked chatGPT to generate about a hypothetical video game (related to a real one as source for what to reference):
> hypothetical video game is sooo fggin trash. like, the parkour is meh, combat is a complete joke, and the whole game just feels off. seriously, what a waste of time.
People only take safety in opinionated communities with the assumption everyone else is a human being offering their sincere take, and due to this assumption it's also much easier to get your bot to "fit in", pushing whatever narrative you want or letting it go wild.
For example, I’m sure you could get an LLM to write about how delicious peanut butter and pickles is as a combination, but most people do not agree.
Reddit doesn't even pass the eye test. Look at /r/programming. over 6 million subscribed accounts, but the activity is like barely anyone is there. The top three posts of the last 24hrs currently have 72, 218, and 39 comments respectively, and after that engagement falls off to like 0-5 comments a post. So we have around 350 comments a day on a board supposedly subscribed by 6 million account holders. The comment making rate of this community is apparently 0.006%, and a good deal of that is liable to be bot traffic.
Dead internet theory is probably right. I should pack up and leave with the rest too I guess.
Now that we have very accessible LLM, that same trend can be artificially created with ease. You can dominate entire subreddits with artificial content, and nobody cares as long as it isn't a giant glaring outlier. You can make people think a certain way, act a certain way, buy certain things, just as long as you seem human enough to pass the sniff test and that a large enough group of "people" were doing so.
These opinionated communities will not survive much longer without implementing more severe human ID verification measures. You can bet against that, but I see the writing on the wall in my smaller public communities. They already don't trust much that isn't coming from someone in real time video.
Hiring just became much harder
Also it's not just knowledge - Insta is rapidly turning into an AI trashfire. AI "digital creators" are taking other people's photos, face swapping another face (AI or another human not sure) and posting it as their own. Very Déjà vu vibes sometimes if you see both versions.
While you do need a constant funnel of new users to sustain a community, I think the end result of this will be a higher quality funnel that will eventually create an environment of people solving niche problems.
All that to say that I now do a first pass through a tuned model and only go to stackoverflow as a last resort. I skip Medium search results because they are mostly people shilling their company.
> [A]ctivity in Reddit communities shows no evidence of decline, suggesting the importance of social fabric as a buffer against the community-degrading effects of LLMs.
This seems rather disingenuous? People go to StackOverflow primarily to find answers to technical questions, usually questions where there is 1 or a small number of valid answers. Reddit, on the other hand, is a place where people often go to share their own opinions/posts/etc, or go to browse not looking for anything in particular. The decline in SO viewership makes sense with the rise of ChatGPT, but I wouldn't expect that to cause any equivalent decline from Reddit. If anything, this looks to me like evidence to the contrary: that LLMs don't degrade communities.
This is a real usage of AI that is there to stay IMO.
What if junior people are able to use LLMs as sparring partners and learn without wasting time of more senior engineers? That leaves more time for online communities to discuss things that are on the edge of what’s known that need worked on.