Timeline for GPT on the platform: Data, actions, and outcomes
Current License: CC BY-SA 4.0
17 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 8, 2023 at 19:50 | comment | added | Paŭlo Ebermann | @Gilles'SO-stopbeingevil' – what do I use a high-rep SO profile for? Pretend I'm a qualified developer and get a good-paying job? Post spam that won't be deleted? | |
| Jun 15, 2023 at 20:18 | comment | added | Gilles 'SO- stop being evil' | @Braiam It only takes one person to automate the process. Then a large pool of unqualified people can do the grunt work. This happens whenver there's money at stake. I confess that I don't know what money is at stake at the moment in pushing AI-generated content on SO: highish-rep SO profiles aren't that much of a market. | |
| Jun 15, 2023 at 18:52 | comment | added | Braiam | For this theory to be true, it needs to presume that the answerers that are posting GPT generated answers are lazy enough to "just ask chatgpt and paste the answer", but diligent enough to automatize the process. This kind of presumption suppose a very high bar to clear to even demonstrate. Most people give up posting at the first hurdle, and this is something that even SE can observe. | |
| Jun 9, 2023 at 6:43 | comment | added | NoDataDumpNoContribution | I like to think of spammers as basically people like you or me that simply have a ChatGPT tab open and copy and paste from there. My first assumption would be that their behavior is simple and does not change over time. But of course it's possible that I'm too naive there and rep farming by posting GPT output on SO is a highly (profitable?) organized crime nowadays. Maybe there are any other indications for that? | |
| Jun 8, 2023 at 20:50 | comment | added | Frédéric Hamidi | @Gilles, I somewhat feel relieved that the Chrome extension you speak of does not work in dark mode for some reason. Maybe the extension itself was written from "AI" input. | |
| Jun 8, 2023 at 14:48 | comment | added | ꓢPArcheon | Upvoted only because you went as far as to verify what I had only guessed: obviously users of chatGPT quickly catch up on the fact that posting an answer too fast is a clear tell-by and developed countermeasures that undermine any assumption the company made in their study. | |
| Jun 8, 2023 at 9:48 | comment | added | Gilles 'SO- stop being evil' | @AJM They'd be false positives if the GPT rate was considered to be the low-draft rate, but it's not. There are many non-GPT posts that don't make drafts for various reasons: very short posts, posts made from older browsers where the draft code doesn't work, etc. We'd expect that rate to be mostly constant over time, so the changes to that rate are likely related to GPT. The idea is that the baseline is from before Dec 2022, then there's a GPT spike, and then the situation is with GPT posters whose behavior may or may not have changed over time. | |
| Jun 8, 2023 at 9:16 | comment | added | AJM | In terms of saving drafts... one thing I had considered recently was using a text editor to work on my answers offline, then pasting the finished product into the edit box. If there's anyone out there who does it that way, aren't they going to be a GPT false positive due to their own lower number of drafts? Sure, there'll be some saved as they correct errors in the Markdown etc. but even so. | |
| Jun 8, 2023 at 2:14 | comment | added | 0-One-0 | What about formatting? I don't use CGPT but from what I see, answers contain formatting and code boxes with a copy code button. Can you do a ctrl-A/ctrl-C and capture the whole text+code in a format that works in meta-wiki directly? How many formatting steps are required if you're trying to reproduce the formatting provided by the AI? Also, does that vary whether you're inside the openAI interface or some other interface, such as Bing? I don't know, but it may be people still want nice looking documents despite using the AI? Some topics might generate more formatting? | |
| Jun 7, 2023 at 23:48 | comment | added | dxiv | @Kevin "I imagine many of them to be unsophisticated users who are playing with a shiny new toy" -- I have seen several on MSE lately who became more sophisticated in disguising GPT contributions. Not all of them match the stereotype of new-account-dozen-answers-then-gone, and I don't know what their end game is, but it certainly happens nowadays. | |
| Jun 7, 2023 at 22:11 | comment | added | Gilles 'SO- stop being evil' | @Kevin The shiny new toy was the Dec 2022 spike. Now we're seeing the second wave who aren't just playing. | |
| Jun 7, 2023 at 22:10 | comment | added | Gilles 'SO- stop being evil' | There is a market for high-reputation SO accounts, although it's a very small market compared to buying political influence on Facebook and Twitter. In the 2010s, for SO, the main methods were: gaining access to an abandoned account (presumably by accessing the email account), voting rings (but I think those are efficiently combatted so it's never really worked), and posting plagiarized content (which still requires effort and is easy to verify). I expect GPT to replace plagiarism because it's so much more effort to filter. The incentive is there to develop efficient tools for GPT posting. | |
| Jun 7, 2023 at 22:07 | comment | added | Kevin | @Gilles'SO-stopbeingevil': While I think that is entirely plausible, I prefer an explanation that doesn't require significant coordination on the part of the attackers, because I imagine many of them to be unsophisticated users who are playing with a shiny new toy. I'm sure some of them are organized as you describe, but I would tend to expect that some of them are just not sophisticated enough to reason about how CAPTCHAs work. | |
| Jun 7, 2023 at 22:05 | comment | added | Gilles 'SO- stop being evil' | @Kevin The first thing I would try is direct copypaste and then I'd run into the captcha. The second thing I'd try is to avoid running into the captcha, because I know captcha makers are going to try to break whatever captcha breaker I might use. And for that simulating a human's typing is a very natural solution. I don't know if GPT posters are doing that, but to me it's a more plausible hypothesis than moderators having suddenly dramatically increased their rate of unwarranted deletions. | |
| Jun 7, 2023 at 22:00 | comment | added | Kevin | @41686d6564standsw.Palestine: IMHO the most straightforward explanation is that the GPT users are editing a word here, a word there, etc., to try and make it sound a bit less like an AI, and incidentally generating drafts. If I were trying to attack SO in this manner, that's literally the first thing I would try. | |
| Jun 7, 2023 at 21:04 | comment | added | 41686d6564 | I mostly agree with your conclusion but unless I'm missing something, the demo of that extension shows the GPT output being printed in real time (not in the answer box) in the same way GPT prints its output in chat mode. So, this doesn't affect the number of drafts. That being said, this doesn't mean that there are no other tools, tutorials, etc. that allows users to avoid captchas, bypass GPT detection systems, and so on. | |
| Jun 7, 2023 at 20:57 | history | answered | Gilles 'SO- stop being evil' | CC BY-SA 4.0 |