In a recent statistics post, the company reported on a large false positive detection rate of a specific GPT content detection tool when applied to old (pre 2023) content. Later in that post they introduced a measure they think represents the amount of copy-pasted AI generated content, the ratio of large to small draft saves, and emphasized that this measure cannot be used to reliably inform about the state of a single post but can be used to estimate the average rate of such content.

That made me thinking if there are maybe more measures, that can be used to estimate the average rate of GPT content since November 2022?

For possible candidates one would basically search for anything that changed significantly around that time and cannot be explained otherwise.

I came up with one example: maybe the average length of a copy-pasted GPT answer is different from the average length of a human written answer. I checked this by calculating the average length of an answer for the last three years on Stack Overflow and saw that indeed starting with November 2022 the answer length of SO increased and is now at roughly 15% longer than it before while before it was relatively stable. The fluctuations in the data are much smaller than the observed effect.

average length of an answer on StackOverflow for the last three years by week

However, I'm not completely satisfied. I could not calculate the average length of deleted answers (in PostsWithDeleted field Body is empty) so this effect would only be for still existing answers and I wanted to calculate percentiles (25%, 50%, 75%) but that proved slightly too difficult for me in SQL. Also when looking at other StackExchanges like SuperUser or Mathematics, I did not see such an effect. It could simply be that in 2023 experts on Stack Overflow decided to provide more thorough answers.

Are there maybe other unexplained changes around that time (number of edits, comments, timings between answers by same author, occurrence of certain phrases in post bodies, ...)?

Maybe the company can help there too, because for deleted posts some data is removed in SEDE but the deleted posts are presumably rich in GPT content and would be nice to analyze too.

  • 2
    Draft saves and revisions can be gamed. One measure that's extremely hard to fake is the amount of responses from the answerer after an answer attracts comments. Now of course the best answers don't need comments (for clarification/correction) but I've found that users that post AI-generated answers (as confirmed later) hardly ever respond to comments. Jun 28 at 19:54
  • 2
    @GertArnoldisonstrike eh, i've seen quite a few that will feed teh comments back into gpt and use it again to post responses
    – Kevin B
    Jun 28 at 19:55
  • 1
    @KevinB Can be done of course. It's just not my experience. Maybe in most cases it requires too much understanding of the gpt response to condense it into something that fits into a comment or improve the answer, or a more than trivial ability to prompt gpt to write something concise (against its habit). I think it could be a contributing measure, certainly not the only one. Jun 28 at 20:02
  • 6
    Your finding that there's likely a substantial amount of AI-generated content on SO that has not been deleted is consistent with a very preliminary look at the same thing using some other metrics. I'm not prepared to go into details, at this time. In fact, I'm not the one doing the actual work, but very preliminary numbers indicate there's quite a bit of AI generated content still undeleted on SO. Unfortunately, as you've found, it's quite difficult to get a better overall picture, due to the fact that deleted posts are not included in the data.
    – Makyen
    Jun 28 at 23:21
  • 2
    Perhaps typing speed of new users? (E.g., the average number of characters in answers written by accounts created <1 hour before posting.) Jun 29 at 1:29
  • 2
    @RebeccaJ.Stones IMX, most ChatGPT answers were coming from accounts that were established but which had sat idle for months (or years) before suddenly and rapidly starting up again. Jun 29 at 6:42
  • Is a long answer better? Did the number of quality answers go up or down, as measured by votes? Jun 30 at 7:19
  • 2
    @user10186832 The problem is that up/down-voting is not a good indicator of actual quality for AI-generated content. That's actually a major part of the problem with AI-generated content on the SE platform. AI-generated content is, nearly universally, written very confidently and is "eloquent bullshit". It requires much more time to evaluate with respect to a quality metric and often requires an actual subject mater expert, or at least someone already familiar with the issue at hand in order to accurately evaluate the real quality and accuracy of AI-generated content.
    – Makyen
    Jun 30 at 15:56
  • 3
    I have often seen AI-generated answers upvoted (assumed to include upvoting by the asker), accepted, and even with bounties awarded to them by the asker, and then a comment placed on the answer by the asker saying something equivalent to "this doesn't work". AI-generated content bypasses a large portion of the metrics which we humans have learned to associate with "bad quality" and is specifically designed" to appear as "good quality" to a human reading it. Appearing as "good quality" on the surface is the *primary purpose of large language model AI generation. That's genAI's whole point.
    – Makyen
    Jun 30 at 15:56
  • 1
    @user10186832 Votes aren't a good indicator of quality. If simply less people vote the quality wouldn't go down. Jun 30 at 16:53
  • 1
    @Makyen "the metrics which we humans have learned to associate with "bad quality"" We surely have to relearn in this regard because in the future there will be more and more "eloquent bullshit" at least outside of SE if not inside. We shouldn't upvote something just because it reads well but rather abstain from voting if we cannot evaluate the content. Jun 30 at 17:00
  • 1
    @NoDataDumpNoContribution I've often seen answers upvoted, accepted and them commented with 'it doesn't even work' long before ChatGPT, so it's not the new problem. Thanks to ChatGPT more and more people stop prenteding not being able to notice the problem. Jul 4 at 13:10
  • @Makyen things about voting are even worse if you take into account how system handles votes to the same account. Three rapid downvotes to same user get quickly reverted, I've seen it myself several years ago. Note that upvotes are handled differently, I saw it myself how 5 to 7 blatantly serial upvotes were totally ignored by reversal script until mod intervened manually much later. This looks like a blessing for a user quickly dumping multiple generated senseless "answers"
    – gnat
    Jul 6 at 11:18


You must log in to answer this question.

Browse other questions tagged .