2
$\begingroup$

Regarding this question, in light of this policy.


I think we can all agree in principle that AI generated text has no real place here in Worldbuilding, with the possible exception of queries about AI in a fictional world context, and now that we've had some examples of purportedly AI generated answers, I think it might be a good time to evaluate not how useful the policy is, but rather how effective the means are of enforcing it.


This answer was deleted by Monty Wild on the suspicion that it is AI generated. On the surface of it, that's just fine. Moderators have the duty and right to delete queries and responses that violate the policy. However. Just for laughs and giggles, I decided to try out a few of the AI text scanners available and found some contradictory findings. These are all commonly used detectors, and are used in other platforms where AI generated text content is frowned upon.


My concern here is not with the policy and not with its enforcement per se, but rather with the means by which the policy is enforced. How many detectors did Monty use? What should policy enforcement look like from the perspective of conflicting or inconclusive evidence? Going forward, how would cases such as this particular answer be dealt with given the nature of the evidence?


Dare I even bring up the considerations of possible ulterior motives (such as an experiment to see if a text can be tweaked sufficiently to either pass as human or deliberately return a false positive / false negative / false ambiguous)?


I went back to review the same author's other response to the same query, which was also deleted, and ran that text through all three detectors. That text came back as unanimously human generated.


Screenshots from the three detector sites I reviewed with the same text:

zerogpt.com returned a score of 0% AI

enter image description here

copyleaks.com returned an ambivalent result:

enter image description here

sapling.ai returned a score of 0% AI

enter image description here

INTERESTINGLY ENOUGH, I decided to run my own query here through the same three detectors and got a very slightly inconclusive result myself from one detector:

enter image description here

$\endgroup$
0

1 Answer 1

Reset to default
8
$\begingroup$

To answer this question, I can say that I used several methods of detecting AI content. Firstly, I used four different AI detection sites:

https://writer.com/ai-content-detector/

https://copyleaks.com/ai-content-detector

https://gptzero.me/

https://contentatscale.ai/ai-content-detector/

I agree with Elementilas that these are far from 100% accurate... more like 50% accurate at best. However, of the answers I deleted, several had "50% human" results, and some had "98% AI" results.

However, I use sites such as these as indicators, not as proof.

The one factor that I (and Stack Exchange) consider to be proof of AI content is the speed at which the answers are posted. I'm a pretty fast typist, but I cannot match the speed of AI, and the account that posted the AI content posted 20 answers in the space of a few hours. Long answers, some with around 700 words. Even typing at 60 words per minute, that means that an answer would take over 11 minutes to type, if the writer was typing at top speed, and in my experience at answering questions on WB SE, that doesn't happen. An answer like each of those should take around half an hour to create, and the user in question created 20 in just a few hours?

Inconcievable!

And yes, I do think that word means what I think it means, Inigo Montoya.

To elaborate further, I have seen other content on this site that is suspected to be AI-generated while acting as a moderator, content that has been detected as "100% human" by those AI detectors. However, since those answers were not posted along with a whole bunch of others at more or less the same time, and contained features not expected from AI, they haven't been deleted. Maybe they were AI, maybe they weren't, but I have to give the user the benefit of the doubt where doubt exists.

$\endgroup$
10
  • 1
    $\begingroup$ So, the deciding factor is speed rather than content. Understood! I do appreciate benefit of the doubt! It's always best to assume the best in others! $\endgroup$
    – elemtilas
    Commented Aug 26, 2023 at 14:45
  • $\begingroup$ Not just speed, but other factors too. I'll leave those to your imagination, @elemtilas $\endgroup$
    – Monty Wild Mod
    Commented Aug 26, 2023 at 14:46
  • $\begingroup$ Well, this is a policy. Don't rely on my imagination! Things could become rather Wild! I am curious now about the speed issue: were all 20 of those speedy answers given to relatively new questions? E.g., question was posted at 10:34 and a 700 word answer was posted at 10:36 kind of thing? Or were they questions asked within the last few days / weeks and 20 responses were posted within a few minutes of each other? $\endgroup$
    – elemtilas
    Commented Aug 26, 2023 at 14:48
  • 3
    $\begingroup$ I've shared my findings with the other mods. If they feel that they're worth sharing with the other users, then we'll share them. There's no point giving the cheaters everything they need to cheat with impunity.... $\endgroup$
    – Monty Wild Mod
    Commented Aug 26, 2023 at 14:51
  • $\begingroup$ Fair enough! I think you've satisfied my question! $\endgroup$
    – elemtilas
    Commented Aug 26, 2023 at 14:52
  • 3
    $\begingroup$ I'll add that I agree with everyone Monty did and has written. Overall, like mods across the network, I think we're fairly reticent to share the details of the heuristics that help us identify AI-generated answers, as that would make it easier for folks to avoid the ban -- there are a number of heuristics that are intentionally kept private. I think this answer is probably at the highest level of detail I'd be comfortable sharing (don't want to speak for the others, though). $\endgroup$
    – HDE 226868 Mod
    Commented Aug 26, 2023 at 16:37
  • 1
    $\begingroup$ Thanks for posting the response, Monty - was wondering why you had deleted the only substantive answer (at that time) to one of your own questions. Makes sense knowing that it was part of a pattern of behaviour. $\endgroup$
    – KerrAvon2055
    Commented Aug 29, 2023 at 23:33
  • 1
    $\begingroup$ +1 and I can't agree more that posting speed is a clever detection method. I type 90 words a minute with < 1% errors - but an answer like the one Elemtilas links to would take me hours to write because it takes time to come up with good ideas and more time to express those ideas well. Posting speed is the most clever idea I've heard yet for detecting AI-generated content. $\endgroup$
    – JBH
    Commented Sep 6, 2023 at 4:37
  • $\begingroup$ As I was both the author of the original question, and the person who reported the answer, I am actually a bit confused why ai content detection is used to determine if an answer was written by AI, instead of just using the (to me) much more obvious method of just putting the answer into chat gpt and seeing what it spits out. Which, in this case, was and still is, basically the exact content of the deleted answer. Not saying that it is a better way to do it, just curious why this wouldn't be the first thing to try? $\endgroup$
    – Amianas
    Commented Sep 8, 2023 at 9:56
  • $\begingroup$ @Amianas I don't have (or really want) a ChatGPT account, and I wasn't aware that it would output an identical response every time to a given input. Having played around with other AIs, they haven't been so consistent. $\endgroup$
    – Monty Wild Mod
    Commented Sep 8, 2023 at 12:39

You must log in to answer this question.

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