Give mods access to the draft count and pasted content of answer posts to help them detect ChatGPT answers

Question

It's been pointed out to me by several people now that giving moderators access to the content of what people paste is bad for privacy. I initially disagreed (and still hold some thoughts of disagreement), but SPArcheon has convinced me. Rather than invalidate the answer posts that have given this feature request valid criticism by editing the feature request into something different, I'll just leave it here as is, and take some time to think of an alternative without the privacy problems, which I'll put in a new post later. Let the following be somewhat of a "cautionary tale":

---

Quoting Slate from their answer to "Ban ChatGPT network-wide":

if any site experiences a volume of GPT posts that are cumbersome to manage, or a site needs any other support managing an influx of unwanted content, we are always happy to help apply the tools we have at our disposal.

It's intuitive that people would start trying to evade detection by editing their answers that they pasted from ChatGPT (several people have pointed this out in the answers to GPT on the platform: Data, actions, and outcomes, including Gilles, me, Kevin, tripleee, NotTheDr01ds, markalex, and more- probably).

If you're happy to help as you've stated, please give moderators the following information for answer posts:

• The draft count

  Time taken to write each revision would also be nice.

• All content pasted into the post body input for each revision

  It doesn't matter if the content has been edited after pasting (that's the whole point of this feature request), or completely removed (I bet that's hard to implement, and it doesn't really matter anyway)

  If we're concerned about user-safety with accidental pastes of sensitive information ^{_{(I see this as an edge case, and personally think mods can be trusted with this information if they can be trusted to redact things from post revisions)}}, then we can significantly narrow down the exposed surface area by making it so at least one user needs to flag it with a custom reason saying "chatgpt" (or similar) in it. If a moderator wants to investigate something on their own, they'd need to ask another user such as another moderator to raise such a flag (and presumably give a reason for asking) before being privy to that info—basically, requiring two users and at least one being a moderator agreeing that there's a good reason to want to see that info.

  If that's still considered too "privacy invasive" (I don't think it is), then I think we can at least settle for showing the length in characters of each draft.

Some sites could really use these tools. Stack Overflow clearly could use it, I've heard Ask Ubuntu is getting numerous ChatGPT answers, and I've noticed one user going on a crazy joyride on Software Quality Assurance & Testing.

Answer 1

As a moderator, I'd be against having access to this information.

Numerous concerns have been raised in the comments (and other answers) already, but it boils down to two reasons why this is a bad idea:

1. It does not fix the problem at hand.
2. Much of the information in there is none of our concern at all.

This very much reminds me of an XY problem with nasty side-effects.

Answer 2

please give mods the following information for answer posts: [...] All content pasted into the post body input for each revision.

I'm against providing mods access to all content pasted into the post body input, as it sometimes contains private information, e.g. when messing up a copy-paste of a text or an image.

Answer 3

Draft contents: NO

As many others have described, there are multiple problems with providing access to this information, including very significant privacy issues. There are also technical issues with this, as it would end up being quite a bit of additional data to be stored for every post.

Number of draft saves: YES

It would be quite useful to have this information, because Stack Exchange relies completely on the number of draft saves for their "gold standard", which relies on the number of draft saves being "small" identifying nearly every AI-generated content post. Stack Exchange's belief that their "gold standard" accurately identifies nearly every AI-generated post is the entirety of their belief that moderators are vastly over-moderating due to moderators identifying at least 3 to 5 times as many posts as AI generated than the "gold standard" says can possibly exist. That belief, and thus the underlying number of drafts saved metric, is the fundamental basis for their current AI policy. We are where we are now, because SE believes that their "gold standard" gives them a completely accurate view of the volume of AI generated posts on the network.

Stack Exchange's beliefs and actions with the AI-generated content policy are based on them believing it's impossible for there to be any significant number of AI generated posts which are not identified by their "gold standard". Their "gold standard" completely relies on AI generated posts never having a large number of draft saves (for whatever numbers they use for "small" and "large" number of draft saves).

Thus, Stack Exchange's position is that the number of draft saves being small is a near-perfect identifier of all AI-generated content posted on the network. It, of course, also identifies a substantial number of non-AI-generated posts, but their position is that nearly every single AI-generated post is definitely identified by the number of draft saves being "small". That position makes having the number of draft saves for each post critical. The number of draft saves can either:

1. be used to extremely accurately define the smaller set of posts within which all AI-generated content exists (vastly reducing the effort to find and identify AI-generated content, because the ones without a "small" number of draft saves can be ignored as not possibly AI-generated), or
2. definitively show that the "gold standard" does not, at this time, accurately determine the number of AI-generated posts that are posted on the network.

Either way, the number of draft saves per post is critical information for us to have. Currently, the number of draft saves are only available to Stack Exchange, so only they can use this amazing metric, which could both save everyone a large amount of time in identifying AI-generated content and make identifying that AI-generated content vastly more accurate. Alternately, they are the only ones who could specifically check to see if the posts which they agree are AI-generated actually have a low number of draft saves, which would either provide some confirming evidence for the accuracy of their "gold standard" (but not actually validate it) or invalidate the "gold standard" by showing a significant number of posts which Stack Exchange agrees are AI-generated do not have a "small" number of draft saves.

Answer 4

I don't see much need for this feature. Seeing every draft would be worthless. Unless you love to see every spelling error I correct, every phrase I rewrite to feel less blunt, and every word I change to better convey my idea.
Even if we limit this to copy and pasted text, it does not get much better. I expect that on Stack Overflow users will very often copy and paste snippets from their code and then edit out the relevant parts like clients names and such. And under your example those would be saved too.

I suppose the idea came from this paragraph in the linked post.

This metric is based around the number of drafts a user has saved before posting their answer. Stack Exchange systems automatically save a draft copy of a user’s post to a cache location several seconds after they stop typing, with no further user input necessary. In principle, if people are copying and pasting answers out of services like GPT, then they won’t save as many drafts as people who write answers within Stack Exchange. In practice, many users save few drafts routinely (for example, because some users copy and paste the answer in from a separate doc, or because they don’t stop writing until they’re ready to post), so it’s the ratio of large draft saves to small draft saves that actually lets us measure volume in practice.

Since this feature came far before what users call the AI bandwagon, I assume it was implemented to provide the ability to return to your incomplete post if you close the browser or navigate to a different page. I therefore expect that only the last draft is saved at any time, and there is no history of previous draft available in any format.

While the company may have temporarily stored the total number of drafts and some other stats for their study, I sincerely hope they did not try to store the full draft history. Not only would this be super creepy and would probably violate GDPR in a thousand ways (do you really think that you removed the picture with the PII before posting? Tough luck, Stack Exchange saved a copy without telling you!), it would also at least quadruple the database space needed for each single post.

Hopefully this simple example will help. Suppose an user includes by mistake some snippet containing personal/sensitive data in a post. Right now we have a process to ask for removal. But would an user ask to remove something that they don't even know it is stored in the first place? After all they never "posted" it, they just copied some code, pasted in the editor and then edited out the sensitive parts with the expectation that it will not be stored in any format and no one will be able to access the "unsaved" draft. Your proposal is IMHO a violation of trust and probably a violation of local data protection regulations too.

I therefore hope that what you are asking not only will be rejected (volunteers and nonemployee moderators don't have any right to see drafts as far I am concerned), but it is not possible in the first place.

Answer 5

I think this is an interesting idea. However, you are taking a metric that was used to try to measure something in aggregate and trying to use it as a yardstick for individual posts and that's not going to work.

The statistics the company is using to guess at the number AI generated posts could be made available for moderators and 20K users though. I would support a feature request to add more site health monitoring metrics. I think the idea of using drafts was creative although I have some issues with the conclusions that were drawn. I'm certain there are many more indirect ways to measure various aspects of a site's community health.

There was already some discussion of what makes a community healthy. Why not be proactive instead of reactive and put more site health data collection in place before the next disruption happens instead of trying to play catch up?

Answer 6

All content pasted into the post body input for each revision
It doesn't matter if the content has been edited after pasting (that's the whole point of this feature-request), or completely removed (I bet that's hard to implement, and it doesn't really matter anyway)

I think there is a good chance that the ChatGPT spammer won't be editing their response in the SO editor, so that may or may not be useful. There is also a privacy risk regarding confidential information as Laurel mentioned in the comments (I personally can't relate either but I suppose that it is possible if you have a lot of tabs open at once).

I would instead suggest instead showing the amount of time it took for the post to be written, as it would be easy to see whether or not it was copy-pasted. Posts that are posted too quickly could be automatically flagged for moderator attention. I think this would also be significantly easier to implement on Stack Exchange's end. This feature doesn't need to be restricted to mods either; sufficiently high-rep users could also see it as well.

Of course, Stack Exchange has kind of banned moderators from moderating AI content, so we may have to wait a bit until suggesting changes.

I just realized that ChatGPT spammers may try to circumvent by waiting a bit before posting, but this can be amended by showing how much time it took to write each block of text or something similar.

Regarding your edit: I believe that is much safer, but it should only be used as a last resort since most ChatGPT spam is relatively easy to detect.