5

Let's say a user has posted a question/answer with the following content. The English is slightly convoluted but it's good content otherwise:

On enormous territory of park there are lakes, rivers, canyons and caves. Lake Yellowstone, one of the greatest alpine lakes in North America, is located in the center of Yellowstone caldera, greatest supervolcano on a continent. Caldera is considered a dozing supervolcano; he erupted with enormous force several times for last two million years. Greater part of territory of park is covered by hardening lava; in a park there is one of five existing in the world of the geyser fields.

We can then use GPT-3 to correct this to:

On the enormous territory of the park, there are lakes, rivers, canyons, and caves. Lake Yellowstone, one of the greatest alpine lakes in North America, is located in the center of the Yellowstone caldera, the greatest supervolcano on the continent. The caldera is considered a dormant supervolcano; it has erupted with enormous force several times over the last two million years. The majority of the park's territory is covered by hardened lava; within the park, there is one of the five existing geyser fields in the world.

Would this mere correction of grammar require an attribution to GPT-3? If yes, would it also be required for Grammarly, which offers similar features without using an LLM (Large Language Model) under the hood?

Note this is not a duplicate of Is attribution required for machine-generated text when posting on Stack Exchange? which asked about novel content rather than mere grammar improvements.

10
  • 10
    Some secondary discussions about attribution on Academia.SE - Is it OK to use GPT-3+ to rewrite your own paper to have better English?. Commented Jan 4, 2023 at 18:27
  • 1
  • 5
    The example used here is of relatively small changes compared to the changes which are, or can be, made by people, including yourself, using GPT or other AI-based technologies. Using such a simple example, when you, yourself, have made much more substantial changes using GPT, seems to be intentionally trying to bias peoples' opinions towards "oh these are just the same as some other known tools, so OK", when the scope really isn't that limited. It feels like you're shopping for an "OK go do this" and then you or others are going to point to it to justify much more substantial edits.
    – Makyen
    Commented Jan 5, 2023 at 0:48
  • 2
    That may be completely different than your current intent, but it's what a substantial number of people do in situations like this. What I've stated is just how this makes me feel about how this question is asked. Overall, where the dividing line is between something that's providing minor assistance and something that's making it such that the text isn't really "your work" anymore is a hard question to answer definitively. It's probably something we're going to be struggling with, both on SE and elsewhere, for quite a while and the answer is likely to change as the technology evolves.
    – Makyen
    Commented Jan 5, 2023 at 0:57
  • 3
    If the "mere grammar improvements" are not substantial enough to meet a reasonable person's threshold for citing, then they're not substantial enough to bother making. We only want users to make substantive edits, not trivial ones. Thus, the same answers apply as the existing question, whose answer concludes, simply: "If it wasn't created by you, attribution is always required here." Commented Jan 5, 2023 at 6:49
  • 1
    @CodyGray is the example in my question not substantive? Commented Jan 5, 2023 at 13:27
  • I would have said that you don't need to say which spellchecker you use but staff says you must identify content that you didn't create; so you'd need to mention "Spellchecked by X" in your post. - So same for speech to text. --- This comment is my own work, I received no assistance in creating or typing it.
    – Rob
    Commented Jan 5, 2023 at 14:42
  • 2
    @CodyGray, I don't think this is a duplicate. The other question asks about posts generated entirely by a machine. This asks about using machines as a tool to help editing. The answer there doesn't clearly answer this question. I object to closure as a duplicate. I think the appropriate response is to write an answer here describing clearly the answer to this question. I understand some may believe that the answer to this is implied by the answer there, but I think that is debatable, and I think the appropriate response is to write the justification for that conclusion in an answer.
    – D.W.
    Commented Jan 9, 2023 at 22:29
  • 2
    @Rob I disagree, regarding the need to cite a spellchecker or Grammarly or whatever. I think that it comes down to the fundamental goal of these services: spell checkers, grammar checkers, copy editors, etc are meant to help someone fix technical problems without changing the tone or meaning of a passage of text. By contrast, GPT is specifically designed to create novel (or at least, novel-looking) text. The "off-label" user of GPT to "fix errors" could easily create new content, which needs to be checked for. Disclosure seems ethically appropriate in this case. Commented Jan 9, 2023 at 23:01
  • @XanderHenderson, it is Johnathan who brought up that point, to which staff replied that is incorrect. I'm not the one to disagree with. --- In addition, a grammar checker would probably change the meaning, a spellchecker rarely so (but more often than never).
    – Rob
    Commented Jan 9, 2023 at 23:54

3 Answers 3

10

I can't set rules, because I'm not an employee. So this answer is just my thoughts on whether attribution should be given, not on whether it is required.

My theoretical side says it probably won't hurt to disclose, it sets a good example and won't hurt until all the dust on what place (if any) machine-generated content can have on the network is settled. What if the LLM makes a mistake or a particular LLM consistently makes many? In such cases, it would probably be nice to know which LLM did it, and it might help pick out problematic LLMs if more of them arise. It never hurts to do the best you can, to give as much information as possible, so add a disclaimer/attribution, for sure.

My more practical side says I've been using Grammarly a lot to help me with my English, and when I start editing a post, sometimes I edit a few things more that I hadn't noticed, but Grammarly did. Yet I don't add a disclaimer that Grammarly was responsible for those particular parts of my edits. For a bit of fun, let's see what Grammarly makes of that first quote in your question:

On the enormous territory of the park there are lakes, rivers, canyons, and caves. Lake Yellowstone, one of the greatest alpine lakes in North America, is located in the center of Yellowstone caldera, the greatest supervolcano on the continent. Caldera is considered a dozing supervolcano; he erupted with enormous force several times over the last two million years. The greater part of the territory of the park is covered by hardening lava; in a park, there is one of five existing in the world of geyser fields.

It's actually worse than what GPT made of it. Grammarly isn't as flawless, so I would still be forced into doing manual checks, and manually doing a few more fixes. Copy-pasting a result from an LLM doesn't really force you into double-checking as much as using Grammarly does. And, another edit after I made one with the help of Grammarly could make sense, because we both may have missed a spot.

Also, Grammarly only works because I already started editing, after recognizing the post needed work. And if the post is already as correct as Grammarly can recognize, it just doesn't suggest further changes when I mash the edit button: I get no 'easy' or 'free' edit from it if the post is already 'good enough'.

BUT! Just for fun, I've run your second quote (the one GPT already corrected, and for which Grammarly gives me no suggestions) through GPT again, with the prompt "Rewrite this to standard English". This is what you get:

The park covers a large area and has lakes, rivers, canyons, and caves. Lake Yellowstone, which is one of the largest alpine lakes in North America, is located in the center of the Yellowstone caldera, which is the largest supervolcano on the continent. The caldera is considered a dormant supervolcano and has erupted with great force several times in the last two million years. Much of the park's land is covered in hardened lava, and the park is home to one of the five geyser fields in the world.

So, basically, even though the quote didn't need changing, the LLM followed my command nicely, and gave me an edit that's over 6 characters long and can be submitted to SE with no problem. It is still flawless according to Grammarly. Which means it's time for a party! 'Free' edit credit without me having to do any thinking/checking! 'Free' post bump and visibility! 'Free' visibility for me through my user card on that post as well! They all may seem relatively harmless, but they could all be abused: edits give badges and count towards the candidate score in moderator elections, meaningless edits aren't allowed so using them to bump a post for more attention certainly isn't, and the user card visibility could be used to push a certain username (think spam, or political opinion) in front of everyone. So there are certainly downsides!

I wouldn't outright ban using GPT to correct quotes like your first one, which, admittingly, does need a lot of work. And GPT did a nice enough job of making it a nice quote. But based on all of the above, I think the following should be the absolute minimum requirements for when you do use GPT or any LLM to edit posts:

  • Don't just copy-paste, give a command, and copy-paste the outcome back. Manually check it. When I use Grammarly to edit posts, I do the same.
  • Abide by the SE guidelines for edits. In particular, make sure your LLM doesn't do anything that's not described under the 'When should I edit posts?' section.
  • Make sure the post needs editing before firing up the LLM. As I showed above, it will follow the command to rewrite, even if nothing needs to be rewritten.
  • Attribute the LLM. It makes it easier to spot patterns of any problematic LLMs. It also helps moderation: If you state you've already used an LLM, any user that shows a pattern of making more meaningless edits using an LLM (they will probably not disclose but be able to be spotted) after yours for whatever reason doesn't have the excuse of 'I couldn't know!'.
2
  • Re "Manually check it.": Yes, that goes for all tools, incl. spell checkers. I have never used Grammarly (only seen the ads), but I am surprised it couldn't fix the Germanism "he erupted with enormous force " (I had one German colleague who very consistently did this (and "infected" a Belgian), not all the others). Commented Jan 5, 2023 at 1:41
  • The real test for GPT-3 is on 'tech' text where only context can decide if it is expect or Expect, meteor or Meteor, google or Google, puppet or Puppet, blade or Blade, selenium or Selenium, etc. Commented Jan 5, 2023 at 2:01
7

I'd say it's a good idea to clarify that a post has been edited (not created!) with ChatGPT at this point of time. The main issue with LLM is that it’s a tool that's been gratuitously misused — and frankly the 'way' it writes has a voice of its own, much like many users have a distinctive writing style.

In this case attribution isn't required — since the substantial content is generated by the user but is encouraged since there's a risk of being mistaken for a bot.

It's worth remembering the 'issue' with ChatGPT is users cynically using questions as prompts, and posting them to see 'what happens' - and its often plausible garbage. If it's used to check a correct answer, well, it's a correct answer that might be mistaken for plausible garbage. This is bad.

If I wanted to play it safe, I'd post the original 'flawed' post, in my own voice, then use the LLM fixed version as an edit. It establishes that the original post as it was meant was done by the user, shows their intent and shows what edits were made. Should the 'automated' correction system be flawed, it gives folks a way to confirm/try to guess at the user's own intentions as well. It's a 'little' troublesome, and as such isn't really something I'd want as policy, but it feels like the best practice in such cases.

At the end of the day, much like a hammer — it's a tool to be used (for good or evil) or misused. Problem is really folks have been throwing hammers at nails, with little heed for what they hit, rather than figuring appropriate uses.

1
  • The way that you suggest users post GPT "edited" content is similar to the way that I would generally require users to post, for example, content translated by Google. Leave the original, and paste in the translated material, with attribution to Google. Commented Jan 10, 2023 at 12:43
3

This is a tough issue, with reasonable arguments on both sides, and I'm not sure that there are any existing rules that provide a clear answer.

On the one hand, we don't require attribution for people who use tools like a spellchecker or Grammarly to improve the English of a post. Also, requiring people to disclose use of a LLM takes a little bit of extra time, which puts a little more work on people who are making edits in this way, which might potentially discourage valuable edits, thereby disincentivizing changes that would improve the site and serve our mission. So we should only impose such a requirement if the benefits outweigh the disadvantages.

On the other hand, GPT makes more substantial changes, which carries the risk that GPT has changed some meaning without you realizing it. Grammarly is unlikely to change the meaning of the question -- but there is a more significant risk that GPT changes the meaning in some small but important way. This is particularly risky if you are submitting the original text to the LLM, using the LLM to rewrite the entire thing, and then copy-pasting the result over without careful word-by-word review. So I can understand the argument that it's important to disclose the use of LLMs, to address this risk.

@animuson, a Stack Exchange employee, claims that "If it wasn't created by you, attribution is always required here." For the reasons articulated above, I am unsure if such a broad policy is indeed a good idea, or if animuson even meant to cover this case, but this is arguably a precedent that attribution is required. I think it is helpful to look at the reasons animuson provides for this policy. animuson states that use of machine-generated text "can be important information in determining the validity of an answer", and that seems like a valid concern, as use of machine-generated text may be relevant for evaluating the credibility of the post.

I find the recommendations on Ask Ubuntu to be helpful to consider if you're going to use GPT or a LLM to generate edits. In particular, I'll copy a few items from that post:

  • Confirm that the post is correct
  • Fully understand the material that you are posting
  • Correct any issues
  • Communicate in your own words
  • Warn of corner cases or potential issues if needed
  • Cite any reliance you made on ChatGPT/GPT/LLMs for the post

Finally, if you are editing posts, make sure you haven't removed links or destroyed useful Markdown formatting. Submitting a post to a LLM to rewrite it is likely to remove all formatting and links. Losing that would likely be a loss, in most cases, so I would be very wary of any LLM-assisted edit that lost links or Markdown.

6
  • 1
    "Also, requiring people to disclose use of a LLM takes a little bit of extra time, which puts a little more work [...]" If re-using a common edit summary message, on my Chrome browser, it takes me 1-2 keyboard presses and a mouse-click to select an edit summary that I've used enough times before for re-use via autocomplete, which is either the same or less effort than hot much it takes me to write any other edit summary.
    – starball
    Commented Jan 4, 2023 at 23:55
  • 1
    @starball unclear if edit summary alone is sufficient, please add an answer on What is the proper way to attribute Large Language Models (such as GPT-3) when posting on Stack Exchange? Commented Jan 4, 2023 at 23:57
  • My comment is not an answer there. That post is about edits and new posts. My comment is only about edits.
    – starball
    Commented Jan 4, 2023 at 23:58
  • @starball, I understand - we could debate how much time it takes, and how much it will discourage edits. But I think the concern has some validity: there there will be at least some impact on edits, and at least the potential for some reduction in useful edits. We could debate whether it is a significant effect or "de minimus" impact. And then we have to weigh the costs of that impact on edits, to the benefits of attribution. I think there are reasonable arguments on both sides.
    – D.W.
    Commented Jan 5, 2023 at 0:17
  • As animuson says in a comment on that answer, "Our policy is clear. You didn't write it. It requires attribution. Period". If you don't give attribution on content you didn't create then anyone copying that material will not be able to attribute it correctly.
    – PM 2Ring
    Commented Jan 18, 2023 at 12:24
  • Also, mentioning that content came from GPT is actually helpful to the next cycle of GPT. The Web is becoming flooded with GPT-3 content, and I expect that OpenAI would like to give a low weighting to such content in the next Common Crawl dataset, since feeding GPT output back into GPT isn't going to give much improvement to the language model.
    – PM 2Ring
    Commented Jan 18, 2023 at 12:30

You must log in to answer this question.

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