Is it acceptable to use LLMs for improving the English of answers and questions?

Question

When writing an answer or a question on this stack, I try to be the clearest possible for exposing my problem or for proposing a solution, but I recognize two main issues¹:

• I use the same sentence structures again and again
• My tendency is to employ the same words, again and again.

I believe that this comes from two aspects: I am not a native English speaker and my styles comes mainly from writing Math scientific paper, where the language is not very rich.

I thought then to ask for the help of Large Language Models (LMMs) such as CHATgpt, by asking something as " Improve the English of < part of question/answer >. Of course, one should check if the LLM produced a coherent writing and did not make up anything. For example, the part of one answer of mine reads;

The options presented in Table "Uthgardt Totems" are available at 3rd level, and such table shows the equivalent of these options to the animals of the original PHB's subclass and to the new option Tiger.

The new options add new totem spirits: you have to follow the same subclass rules depicted in the PHB for the choices you are taking during leveling up. This means that you can choose a Tree Ghost totem at 3rd level, and subsequently the Eagle spirit at 6th level, since the PHB rules states, for the Aspect of the Beast feature, that:

you can choose the same animal you selected at 3rd level or a different one.

and a possible outcome of chatgpt is

The choices outlined in the "Uthgardt Totems" table become accessible at the 3rd level. This table not only illustrates their counterparts among the animals featured in the original Player's Handbook (PHB) subclass but also introduces the new option, the Tiger.

These new options introduce additional totem spirits, and it's essential to adhere to the same subclass rules detailed in the PHB when making your selections during character advancement. This means that you can initially opt for the Tree Ghost totem at 3rd level and subsequently choose the Eagle spirit at 6th level, in accordance with the PHB guidelines for the Aspect of the Beast feature, which says that

you can choose the same animal you selected at 3rd level or a different one.

The meaning is the same, but the English is much better.

Can we use such LLMs for improving the English of our posts? Or should we ban it for this purpose too?

Note. I am not asking if we should allow the employment of LLMs for generating questions and answers (we already have a meta for that), but just for improving the language of a human written answer.

---

¹ Ça va sans dire, the third issue regards typos and grammar errors...

Answer 1

To sidestep your question for a moment, I have long believed, longer than the existence of LLMs, that you have provided some of the most consistently high quality answers on this site. I have never read one of your answers and thought “this needs some composition help”. Further, I’m pretty sure you are the top recipient of my bounties, both in reputation and number of bounties. I have an abundance of confidence that you don’t need LLM help, to the point that I’d be worried they might actually make your answers worse, at least in my opinion. Now, on to the actual question.

LLMs present a plagiarism problem for us.

I won’t reinvent the wheel here. Xirema summarizes perfectly my trouble with LLMs, reproduced below in its entirety.

---

I'm going to reiterate the points I made when this question was proposed at the network-wide level in a more succinct form; you can read my answer in the link if you want the long form.

Yes, ChatGPT (and other AI-generative tools) need to be banned on RPG.SE

And it's for the simple reason that these tools, at least as they exist today, are incapable of producing output that does not commit plagiarism.

There's a few criteria I would line out that I think are necessary prerequisites before they should be approved for use on this site (or any site in the Stack Exchange network), and they are

• That the tool clearly cites every single work that was used as part of its training and ingested into its dataset
• That every single work that was used provided explicit, specific, opt-in permission by its original author to be used as part of that specific tool and its training process
• That the evidence of permission for each work that was used is easily accessible
• That the tool is able to cite, in its resulting output, the works sourced that were used to generate its output.

To my knowledge, none of ChatGPT nor any other publicly available AI-generation tool is capable today of satisfying all these criterion (and it's my understanding that most-if-not-all of these tools are technologically incapable of doing so), and they should not be permitted until/unless they are developed to be able to satisfy these criterion.

Answer 2

No, you should avoid using LLMs and ChatGPT for proofreading and for attempting to improve posts and their English. This is because it will make your posts worse, not better. Proofreading is not the function of ChatGPT and it is not something it will do for you.

The example post you gave us was already 99% fine. The GPT post doesn't make it better. It introduces weird phrasing like “at the 3rd level” which we don't use because it doesn't understand how we write in our domain. It turns a sentence into marketing talk (“not only ... but also”) which has no reason to be that way. ChatGPT also makes several things wordier for no reason and rephrases several things for no actual benefit—these are not improvements! If this were submitted as an edit, I'd roll it back and ping the editor in comments to have words with them about editing guidelines.

The only things I'd have changed in the original were the phrasing of “and such table” and probably dropping the quotes around the name, and I don't need GPT to detect that or fix it for me.

ChatGPT was made for chat conversations. Like, you visit a company's website and want to speak to support and it offers an instant chat with a bot to route your query — ChatGPT was made for doing that! That's why it has “chat” in its name. It is not a proofreading AI, it's a chat AI. If you give it input, it will generate chat, not proofreading. It will not make your post better.

You also don't want to pass your otherwise completely legitimate post through a filter that will turn it into ChatGPT output. Not only has that made your post worse, it's going to draw a target on it: you submitted a GPT output for a post and we ban those, and people are going to be looking out for GPT-ish posts to remove them. Passing a perfectly good post through ChatGPT is the best way to sabotage that post. (Or, worse yet, you'll create traps and problems for people trying to hunt down actual GPT-based misuse of the site.)

You already have perfectly adequate proofreading tools available: your own judgment and a spellcheck. You're doing well with them — keep using them and keep improving.

Answer 3

Yes, that’s fine. All critiques of plagiarism are irrelevant if you are providing an answer and having it punch up your grammar. Make sure it doesn’t change your meaning though before posting

Answer 4

I Think It's Fine For That Specific Purpose

There are two broad classes of objections to clean-up AI, that I've seen (not limited to this forum): The argument from copyright, and the argument from the virtue of labor.

To the former (copyright), I say this: What matters more is the output of the tool, than the input. No content producer has ever had the right to know whether another content producer has read their work first, or to know if they read a borrowed copy, or to demand attribution if that's the only link. That's simply not how this works, how this has ever worked, or could ever work.

What copyright law protects is concrete expressions. And while I acknowledge that the academic literature is still being written on the copyright (and, interestingly, security) implications of how LLM models work, and the possibility of direct copying from the training corpora, the risk of direct copying seems very low if you're giving it your own primary text and just asking for grammatical and stylistic clean-ups.

To the latter (virtue of labor), I have heard that argument so many times in so many ways over the course of my life that they all just blend together into an undifferentiated muddle. I collaborate on radar and satellite designs-- my work is improved, not degraded, by using calculators, spreadsheets, and numerical modeling packages. I collaborate on writing their specifications-- my work is improved, not degraded, by spellcheckers and grammar checkers.
(Ironically, but truthfully, I misspelled 'specifications' and 'misspelled' just now. I've never been able to spell 'misspelled' right on the first attempt.)

These tools, for this specific purpose are no different. Like any other tool, they need to be used wisely, for instance a minimal check to make sure the meaning hasn't been changed. But they are just tools designed to reduce the burden of certain tasks.

But Note Well: I am not in any way changing my position on the topic of using these tools to generate questions or answers wholesale. They are nowhere near reliable enough to consider that application. And when they are, people will bypass Stack Overflow entirely.

Answer 5

What's commonly called artificial intelligence on the Internet nowadays is a method that, as with the machines it's run on, lacks will and intent. It is entirely responsive, entirely a reaction to inputs based on other inputs. It does not understand how people communicate, only the arrangements of words, because it does not have sentience.

When a real person tells you what they think somebody else is trying to say, there is interpretation founded on reason and expertise, from a person who understands that other people exist and tries to understand them. Someone who is capable of empathy because they're bought into the system of empathy as an actor in the world.

When you feed text that's lacking information into a machine, that machine can't uncover the information that's missing, it can only guess (the same way that when you feed a low-resolution photograph to a machine and ask it to upscale the photo to a larger size, it will make up elements of the image to fulfill your request, without truly having the information that would give it a better picture).

Most importantly, the machine can't be held accountable when it guesses wrong, because it isn't a being involved in the community, merely an automated service. The only person that can be held accountable is the person who decides to involve the machine in the first place.

For these reasons and more, it is less than reasonable to allow ML-parsed text to replace genuinely authored text, even in cases where a person is having difficulty expressing themself. The machine will not reduce difficulty except insofar as it suggests what's common. Sometimes that's what you need, but common text deserves no special priority on a question-and-answer site.

People will want to help you even if your command of the language you're using isn't the best.