Stream: Miscellaneous

Topic: How do you precieve AI usage for writing papers


view this post on Zulip walker (Aug 14 2023 at 22:12):

So, I imagine people here are involved a lot in academic work, publications and paper reviews, and I wanted to know how would you perceive a paper that is written with assistance of AI. Now I am just asking about the writing part, and not making up results, in particular for improving paper quality.

for instance I just opened some AI prompt that improves writing and I typed random stuff like this:

In this paper we present a full recipe for creating a FOO. FOO has been known in literature
for long time, it  has been described in [1], [2]. [3] claimed that creating a FOO has lots of
details that one must get all at once otherwise the procedure fails. That is the key
contribution of this paper, a procedure for making a FOO that is reproducible, and precise
and produces FOO in high quality with large percentage yield.

And the AI turned it to this:

In this research paper, we are pleased to provide a comprehensive recipe for creating a FOO. The
concept of FOO has long been recognized in academic literature and extensively discussed in
previous works, such as [1] and [2]. Moreover, it has been acknowledged by [3] that the creation
of a FOO requires careful attention to numerous intricate details, as missing out on any one of these
aspects would likely result in a failed procedure. The primary contribution of this paper lies in the
development of a reproducible and precise procedure for manufacturing a FOO, which guarantees
a high-quality end product with a significant yield percentage.

Now it is clear that the second paragraph is much much better then the first, but then I didn't exactly write the second one myself.

I wonder what people in formal methods/ theoretical CS think about that?

P.S. personal opinions are welcomed!

view this post on Zulip Paolo Giarrusso (Aug 14 2023 at 22:40):

Crediting the AI should absolve you of (some) concerns about plagiarism. But are you confident that AI will not mangle key parts of your claims? Even if you are, can a reviewer share that confidence?

view this post on Zulip Paolo Giarrusso (Aug 14 2023 at 22:42):

I expect some reviewers will have stronger feelings against this; I don't know if they'd be wrong.

view this post on Zulip Patrick Nicodemus (Aug 15 2023 at 00:27):

I would be on the lookout for subtle changes in the meaning of what you are saying, which come as a result of the LLM swapping out boring phrases for more flowery ones.

Example: someone asks: "Can you give an example of X phenomenon in mathematics?"
You first write: "An example of X phenomenon in mathematics is (...)."
The LLM spruces it up as follows. "Of course! Examples of X abound in mathematics, it is a widespread and pervasive pattern. We will concentrate on the following example."
Now, do examples of X really abound in mathematics? Or is that just a common trope you observe in people talking about mathematical patterns. Maybe X is frankly a rather rare phenomenon without too many examples, although important examples, but the story sounds better if we say "Examples are everywhere!"

As the writer you are responsible for the meaning you want to communicate. I think it would be dangerous to think of the changes made by an LLM as purely stylistic, because meaning and style are bound together. I really think that it is difficult to co-write a paper with an LLM without it supplying a nonzero amount of the meaning, the semantic content of the paper that you really are supposed to be responsible for as a scientist. I don't see another way to communicate precisely than to choose your own words carefully.

view this post on Zulip Patrick Nicodemus (Aug 15 2023 at 00:30):

I also think that a lot of what is being supplied here by GPT is cruft/verbiage. The scientific papers you really remember actually say something with every word. The words added by GPT often don't really communicate anything at all, they're just boilerplate, a little bit like a teenager writing an essay and trying to pad it out. "Since the dawn of time man has asked..." "Webster's dictionary defines "irony" as follows...". Like, good writing doesn't need boilerplate. You just say what you want to say.

view this post on Zulip Théo Zimmermann (Aug 15 2023 at 15:54):

FWIW, the recently published FSE call for paper contains a note on when to acknowledge or not the use of generative AI when writing a paper: https://conf.researchr.org/track/fse-2024/fse-2024-research-papers#submission-policies

view this post on Zulip k32 (Aug 16 2023 at 21:09):

I used it as an advanced spell checker and thesaurus. "Please rewrite the following paragraph in a style of scientific paper" prompt followed by informal description with engineering jargon.

view this post on Zulip James Wood (Aug 16 2023 at 22:28):

With page limits being what they are, the second paragraph is 2 lines worse than the first.

view this post on Zulip James Wood (Aug 16 2023 at 22:46):

In terms of content, the second paragraph is more grammatical, but the style sucks. I think the first would be much easier to redraft into something acceptable, given that the required grammatical changes are routine (like, “known in literature” ↦ “known in the literature”) and the required rephrasings largely preserve the content words and intended semantics (e.g, “That is the key contribution of this paper, [...]” ↦ “The key contribution of this paper is [...]”).

view this post on Zulip Paolo Giarrusso (Aug 16 2023 at 22:49):

Honestly http://www.denisdutton.com/bad_writing.htm came to mind — while the AI output goes nowhere as far, both go in the same direction, in part because such writing often passes for smart.

view this post on Zulip James Wood (Aug 17 2023 at 13:42):

Paolo Giarrusso said:

Honestly http://www.denisdutton.com/bad_writing.htm came to mind — while the AI output goes nowhere as far, both go in the same direction, in part because such writing often passes for smart.

That site seems to be down, so here's an archive link: https://web.archive.org/web/20220425131727/http://www.denisdutton.com/bad_writing.htm


Last updated: Jul 24 2024 at 11:01 UTC