I’m a copy editor/proofreader and this is a report of my experiences reading papers by researchers using or not using large language models (LLM) to write them and of conversations I have had with their authors. I’m NOT going to report the conversations with AI-using authors, but just describe my somewhat subjective impressions of their products.
Although the LLM may be competing with me, and I’m getting less work as a result, a reason for opposing their use,, I think there are good reasons to use them (and good reasons not to). If you can’t beat them, join them.
TL;DR: Perhaps browse my GoodBadUgly AI link collection, rather than continuing reading?
Keywords:
Features of AI-assisted papers:
AI play with rhetorical conventions
AI-assisted papers observe conservative rhetorical practices common to papers in the area of interest and they avoid developing new conceptual apparatuses for the phenomena under discussion. However, they may surprise by “puncturing the buttressing shield” of the literature with references supporting key observations from scholars about whom readers are unlikely to be knowledgeable.
An example would be the citation of a 100-year-old seminal paper by a scholar whose observations have become foundational in many fields distant from that in which the scholar worked.
These scholars work in disciplines remote from that of the paper being written (e.g., in the arts versus the sciences, and vice versa) and depending on the degree to which they surprise readers, there may be distrust of them and a failure to accept their assumptions among researchers working in the area of the paper.
AI support for the observation is not likely to extend beyond the novel “drive-by” reference itself, and readers are likely to discount it. A similar phenomenon is seen in results from search engines using semantic search that find patterns and connections that no one has seen or made up until that point. Search users may or may not see the search results as relevant.
If authors see great value in the AI’s introduction of ideas from the other discipline, the rhetorical structure of the paper will need to be reworked to reflect the change in focus, which will require successive rewrites with AI assistance. This rewriting will be more difficult for the AI to do than in the case of a paper that follows the style and adopts the conceptual apparatus established in the literature on the topic.
The look and feel of the resulting paper are unlikely to be as acceptable to editors as a “normal” paper, damaging its chances of publication. Of course, it may also be seen as a major breakthrough, establishing a new paradigm.
My job as copy editor is to identify and repair the deficiencies of the text, while handicapped by a shaky understanding of the subject of the paper, made more difficult by the need to integrate two different conceptual paradigms.
In the case of extensive additions/accretions to overworked research questions of more closely related but independent single-topic models, particularly from the research papers of single researchers or research teams without general repute or not widely taken up within the discipline, it is less clear that this is a response to an LLM suggestion and not a result of the researcher’s own initiative.
The LLM may have glommed onto the other papers as a supplementary topic and have failed to develop arguments that adequately support those of the paper. Or the failure to integrate the additional model so that it bears its weight may be a failure of the researchers themselves.
I think this is a valuable use of AI, creating new results and extending knowledge by means of a mechanical compare-and-contrast procedure, and a good reason to use them for brainstorming.
The AI revolution in text production
I don’t know if authors are 1) writing their texts in English and then having LLM correct the grammar, of if 2) they’re writing them in their native language and having the LLM translate them for them.
If they want to improve their own English writing skills, they should be trying alternative 1). Given the time pressure under which they work, conducting research, however, alternative 2) is probably more common, unless students are being given this task.
In any case, the texts produced are “word perfect”, as far as sentence-level grammar is concerned. Within sentences, there are no grammatical mistakes. In terms of paragraph-level rhetorical structure, and vocabulary, they are also good, if not word perfect.
They are indistinguishable from well-written technical texts, and nothing suggests their authors are not native speakers of English.
The problem is the plain, uniform, formal, straight-forward style of the text is a “tell” that it is LLM-generated. As text, it is not interesting to read.
When writing, it is necessary to decide whether to minimize the effort needed to understand the text or to maximize the interest of the way it is expressed. Of course, if the aim is to convey one’s meaning in a way that is as easy to understand as possible, simple uninteresting text may be the best.
I like reading text that is NOT word perfect. As a language teacher and language learner, I think it is important to grapple with the inability to express oneself adequately.
It is not good to cede this struggle to the machine.
As a copy editor, I like trying to express an idea that a writer is having problems with in their second language in a way that a native speaker would use to express it.
I find enjoyment in solving these grammar puzzles.
So I don’t welcome AI use to produce word perfect texts.
I also encourage researchers to practice their writing skills by writing their papers in English themselves, or giving their students this practice.
The native speaker knows how to fix the words the LLM reach for but fail to grasp.
I believe the LLM are able to machine translate CJK texts into English on the fly. They do this by means of a fill-in-the-blank-style process finding words in English to fill slots with the same characterization as the slot.
This works well for words that are used in many academic texts. Words that are less common, or whose proper use depends on particular contexts, may not be characterized well enough in language model data to be found by the model.
For example, in a sentence from a LLM-generated text about city governments’ intention to take action with some goal, I changed “in the near future” to “in short order”. Their intention was not to take action at some point in time, but to take action quickly.
In another, drawing attention to an over-simplified treatment of an issue by previous studies, I changed “fail to capture practical complexities” to “fail to recognize the complexity of actual practices”.
I make many such changes editing LLM-produced text.
Perhaps the LLM choices are “good enough” already. With the continued development of LLM, their vocabulary handicap is also likely to gradually disappear.
For the moment, however, I recommend a read-through of the words by an editor before submission.
Not often matched by human writers
The detail in LLM descriptions of the subject of interest may often be more than researchers themselves are ready to provide. Although perhaps relevant, readers not expecting such explicitness of exposition might find reading the paper heavy going. Determining whether the detail exceeds a level involves deciding whether it is necessary to establish the claims of the paper.
As an example in a cryptology paper,
Encryption/decryption is a two-way process, involving three entities, a plaintext message, cipher text message, and method(s) of encryption/decryption, while engaging two persons, a sender and receiver.Readers of the paper will already know this, and depending on their purpose reading the paper, find its statement useful or better not included.
Asking the LLM to provide a lot of detail makes the paper a resource for further study of its subject by the researcher and others, but lengthens the paper and the ease with which it can be read.
But if this is what is desired, the LLM will do a good job.
I don’t have a recommendation.
I’ve had 2 conversations with non-users of LLM. One researcher said they worried that LLM presentation of their paper’s ideas might not be faithful to the ones they intended. The LLM interpretation of the task set them might not align with that the researcher intended. They also worried that journals may have unspoken, as well as explicit, policies banning AI use.
I had a longer conversation with the other researcher, Dr Fang Ying-Chih (NSYSU). I asked him if he would go on the record with his views.
He said he didn’t mind if I quoted him in detail if his statements were useful. And he was willing to put his name (and that of his school, NSYSU) to it, because they were not doing bad things.
Read what he said at ResearcherCriticalOfAI.
Me atBack to AI