On writing and computers (perhaps also ChatGPT and large language models) (#1254)

Topics/tags: On writing, Generative AI

I drafted much of this musing on June 20 and then put it aside for nearly two months. Some time-based comments are from that date, particularly the first line.

Last night, I dreamed that I gave a talk about my Mathematical Images Synthesis Toolkit project that began with a reflection on the ways in which computers change processes. I spoke a bit about writing, in particular, and how I’ve seen writing processes change over my lifetime. As I noted in the dream, I’m one of the last generations not to have access to computers as writing tools in high school [1]. It was one of those nights in which I had a lot of different dreams; however, each came back to thinking about how I might turn that earlier reflection into a musing. I’m not sure how much I recall or even whether the original dreamt [2] reflection was all that good. But my muse calls upon me to write. Here goes.

Computers have transformed the ways we write; at least, they’ve transformed the ways I write. In high school, I wrote—or at least wrote essays—in much the same way as high schoolers had written for the prior half century or so. That is, I gathered information, made an outline, wrote the paper, read through the paper and made notes on changes, rearrangements, and new sections, rewrote the paper, repeated the process as many times as appropriate, and then typed up the result. That’s not quite accurate. When I say wrote, you probably imagine me putting pen to paper. But I’ve typed faster than I could write since I was in seventh grade or so, so most of my writing happened on the typewriter. I suppose I could have written typed; however, retyped doesn’t have the same meaning as rewrote.

Note that this writing model leads to some practices that are still common. Why do we double-space text? Because it leaves more room to insert comments and make changes. Why do we call the operations cut and paste? Because that’s what some people did when rearranging their essays [3].

When I started college, I still typed. However, I quickly learned how to use the text formatting system on our mainframe. At least I think I did; it’s been a long time. I even got to use the legendary TECO-based Emacs, although I didn’t learn TECO. At first, using a text editor meant that I had a more convenient way to follow the processes I’d been following. In particular, when rewriting, I could use commands to move text around, rather than literally rewriting (or retyping) text. As most modern writers know, it’s much easier to delete and insert—or copy and paste—on the computer than it is on paper. But the process was still much the same: Write, review, make notes on changes, make the changes, repeat as necessary.

Some obvious things changed. I could now play with fonts and faces. Where I had previously been limited to underlining for emphasis, I could match the boldface and italics I saw in printed books. I suppose I could have retyped text for bold, but italics required a fancy typewriter, such as an IBM Selectric, and a willingness to switch the type balls for a word or two. Underlines don’t feel quite the same. I use bold and italic much more than when I was a high-school or college student; I don’t know when I made the full switch.

Midway through college, the Macintosh and the IBM PC appeared. Since I worked at our Central Users’ Site [4], I had ready access to both computers and printers; first dot-matrix, then LaserWriters. As with Emacs and a text formatter, MacWrite and the early versions of Microsoft Word made the traditional processes much more efficient. They also made fonts and faces even more accessible, leading too many people to horrendous design choices.

At some point, my processes started to change. The reading, commenting, restructuring, and more that used to happen at the end of a writing cycle started to become part of the writing process. I’d reach the sixth paragraph of an essay and realize that I wanted ot make changes to the second. I’d get to the twelfth paragraph and find that I wanted to move the third paragraph there. I think others experienced similar transitions.

Then the software evolved. Now, it helped with the less fun and less creative aspects of editing, particularly finding misspellings and awkward phrasings. At first, I used these new tools much like I had used my brain (or my friends’ brains) before: At the end of a writing cycle (a.k.a. a draft), I’d check spelling and style. Most frequently, I’d mark up a printed copy. I didn’t physically cut and paste, but I did draw arrows to indicate how I should rearrange a paper.

At some point, the software evolved enough that I’d fix those live, while writing, rather than waiting until the end of a writing session. I still printed out drafts to make comments, but I did so much less frequently.

These days, I’ve moved backwards. I edit in vi [5] or, more precisely, vim [6,7]. I don’t get live comments on my writing from my editing software. I prefer to separate writing and that kind of editing. Grammar and spelling issues seem to occupy different areas of my brain. Rearranging and some kinds of rephrasing, on the other hand, seem to occupy the same regions of the brain as creating and fit well into the live process.

I know that others’ processes changed even more significantly. In particular, free writing became a much more common practice once word processors became more commonplace. If you can grab useful ideas and phrases from your writing and easily restructure them, it appears more reasonable to write freely. At least, I think it does.

During the past few months, the writing process has potentially changed much more significantly. Or at least the options available for the writing process have changed. Now, we can ask chat-based large language models for advice on how to begin a paper, or how to rewrite our paper in the voice of another author, or even to provide us with the first draft of a paper. What will those opportunities do to our writing? [8] I have no idea. I haven’t explored these tools sufficiently much to understand.

I may also lack the creativity to think about how to use the tools [9]. What do I mean by that? My first inklings of how I might use ChatGPT were generally of the form I could ask it do draft an essay, or at least a paragraph of an essay, and then I could rewrite it in my own voice. Then I learned that folks ask ChatGPT to rewrite their own work in the voice of another author. I’m not sure I’d use that option much, but I can see that as a fun way to explore other ways to phrase things [10].

My colleagues in English have suggested ideas like, Advise me on how to remove forms of to be from my essay. That seems like a good idea. There are likely many other similarly valuable questions for ChatGPT, questions that won’t flatten our text.

I also read an essay by a Columbia student [11]. Rather than asking it to write a whole essay, the student asked smaller questions: What’s a good thesis? What kinds of evidence might I present? How should I organize my essay? Those are good questions to ask. And it struck me that these systems could help equalize the writing process. Some high schools help students learn how to organize a longer essay and give them practice doing so. Others do not. Presumably, chat-based LLMs could help address such differences.

What will I do with ChatGPT and other LLMs as a writer? I’m not sure. I suppose I will see.

I know that I’ll do my best to turn off the use of my inputs as training data [12]. Unfortunately, although that is possible in ChatGPT, it does not seem possible in Bing Chat. Given that ChatGPT is not on Grinnell’s approved software list, and the College has not licensed Bing Chat Enterprise, it may be that I should forget about it all for the time being.

What will I do with ChatGPT and other LLMs as a teacher? I’m even less sure about that. Fortunately, I’m not teaching writing this semester (or at least not the writing of English). For writing code, I’m asking my students to avoid LLM-based tools. Why? Because I still believe that you learn from writing the code yourself, and you’re unlikely to be able to write more complex algorithms and programs until you’ve mastered the basics.

In any case, it will be an adventure.

[1] I graduated high school in 1982. Damn! That’s more than forty years ago. I’m old.

[2] dreamed?

[3] Rather than physically rearranging my text, I tended to rely on circled text and notes about rearrangement.

[4] Usite. Unfortunately, too often rephrased as Lucite or Lusite. You can guess what the L before user implied.

[5] vi is short for visual.

[6] I think vim stands for vi, improved. I tend to think of it as vi, mangled.

[7] R.I.P. Bram Moolenaar.

[8] Perhaps I should be asking what these tools will do for our writing.

[9] I am amused to learn that Grammarly thinks I also lack the creativity to think about how to use the tools is more confident than I may also lack the creativity to think about how to use the tools.

[10] Stay tuned for a musing in which I put a paragraph through the ChatGPT wringer.

[11] Terry, Owen Kichizo (12 May 2023). I’m a Student. You Have No Idea How Much We’re Using ChatGPT. The Chronicle of Higher Education. Available online at https://www.chronicle.com/article/im-a-student-you-have-no-idea-how-much-were-using-chatgpt.

[12] I realize that that’s a strange attitude, given that I release my text under a Creative Commons license. I’m looking forward to a new CC license that says, This text may be used for all purposes except the training of large language models.

Version 1.0 released 2023-08-16.

Version 1.0.1 of 2024-06-03.