Generative AI and Writing--Why Not?
AI's been in the discourse a lot recently. I thought I'd summarize my take on both its general nature and its "use" (or rather, preferably non-use) in creative writing.Â
This ended up being a lot longer than I'd planned. If anyone has any additional insight, please feel free to comment.Â
Disclaimer: I do have a background in computer science, but am a far cry from an expert in artificial intelligence and don't currently work professionally in the industry.Â
What is Generative AI?
Artificial Intelligence (AI) is a broad term--it can refer to any computer system that can do tasks that we might typically associate with human intelligence, like processing data, recognizing patterns, or making predictions. Generative AI is also broad, and simply refers to a type of AI that can create "novel" content based on existing data. The models that we tend to be talking about today when we refer to AI, such as ChatGPT, Claude, Gemini, and such, are specifically something called "Large Language Models" (LLMs)--models trained to do something called "Natural Language Processing" (NLP) by utilizing a specific type of "neural network" (a computational model based on a series of interconnected nodes, loosely modeled off the way we think the human brain works) to "learn" from vast amounts of pre-existing data. To put it simply, these models use what they've learned from their training to "predict" what the next pieces of text (represented as tokens) will be. They are not thinking or being truly creative in the way we think of human creativity--simply using the patterns from the vast amount of data they were trained on to make these predictions.Â
The Art of Writing
I probably don't have to preach this to this audience, but creative writing is more than just slapping words onto a page. There's an art to it, and creativity and learned technique in equal measure. Moreover, the overall structure of a work is important, not just what word comes next after another word, what sentence next after a given sentence. Language transmits meaning, and that meaning can take the form of stories--themes--feelings--metaphors--consistent characters that leap off the page. Doing this well takes careful iteration and practice to master, and is rewarding for writer and readers alike--establishing in the process a creative bond between one who writes and one who reads. There's meaning being transmitted from human mind to human mind.Â
'But wait,' you might be saying. 'If AI is advanced enough to generate work that feels just as meaningful, what's the harm in using it?'Â
SoâŚWhy is AI an Issue When Used to Write "Creatively"?Â
These models can only generate their predictive text because they're pulling from a vast array of training data--data which, in the vast majority of cases, was written by real human authors and was used without the permission of those authors. Not only that, but these models even sometimes reproduce almost verbatim copies of copyrighted material when prompted in a certain way. This isn't the same as humans learning from vast arrays of material and then using their own brains and human experiences to create something new. This is a machine essentially remixing other people's work. I think there's a strong argument to be made that this is plagiarism, or if you want to quibble over the word, at least dubious from an ethical standpoint. We should probably err on the side of not stealing people's art. ButâŚwhat about the utility of it? Can it actually write well? Is using it even a good idea from a purely practical standpoint, setting ethics and artistic integrity aside?
I tested models like ChatGPT and Claude on various writing samples to investigate this for myself.
The Major Use Cases:
Prompting the AI to write a story for you.
This is the use case that many people instinctively bristle at. So, is AI even any good at this?
It can certainly generate things that look okay on the surface. The grammar tends to be largely correct. It can generate sophisticated-looking story arcs, though on inspection they're often clichĂŠd.Â
However, given that these models really only use previous tokens to predict the next tokens, they aren't thinking about overall story structure, or coherence over a long-form narrative, or character consistency over time. You'll find that these models make "mistakes" or cause certain inconsistencies to surface that are unlike those a human would make--a sort of uncanny valley effect, where everything's just a bit "off". The character says something that seems reasonable--quite like what characters in fiction would often say in such a situation--but doesn't quite fit their prior characterization. Or the story subtly contradicts itself in a way that you wouldn't expect from a human author.
Can you try to catch all these subtle mistakes and fix them? Sure. But at that point, I would argue that (a) you're unlikely to catch them all, especially when the soul of the work is missing--it's not really about each individual error as it is about the work as a whole, (b) you're robbing yourself of the ability to come up with the story yourself, while also still having to rewrite it because the AI can't do it properly.Â
Doesn't seem like the best use of your time, or a reader's.Â
But what aboutâŚ
Having AI read your work and suggest high-level edits, like rewording things or adjustments to the plot, or inconsistencies?
I've noted that its "edits" tend to be hit-or-miss at best. It often clocks plot holes that aren't real plot holes, suggests changes in the plot that subtly break other parts, or endorses rewording sentences in a way that actually makes them weaker.
Of course, if you're already a decently skilled writer, you'll likely notice when it does these things and reject silly suggestions. But if you're a skilled writer, do you need AI? Conversely, if you're still learning, might you end up taking its bad advice at times, and learning the wrong lessons?
Having AI read your work and help with grammar and style adjustments?
This often feels the least offensive to people at first glance.
However, I've noted it often falls short here, too--sometimes it picks up on real mistakes you made, but often points out mistakes that aren't really mistakes--exceptions, or stylistic choices, or wanting you to reword things in a way that makes less sense.Â
Again, if you don't already have a strong handle on grammarâŚyou might end up adopting even its incorrect or dubious suggestions.
Learning writing from an AI may not be the best idea.
To Be a Writer, You Need to Write
It feels like I shouldn't have to make this case, but maybe I do. Putting ideas into a machine and reading the stories it spits out, then asking it to make changes or even making some edits yourself, doesn't make you a writer. And it doesn't make you better at the actual task of writing--sitting down and doing the work, coming up with not only your theoretical story but turning that into a series of words, sentences, paragraphs, and chapters--learning how to measure your ideas, how to bring them to life through words, how to spill them across the page, free-flowing yet meticulously contained. The only way to get better at that is to do it. And practicing that skillset will serve you far better over time than having a machine do an inferior job at it for you.
Bottom Line: Stop Using AI For Creative Endeavors!
Using language models to predictive-text your way into telling your story is not only bastardizing others' hard work, but also teaching you bad habits, likely producing an inferior product, and cheating yourself out of the art itself.  Here's where I must admit to some bias. I did make some arguments above from a utilitarian standpoint--arguing that AI isn't actually all that good at writing. But what if it was? What if in some years, what it spits out is indistinguishable from something a human wrote, even to professional writers? I'd still take issue with the ethical issues of using human writers' work to do it--and with trampling on the craft, the human element of creation. As a writer and a reader, it simply leaves a sour taste in my mouth. Just don't do it. And support human writers. Thank you for attending my Ted Talk.
References
1-Nasr, M., Carlini, N., Hayase, J., Jagielski, M., Cooper, A. F., Ippolito, D., Choquette-Choo, C. A., Wallace, E., Tramèr, F., & Lee, K. (2023, November 28). Scalable extraction of training data from (production) language models. arXiv.org. https://arxiv.org/abs/2311.17035Â
2-Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Akhtar, N., Barnes, N., & Mian, A. (2024, October 17). A comprehensive overview of large language models. [2307.06435] A Comprehensive Overview of Large Language Models. http://export.arxiv.org/abs/2307.06435Â
