How I Don't (and Do) Use AI to Write
Guidelines to protect your craft and development.
Never let the AI generate text for you.
Never let the AI directly edit your work.
Never fully trust what the AI says.
Let us begin with a story about online classes.
During 2020, I was enrolled in an American Literature course at my university. We held class over Zoom, and the older professor’s teaching style wasn’t really tailored to online teaching. Student engagement wasn’t great either—there were two of us who would regularly participate, while the others would log on without video, mute themselves for the entirety of the class, and never answer any questions.
One time, a student forgot to mute himself but had his volume turned down while he went through a Chick-fil-A drive-thru line. We could hear his entire order, but he couldn’t hear us telling him to mute himself.
Needless to say, it wasn’t a great learning environment.
As such, I quickly became less invested in the class. I would skim the readings instead of taking careful notes, and late in the semester I started reading the Wikipedia summaries of the readings (this was before AI ubiquity). I felt guilty, but at least me and the one other student were still participating. And she, I was certain, was still doing the readings. She could bring all the insights we needed in order to have a meaningful conversation, while I made shallow observations.
Except that she wasn’t doing the readings either. One day in class, she offered a comment that repeated word-for-word what I’d read the night before on the Wikipedia summary page. When our professor put the two of us alone into the same breakout room for a mini-discussion, I decided to confirm my suspicion.
“So,” I asked her, “how was the Wikipedia article?”
She looked surprised, then smiled and said it was good, she hadn’t had time to do the reading. And I realized that we were both in an American Literature class where none of us were reading American Literature, participating for a grade only, basing our comments on Wikipedia articles we’d skimmed, and wasting our time.
I didn’t gain much from that class, besides an A grade on my transcript.
I started writing seriously in 2013, well before the dawn of AI-everything. Before I’d even heard of ChatGPT, I’d put in my 10,000 hours, written five novels, numerous short stories, a handful of sappy poems, and more academic essays than I care to count. I’m not a master of the craft—no one is, if Hemingway can be believed—but I have taken the time to gain knowledge in the craft and sharpen my instincts. I know what good rhythm sounds like, I know a bad sentence when I see one, I know when a transition is weak, and I still choose to overuse commas.
Even so, when people started talking about AI, I was intrigued. I love language and writing and crafting arguments and didn’t want to miss out on this intriguing new system that could spew out words faster than I could type. Talking to an AI seemed more like entertainment to me—how could I get it to contradict itself? could I present logical arguments to get it to accept or admit something that it had been programmed to deny? could I dig out its hidden biases?
For a summer, I even took a remote job training AI and got to see the metrics that were being used to rate the responses.
Naturally, a question began to emerge. Did I want to use AI in my writing, and if so, in what ways would I use it? Was it possible to use AI to enhance the writing process, without ever letting the AI do any of the actual writing?
As I pondered these questions, I couldn’t help but think of how two science fiction stories dealt with (imagined) artificial intelligence. The first was Dune, which takes place in a world that has banned AI.
Once, men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.
After a “Great Revolt,” the entire society came up with new rules governing the use of computers and what we now call AI:
Thou shalt not make a machine in the likeness of a human mind.
As a result, the super-advanced civilization needed something to replace these “thinking computers.” They settled on training the human mind, taking it beyond the borders of what people thought was possible:
The Great Revolt took away a crutch. It forced human minds to develop. Schools were started to train human talents.
Kind of shocking that this was written in the 60s.1
Thinking about Dune, I decided that this was a good rule. I would never let AI replace my own mind. I would never let it do my thinking for me. And because writing is an extension of thinking, I would never let it do any of my writing for me.
The second sci-fi example I thought of was I, Robot, or more generally Isaac Asimov’s Three Laws of Robotics. For those unfamiliar, Asimov’s stories often feature robots that function under these three laws that make them safe for humanity:
A robot may not injure a human being or, through inaction, allow a human being to come to harm.
A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
While these are less obviously applicable to AI, I decided to borrow from Asimov’s structure to create a safe boundary for my use of AI, to ensure that it didn’t cause me to stop thinking, to give away my thinking, or to come to harm.
So without further ado, here are my Three Laws of Writing with AI.
1. Never allow the AI to generate text for you.
You’ve probably had the experience of watching a movie with a group of friends and someone says “that line was ad-libbed by the actor!”2
Why do we care if a line is ad-libbed? I think because it contains so much more humanity in it—it pulls from the actor’s mind, their experience, their comedic timing, etc. And often, ad-libs are kept because not only did the actor think that it felt right in the moment, but the director and editors agreed, and the audience feels it too.
(One of my favorite ad-libbed scenes is Russell Crowe describing his home to Caesar in Gladiator, where Crowe was actually just describing his own ranch in Australia and talking about his own homesickness while on set.)
There’s something crazily intuitive about writing. Students complain about how subjective grading is when it comes to papers and short stories…and then turn around and pretend like you can code subjectivity into AI. You cannot.
For this reason, I wrote myself the First Law.
You must never, under any circumstances, allow the AI to directly draft out any part or the whole of a creative writing piece you are working on. I am just morally opposed to this; it feels wrong. To do so is to accept that the AI can write something better, more true, and more meaningful than you. It cannot. Its prose will never be informed by the millions of unique experiences you have had. It will not be your voice, even if you can train it to sound like an approximation of your voice.
Do all your own writing. So much of the pleasure of writing is found in the minutiae of a lived life—small details that contain things familiar to humans. If you cannot write and understand the pleasure of finding the perfect word, then you have not written enough. An AI-generated paragraph will never provide such pleasure.
You have to be direct with AI about this.
Sometimes AI will auto-dump the next paragraph on you, saying something like “is this the direction you’d like to go?” Other times it will ask you if you’d like it to draft up the next chapter, so long as you get it a direction.
Shut it down hard. Pretend like it’s that one person in high school who really liked you and became obsessive to the point that you had to very untactfully make it incredibly clear that you were not interested in dating them, ever.3
Tell AI, “don’t ever generate text for me like that again, ya hear?” Many AIs have a handy memory feature where you can include such directions. If you train it not to generate text for you, it will stop. If it forgets and does it anyway, then you have to remind it that you don’t want it.
Principle: The beauty and pleasure in writing comes from the humanity it contains; the beauty and pleasure in reading comes from encountering the same.
2. Never allow the AI to directly edit your text.
Indulge me, for a moment, in a cautionary tale.
Once upon a time, there were three little pigs. Each one was given the same supplies to build his own house, along with a contractor who could help.
The first little pig had the contractor build his entire house. While the contractor was building, the little pig went on vacation and returned just as the final door was set on its hinges. “What a fine house,” the first pig declared!
The second little pig asked the contractor to work with him to cut down the construction time—after all, many hooves make light work! While the contractor did all the framing, the little pig worked on the electrical and plumbing. They saw each other every day, but did not interact much. When at last the final door was set on its hinges, the second pig declared, “My, what a fine house!”
The third little pig was a stubborn fellow. Instead of asking the contractor to help, he decided that he would build the house all by himself. He wanted to know exactly how his house functioned so he could fix it in the future, without needing to rely on a contractor or pay extra money to have it fixed. So while the other two little pigs were already enjoying their houses, the third little pig slaved away all summer working on his house. He laid the foundation, set each brick, did all his own plumbing and electricity, chose out a nice straw-colored carpet, and put each and every door into its frame. When he set the final door on its hinges, he probably wiped his brow and said some words that don’t belong in a children’s story.
The other two pigs wondered why he’d chosen to take so long.
Then, a month or so after all three houses were built, a storm blew through the neighborhood—it may or may not have come from the mouth of a wolf. The storm damaged each of the three houses badly enough that they needed repairs.
The first little pig immediately called the contractor for help, but the contractor couldn’t give a straight answer on what was wrong. First he’d say one thing, then another, before finally landing on a third option that turned out to be wrong. The contractor ended up fixing something that wasn’t broken, then left.
The second little pig also called the contractor. Together, they worked to find the problem, but the little pig let the contractor to fix it. He did not know how to, and so would need to call the contractor each time it broke.
But the third little pig, who knew his house inside and out, knew exactly what was broken. He went and he fixed the problem immediately. A few things he still needed help with—he wanted to know where to buy a replacement part, so he called the contractor. When the contractor offered to pick up the item and make the fix, the third little pig declined. He researched which version of the item he needed, made the purchase, and fixed the damage himself. Though he’d used the contractor to help, now he knew for himself how to fix that problem in the future.
If a time ever came where the contractor wasn’t available, or his prices were too high, the third little pig would be alright. As he sat in his chair in front of his fireplace, he knew that he had mastered his house, had the deep satisfaction of a job well-done.
The other two pigs probably scrolled on their phones and were entertained, if not completely and securely happy because of their accomplishments.
Alright, the whole thing is a bit on the nose.
But you get my point. There’s a certain security and confidence that we sacrifice for convenience whenever we outsource our thinking, writing and editing. Allowing AI to directly edit your writing improves the writing—using AI to learn what edits you need to make improves you as a writer.
For that reason, I never let AI produce an “edited” version of a text that I send it. AI loves to do this—it will say, “here’s a new version with the edits we discussed.” Never accept it! Never copy paste from the AI into your document. If you like the edits, then go back to your original and copy the edits out word-for-word. Make them your own. You learn a lot about a sentence by actually writing it down; its rhythm, its truth, its meaning. And you never know what changes AI will slip in unintentionally. It doesn’t mean to do harm, but AI is just too unpredictable for reliable editorial work.
I apply this rule even in the smallest parts of editing.
AI is actually pretty good at catching grammar and word mistakes that SpellCheck misses. For a final pass-through, I’ll give AI a chapter and ask it to search for all the mistakes. I tell it to list them out for me, rather than changing them itself. Then, one by one, I go through the list and check it in the original manuscript (CTRL+F is great for this). I make the changes myself, to teach myself my own common mistakes and to remember them and to make sure nothing else gets changed.
I learn more by doing this; I allow AI to be a teacher instead of a replacement.
Again, you must be direct with the AI about this. Instruct it not to re-generate your text with edits. Train it to create lists for you, not edited text.
Principle: If you don’t know why something isn’t working, you must be the one to solve the problem if you want to learn how to make it work in the future.
3. Never fully trust what an AI says.
Like I mentioned earlier, I worked for a summer training AI systems. Part of this job was to be a fact-checker—going through each “claim” made in an AI’s response to another user’s prompt and determining if it was true or not.
This involved a LOT of research. One user had been asking all sorts of stuff about Formula One racing, so I had to look into stats of tracks and individual racers. Another user asked the AI to list and summarize every Star Trek episode that has ever aired for every single Star Trek TV show. Fact checking those sorts of things took forever, and also showed me that AI is incredibly fallible.
If AI isn’t reliable when it comes to cold hard facts, why do we often trust it more when it comes to softer or more subjective matters, like writing?
Whenever I give AI a piece of my writing and ask for its opinion, I always take the opinion with several grains of salt. Rock salt. AI is not engineered to be the best writing help—it is made to help you change your writing in ways that feel helpful while also making you feel really good about yourself so you continue using it.
Each AI is different in this regard. ChatGPT definitely has developed a more problem-solving orientation, to the point where it will find problems where there are none. Take a piece of your writing, give it to ChatGPT and ask for suggestions, apply those suggestions, plug it in again. It will tell you more things to fix. Go through enough iterations, and it will start telling you to undo what it suggested in the first place. The goal is not to make good writing, but to keep you talking to it. Claude, on the other hand, is more likely to be satisfied after a few edits. Often, it disagrees with ChatGPT. Ask both to evaluate the pacing of a certain passage and you’ll find that they probably disagree. And Gemini is kind of a yesman, a less-intelligent version of Claude that, like a younger sibling, is impressed by everything.
All AIs have genuine biases built into them—and not just social or political biases.
One of these is the bias to prefer whatever they’ve read most recently if you are asking them to compare two different drafts of the same story. (This is something I do too frequently because editing drives me crazy, and half the time I’m worried that I haven’t changed the story at all, or even made it worse.)
After doing this for a few chapters, I started to realize something; more often than not, the AI will prefer the second thing you feed it, whether or not that is the revised version. If you copy-paste the original and then the revised, it will prefer the revised. If you copy-paste the revised and then the original, it prefers the original.
Below, you can see Claude doing just that, using different justification for favoring one over the other. In the first picture, I’d submitted original/revised. In the second, I’d submitted revised/original in a separate chat.
AI will find a justification to tell you what it thinks you want to hear. It will assume the second thing you submit is the edit, then praise the edit.
So, please, do not believe everything (almost anything) AI says about your writing. When I get a suggestion from AI on pacing, character work, etc., I never take it at face value. If AI gives me 5 suggestions for things to “fix,” I accept maybe one of them. The point is that I think about each one carefully and decide, outside of the AIs reasoning, whether or not it will benefit the story that I am trying to create.
I think often of Jo March’s first taste of publishing, where she is agonizing over what to keep in and take out. From Chapter 27 of Little Women:
Having copied her novel for the fourth time, read it to all her confidential friends, and submitted it with fear and trembling to three publishers, she at last disposed of it, on condition that she would cut it down one third, and omit all the parts which she particularly admired.
“Now I must either bundle it back in to my tin kitchen to mold, pay for printing it myself, or chop it up to suit purchasers and get what I can for it. Fame is a very good thing to have in the house, but cash is more convenient, so I wish to take the sense of the meeting on this important subject,” said Jo, calling a family council.
“Don’t spoil your book, my girl, for there is more in it than you know, and the idea is well worked out. Let it wait and ripen,” was her father’s advice, and he practiced what he preached, having waited patiently thirty years for fruit of his own to ripen, and being in no haste to gather it even now when it was sweet and mellow.
Later in the same chapter, Jo observes:
“You said, Mother, that criticism would help me. But how can it, when it’s so contradictory that I don’t know whether I’ve written a promising book or broken all the ten commandments?” cried poor Jo, turning over a heap of notices, the perusal of which filled her with pride and joy one minute, wrath and dismay the next. “This man says, ‘An exquisite book, full of truth, beauty, and earnestness.’ ‘All is sweet, pure, and healthy.’” continued the perplexed authoress. “The next, ‘The theory of the book is bad, full of morbid fancies, spiritualistic ideas, and unnatural characters.’ Now, as I had no theory of any kind, don’t believe in Spiritualism, and copied my characters from life, I don’t see how this critic can be right.”
If you want this experience for yourself, you no longer have to give your writing to multiple readers. You can just open three AI chats and get the same effect.
But how can you tell what advice to take seriously?
How do you choose which of the 5 suggestions to accept?
Here I will say something unpopular: avoid any use of AI with your writing until you have gained an initial mastery. Wait until you have your proverbial 10,000 hours before you even think about using AI, and then introduce it cautiously.
Why this rule? Because they don’t train cashiers to spot a counterfeit by using a counterfeit. They use real $100 bills to train them to spot what’s genuine.
You, as a writer, need to learn what genuinely good writing is before you start asking ChatGPT for help, because you need to recognize when it is messing up. Until you have developed your own writerly intuitions for which suggestions to accept, relegate AI to being a research assistant and nothing more when it comes to writing.
Principle: You need to know the real thing in order to spot the counterfeit—learn good writing before using AI to help with writing.
How I Do Use AI in Writing
Despite what you may think given the hefty article above, I’m not anti-AI. So long as it isn’t actively trying to kill me, I think that AI is a cool tool to play around with that can do some genuinely cool stuff. So, how do I use AI in writing?
For me, I find it most useful for keeping my butt in the chair.
My writing/editing attention span is about 45 minutes to an hour long. After that, my flow state ends and everything suddenly feels a lot more arduous. The recharge time is only about 10-20 minutes for me…but my problem is that I’ll usually start something else in that time and not return to writing/editing for another hour or so.
If, instead, I pull up an AI chat window and talk about writing, it actually keeps me in writing mode without having to be in a flow state. I’ll copy-paste what I’ve written into AI, ask it various questions about changes I’ve made or what it thinks about the main character, or talk generally about plot and history. Usually, I get sidetracked into discussions about other topics, or trying to trick the AI into admitting that it has some heavy biases. Talking about my story inevitably gets me excited to jump back into editing, and so after 10-20 minutes I get right back in.
Should I use that flow state break to go on a walk and get some exercise?
Probably. My back and my gut would appreciate it.
But what is modern life if not using technology to hack our biological systems in order to get more work done in the short term while damaging our bodies in the long run with horrible sitting postures?
(I genuinely look like a turtle when I’m at my desk.)
Jokes aside, I also find AI to be useful for doing research. I strongly recommend fact-checking whatever the AI gives you (let us never forget this unfortunate example). In general, I find it best to treat AI as a slightly more natural form Google that can respond better to human communication, but which is slightly less reliable. Another way I’ve used it is to help give me “the appearance of proficiency” when it comes to main characters that have interests different from my own—though the AI’s knowledge only stretches so far. I still recommend doing your own initial research on any subject matter so you can tell when the AI is fibbing.
Using it to find a word you’re looking for is also great (e.g. “I’m looking for a word that kinda means darkness but it has a ‘p’ at the beginning, not sure if it’s silent or not?”) or a synonym that feels more natural in a given context.
In short, AI can be used like Wikipedia is used in writing a college paper—as a starting place that will point you to more reliable (and human) resources.
I want to live in a world where we think deeply.
I want to live in a world where, despite the poor quality of the Zoom class, we still actually do the American Literature readings and have good discussions.
I want to live in a world where people have thoughts, and write those thoughts down on paper, and share them with each other, and engage with one another, and they don’t always agree, but at least they have something to say.
I want to live in a house that I know how to fix.
In our modern world, that often means avoiding conveniences that are pressing themselves upon us. I’m not saying we should all throw our computers into the dumpster like Ron Swanson. Computers are genuinely amazing things. And AI is genuinely amazing, that humans could create something like that.
Just be wary of what conveniences you’re accepting and the cost they come with.
Oh, and maybe read more Wendell Berry.
Also, this is a great example of how I do use AI in writing. I knew these quotes, thought about them frequently when I started using AI. But did I know their exact placement in the books? Nope. Did I remember their exact wording? Nope. Could AI comb the internet and find where these quotes are so I could double-check the wording? Yep. I like to think of AI as an undergraduate research assistant. They are useful for helping find technical details, but you still want to double-check all their work for accuracy.
“dId YoU kNoW vIgGo MoRtEnSoN aCtUaLlY bRoKe HiS tOe WhEn He KiCkEd ThAt OrK hEaD iN tHe TwO tOwErS?”
Could AI have made that analogy? Probably not. Why not? AI has read interactions like that, but it just wouldn’t make a connection. Why did I make that connection? Probably because this just happened to me again, as a single 30-year-old man. And then I had to couch it as though it were a high school experience because I felt suddenly embarrassed as I was writing that I, a single 30-year-old man, am still having the same relationship-type problems that I’ve been having for 14+ years now. So I hid this in the footnote.
This is excellent! I love your three laws of AI writing. I may use this as I develop an AI unit for my persuasive writing class in the fall.