Research notes on AI & writing

Over the past month, we conducted a bunch of interviews with friends and writers across diverse disciplines: children’s book authors, Substack personal essayists, AI policy researchers, and nonfiction writers. We hoped to capture a little smidge of the vibrancy of what it means to be a human writing, and how our friends envision their identities and craft evolving alongside AI.

Collected notes below:

Why write?

To think:

M: "I write to think."

J: "My target audience is my future self. How did I handle/overcome this situation?"

J: “I view writing very instrumentally – some write to entertain, but for me writing is good bc it helps me to clarify and communicate ideas. For me, good writing is close to novel scientific discovery.”

For joy & beauty:

R: "With writing, you don't write to perform some productive output. Most writing is antiproductive. So that's where the good part of it is. If the ultimate outcome of writing is that you write something beautiful and move something else, then it already has this antibody built in. It's already not built into AI writing."

M: "It's kind of artistic."

We think that the intellectual and aesthetic dimensions of writing are closely intertwined —

What we found interesting was the consistent statement that writing emerges from necessity and life force.

M: "Almost always, my articles start from an itch. I usually write out of necessity. I want to scratch the itch. This is more sparse and intermittent, but when I write it has more weight."

E: I don’t actually seek out new ideas. Most of them come to me when I lie down. Then, I have to move immediately to writing.

Workflows

Nonlinear, or linear?

M: "Putting ideas into nodes is so much more useful, because writing is not a linear process, but reading is. Sometimes people fall into the trap of thinking that they need to write how they read. But that's not true, especially in nonfiction. You're trying to link together nodes, and you don't know how they relate... You're in the world of ideas and there's nothing linear about an idea."

E: "I try to get it right in one go. No one's really going through paragraphs of point evidence explanation—when I'm writing it always comes more naturally. Flows more naturally."

Friction and speed:

J: "When I need to jot down something really quick, I just get it down into Google Docs. Digital tools are more for 'capturing thoughts,' and writing is more for 'digesting' them... But when I write, I write with my hands. The latency is important. Concise expression is important. With a pen, you need to get it right the first time."

When to collaborate?

Protecting the creative core:

M: "I actually find that writing for me is a very personal process. I would rather that people don't get involved at all. When they do get involved, I find that the article gets delayed: I need to wait on feedback, want to perfect things. If I don't have people involved, then I can just get it out."

J: "The digestion and writing is relatively solitary. I don’t show other people my journal, but I do talk a lot about the ideas that I’ve gleaned from it."

Strategic collaboration before and after:

A: "Talking to people is 40% of my inspiration... Before and after the writing should be more collaborative, because it's when your ideas are starting to form, and also when you're looking for parts that should be missing. But the part in the middle, for me, is the part I'd like to keep private."

The Challenge of Finding Good Feedback

Writers consistently noted the difficulty of finding truly critical reviewers. People are too nice!

M: "Reviewing other's works only works when I have a fundamental, deep connection with them... there should be a point where we can establish connection with that person."

A: "It's really hard to find people to read your work and respond meaningfully to it, because that also requires a level of understanding of you as a person to an extent; and also just patience."

R: “I’ve done a bunch of writing workshops over the years… and you’d hope that they give feedback with your writing… help make it better. But that does not end up being the case most of the time, bc ppl have a lot to read, and most people aren’ very good at editing. they’re nice most of the time. they’ll just give you a bit of surface-level criticism, and that’s really what you have to go along with.”

But when it’s good feedback, it’s amazing.

M: “The writing group was awesome… Getting input can change a trajectory of an article.”

On AI: what works

Feedback and structure:

R: "I've been trying to use chatbots as a stand-in for workshop criticisms, and I think it does better than your average person that you're in a workshop with. It's able to give you a summary of the themes, and you can figure out if what you're writing is making sense."

Mechanical tasks:

R: "It's great at copywriting stuff. Let's say that I have this passage that I now want to switch to past tense. It'll do it for me; anything in general a copywriter will do."

R: “The tasks that require the least amount of rumination in your head are best suited to offloading to AI. Anything else, and you’re shooting yourself in the foot, almost.”

J: “It does things that are impossible for us. Would take us ages to understand thousands of words of notes, but that becomes trivially easy with LLMs.”

Brainstorming partner, especially for nonlinear thinking:

M: "A lot of the bigger ideas I tackle—I almost always talk with Claude... it's just a genius. It's so smart... What I've done with Claude before is ask it to draw conceptual diagrams and concept maps for my essays, and it'll do it."

The accessibility of LLMs is also useful chiefly:

E: "Claude helps a lot with momentum—when you need to go go go, at any time in the day. Getting bullet points from Claude fits in easily with my flow — I don’t have to wait on anyone."

Good for summarization and BS writing:

E: “there’s regular writing, where I want to get good, and then there’s bullshit writing. For BS writing, I use Claude for everything.”

On AI: what doesn’t

There seems to be a tradeoff between statistics and creativity. AI writing tends towards the latter, so it nukes the former.

R: “I’m looking at my bookshelf right now, and I have this book written by someone, where the sentences are really outside the mean. They’re not succinct; the words are crammed together. And the effect is that the book is really wonderful to read, but LLMs would neuter it.”

J: “AIs are pretty good at synthesis, but pretty bad at coming up with new ideas.”

And when we really want to read, we want to read real people — and for writing, we can tell:

R: “When you write something, it’s not the arrangement of the words that matter. It’s the person thats putting it into the paper. And that’s a critical aspect of creative writing.”

A: “I absolutely don’t want to include any AI-generated content in any post or anything I claim is my own. I feel like personally, I can read something and feel, “that’s really AI coded”.”

J: “AI writing has the look, but not the smell.”

I don’t know how people can use AI and feel it matches their text, which is interesting. but sometimes i feed my stuff to aI and let them tell me what big themes or general vibe their getting, and see whether it matches their expectations. when i’m writing there’s so much more i could add, stuff i leave out of the camera. but i want others to get what im trying to say.

The value of doing the hard thing yourself:

R: “Part of the work around good writing is making those connections yourself... There’s something to be said about doing that manual labor of thinking, which creates better writing. My worry is that when you offload too much of this creative writing to external agents where it doesn’t leave you time to just dwell on it, you’re gonna lose a big part of what makes writing good.

How can we use AIs better?

We should be prompting differently:

L: “Generally, AI exposes the problem in the way that you do things – tech isn’t inherently evil, but it just starts catering to capitalistic incentives for efficiency.”

J: “With reasoning models, it’s the case that a lot of times, you’re not giving it enough context and instructions. Claude has great synthesis and understanding skills. The gap is often the correct delivery of context. I also think to better help in writing, they’d need agency.”

M: “I would argue that AI isn’t inherently narrow and productive. It’s just raw intelligence, so it can be used for divergent, creative thinking, lateral thinking. It’s about how you prompt it.”

LLMs should be more obstinate:

R: "The only way to replace a human editor is a LLM that has its own set of beliefs, and I can't get around it... I don't want to be able to game it. I don't want to be able to fit the feedback to where I want it to be... A virtual workshop, where every person is its own LLM who's obstinate in their opinions. Then it's on you to synthesize."

LLMs should prioritize human connection:

M: "Some sort of hub. I write and upload my writing... when there are other people that are writing about the same thing as you, and it connects you with people similar to you."

On Taste

Writers worried about AI's potential effect on readers' ability to distinguish good writing.

R: "This feedback loop pollutes our heads with this cold writing. Before, you could have faith and trust that there’s someone in that writing. But now you can read something and people will say, 'that sounds AI generated.'"

R: "Not everyone has taste. Probably people will just forget taste as they once had in the earlier years, and that'll end up with us having a lot of shit writing. We need to all be a little more highbrow, a little more discerning about what we consume."

However, some remained optimistic:

R: "Even now, there's an oversupply of writing... that supply has never really been much of an issue, and it's always been discoverability—how do you discover what's good? I'm optimistic that the AI slop won't break through—it'll just go into the background."