January 2, 2025

How to use AI for writing

If you let AI do the writing, you'll get slop—or popcorn. But if you use it as a tool in your writing, you can write faster, and better.
January 2, 2025

How to use AI for writing

If you let AI do the writing, you'll get slop—or popcorn. But if you use it as a tool in your writing, you can write faster, and better.
January 2, 2025
Briana Brownell
In this article
Start editing audio & video
This makes the editing process so much faster. I wish I knew about Descript a year ago.
Matt D., Copywriter
Sign up

What type of content do you primarily create?

Videos
Podcasts
Social media clips
Transcriptions
Start editing audio & video
This makes the editing process so much faster. I wish I knew about Descript a year ago.
Matt D., Copywriter
Sign up

What type of content do you primarily create?

Videos
Podcasts
Social media clips
Transcriptions

AI writing gets criticized a lot — sometimes deservedly, sometimes not. But no matter how you plan to incorporate AI into your writing process, understanding why it produces the outputs it does, plus its strengths, limitations, and quirks, can help you get the most out of it.

It’s all about making the process faster and smoother so you can spend less time wrestling with AI, and letting you focus on what matters: getting your ideas into the world.

Not all AI are created equal

There are dozens of open- and closed-source models out there, and each has its own quirks, strengths, and style. Some people swear by Anthropic’s Claude, while I tend to gravitate toward Open AI’s ChatGPT. We can each choose the best tool for us based on our personal preferences and what we’re trying to create. We can choose not to choose: you might find yourself switching between tools depending on what you're working on.

Size isn't everything, but it often helps. Larger models tend to be more sophisticated and generate higher quality output, but that's not a hard and fast rule.

I recommend you try the models side-by-side on the same task and see which one works best. You can see my testing of Claude vs. ChatGPT here. 

AI is best at writing what it knows

Large language models are trained on massive text datasets, and their best outputs come from close alignment with their training data. In other words, they're better at reproducing the structure, tone, and logic of what they’ve been fed — sticking to what they "know" always yields stronger results.

Although each model’s training set is unique, many rely on the Common Crawl — a massive dataset of internet content that’s been growing since 2008, collecting billions of new webpages every few months.

But here’s the catch: the Common Crawl isn’t the whole internet. It’s not even a balanced slice of it. At 9.5 petabytes (1,024 terabytes) of data, it’s enormous, but it’s biased by design. The dataset is built by “crawling” links between sites, which means popular, highly-linked pages are overrepresented while obscure, unlinked ones are left out.

What does this mean for AI tools like ChatGPT? They’re pretty good at mimicking the kind of content that dominates this dataset: marketing copy, corporate websites, blog articles. So if you're using AI to write that kind of content, the output might already be close to what you need — after all, it’s packed with writing from professional copywriters and top bloggers over the years. But if you're writing something quite different, you’ll need to teach your AI to write like you

Vocabulary and format

LLMs work by continuing the text word-by-word, and each word is chosen by a statistical relationship. This ends up giving it a strange quality you've no doubt noticed: the words it chooses are usually common, with nothing resembling the efflorescence you sometimes get from creative human writing (I got "efflorescence" from Edgar Allen Poe, who referred to an "efflorescence of language" in poetry: language that was flowery, or overly rich and colorful). Using simpler words is not necessarily a bad thing — after all, we want the writing to be clear and not stuffed with overcomplicated vocabulary — but it can strip out what's unique about the voice of the writer too.

AI vocabulary has quirks of its own, and it can also overuse words that many humans are unlikely to. That's because LLMs go through several other steps to add the conversational chat capability, and finally to learn to generate good responses through Reinforcement Learning with Human Feedback (RLHF). (Looking at you, Delve). 

This gives it other formatting quirks too, like its tendency to try to put everything into "Short Title: One to three sentences about that short title." Basically, whatever the humans giving the feedback like gets baked into the model. So if you don't want that, you need to give explicit instructions to move the model away from its default tendencies.

AI can't play jazz

The token-based continuation of transformers used to create LLMs give it another, stranger, macro effect: the sentences it generates are often similar in length.

For instance, here's the first paragraph from ChatGPT 4o's version of this article (Prompt: "Can you please write me an article about AI writing?"): 

Artificial Intelligence (AI) has become a transformative force in modern life, and one of its most fascinating applications is in the realm of writing. (24 words) From drafting concise emails to crafting intricate novels, AI writing tools are reshaping how we communicate and express creativity. (19 words) These tools leverage natural language processing (NLP), a subset of AI that enables computers to understand and generate human-like text. (20 words) Using deep learning models like OpenAI’s GPT (Generative Pre-trained Transformer), these tools analyze massive datasets to learn the structure, tone, and context of language. (24 words) The result is content that is coherent, relevant, and stylistically adaptive, making AI writing an increasingly indispensable resource across various industries. (21 words)

Here's a graph of the sentence lengths for the whole article it wrote.


If you let AI do the writing, sentence length stays fairly consistent, which isn't great.

Although the sentence lengths start varying a little more as you get farther into the article, there are only a couple of short sentences.

Meanwhile, here are the results for the first few paragraphs of the article that won the 2024 Pulitzer Prize for feature writing. The difference is striking. The Pulitzer winner alternates between short sentences, medium-length ones, and the occasional sprawling “monster” sentence.


When a good human writer is writing, you get varied sentence lengths.

Just like great music, good writing is surprising and repetition is deliberate. If you look into the parts of the Pulitzer winner, the author deliberately uses similar sentence lengths to create a rhythm in the piece. So if you are using AI writing, listen to its rhythm. We want momentous, not monotonous.

AI writing is like popcorn

When I think of AI writing, here's the image that comes to mind: an old popcorn machine in a movie theater dramatically shooting kernels all over the place.

Here's what I mean: often the AI will have good points to make in what it writes, but it brings in new ideas seemingly at random, and doesn’t fully thread together the arguments to make a resonant point.

My theory is that this largely stems from LLMs' challenges with logic. Because these models struggle to reason through an argument well, they also struggle to structure written prose in a way that's compelling. They don't take you on a narrative journey like good writing should.


Because writing is linear—you read from the beginning to the end—your ideas have to be smushed into a line. To make that line worth following, you need to be really clear about what goes where on that line. I've found that when I'm getting AI to help me write, I have to do this myself, without AI. (Also note that AI doesn't usually give great criticism on this, either; so if you're using AI as an editor, you probably need to spend more time here.)


I will give you a caveat, though: we're seeing a lot of improvements in the ways AI systems reason, so this might be temporary. But until then, even if you have your argument figured out, the argument structure is going to be up to you.

If you didn't take the time to write it, why should I read it?

I think a lot of the angst about AI writing really boils down to this: often people think that the writers using AI push a button to create meaningless AI slop. And as a writer who uses AI heavily, let me tell you: I hate slop too.


I heard this quote once and I think it speaks volumes: "If you didn't take the time to write it, why should I take the time to read it?" Readers can get frustrated with AI-generated writing because it is no longer a given that there is intentionality with what is communicated between the reader and the writer. The eunoia2 is missing.


So to overcome the growing skepticism around AI writing, it's worth really thinking about what I call your creative core — that is, what is the part of the process that, if you gave it up, you would no longer consider yourself the creator of the output. This is different for everyone and figuring it out for yourself lets you hone in on the parts that matter to you, so you have a good idea of where to spend your time using AI and when not to.


The other aspect is the relatively common belief that AI can only badly copy existing works — see Ben Affleck's comment above. While I fundamentally disagree that it can only be a bad copy of something human-created (for example, my 100% AI generated creative fiction piece Regenerate absolutely does not copy any existing works), I agree that a lot of AI writing does indeed badly copy existing writing. And so developing and honing your own voice is even more essential now that everyone has access to AI.


That takes time and effort. Many of the best writers I know who are using AI got there by spending an enormous amount of time and energy thinking about their voice, and teaching the AI to use it as closely as possible so that they could save time on edits and spend their time on the ideas, the structure, and plot and character development, etc. It's not insurmountable, and here are some tips from a professional ghostwriter.


2Yep, I found “eunoia” on that rare words list too and I thought it was too interesting not to include. It refers to the audience’s goodwill toward the speaker, their trust that they have the audience's best interest at heart. 

How to write well with AI

I think it comes down to three things:


  • Know your voice
  • Know exactly what you want to communicate
  • Figure out the structure to effectively communicate it


Writing is hard, and doing it well takes practice. AI can make that process faster in some ways, but it also forces you to think about your writing in a much more sophisticated way.


Beyond the output, though, in this moment where AI can do so many things for us, we need to be more intentional about what skills we want to build and why they are important. Writing helps you think things through, so it might be worth thinking about whether you want to write something yourself even if AI tools would make it faster.


Ultimately, the question isn’t just how AI can help you write, it’s figuring out why you’re writing in the first place. AI can accelerate the process, push you to think deeper, and even surprise you with creative solutions. But it can’t replace the intention behind your words or the clarity that comes from wrestling with your own ideas.

Briana Brownell
Briana Brownell is a Canadian data scientist and multidisciplinary creator who writes about the intersection of technology and creativity.
Share this article
Start creating—for free
Sign up
Join millions of others creating with Descript

How to use AI for writing

AI writing gets criticized a lot — sometimes deservedly, sometimes not. But no matter how you plan to incorporate AI into your writing process, understanding why it produces the outputs it does, plus its strengths, limitations, and quirks, can help you get the most out of it.

It’s all about making the process faster and smoother so you can spend less time wrestling with AI, and letting you focus on what matters: getting your ideas into the world.

Not all AI are created equal

There are dozens of open- and closed-source models out there, and each has its own quirks, strengths, and style. Some people swear by Anthropic’s Claude, while I tend to gravitate toward Open AI’s ChatGPT. We can each choose the best tool for us based on our personal preferences and what we’re trying to create. We can choose not to choose: you might find yourself switching between tools depending on what you're working on.

Size isn't everything, but it often helps. Larger models tend to be more sophisticated and generate higher quality output, but that's not a hard and fast rule.

I recommend you try the models side-by-side on the same task and see which one works best. You can see my testing of Claude vs. ChatGPT here. 

AI is best at writing what it knows

Large language models are trained on massive text datasets, and their best outputs come from close alignment with their training data. In other words, they're better at reproducing the structure, tone, and logic of what they’ve been fed — sticking to what they "know" always yields stronger results.

Although each model’s training set is unique, many rely on the Common Crawl — a massive dataset of internet content that’s been growing since 2008, collecting billions of new webpages every few months.

But here’s the catch: the Common Crawl isn’t the whole internet. It’s not even a balanced slice of it. At 9.5 petabytes (1,024 terabytes) of data, it’s enormous, but it’s biased by design. The dataset is built by “crawling” links between sites, which means popular, highly-linked pages are overrepresented while obscure, unlinked ones are left out.

What does this mean for AI tools like ChatGPT? They’re pretty good at mimicking the kind of content that dominates this dataset: marketing copy, corporate websites, blog articles. So if you're using AI to write that kind of content, the output might already be close to what you need — after all, it’s packed with writing from professional copywriters and top bloggers over the years. But if you're writing something quite different, you’ll need to teach your AI to write like you

Vocabulary and format

LLMs work by continuing the text word-by-word, and each word is chosen by a statistical relationship. This ends up giving it a strange quality you've no doubt noticed: the words it chooses are usually common, with nothing resembling the efflorescence you sometimes get from creative human writing (I got "efflorescence" from Edgar Allen Poe, who referred to an "efflorescence of language" in poetry: language that was flowery, or overly rich and colorful). Using simpler words is not necessarily a bad thing — after all, we want the writing to be clear and not stuffed with overcomplicated vocabulary — but it can strip out what's unique about the voice of the writer too.

AI vocabulary has quirks of its own, and it can also overuse words that many humans are unlikely to. That's because LLMs go through several other steps to add the conversational chat capability, and finally to learn to generate good responses through Reinforcement Learning with Human Feedback (RLHF). (Looking at you, Delve). 

This gives it other formatting quirks too, like its tendency to try to put everything into "Short Title: One to three sentences about that short title." Basically, whatever the humans giving the feedback like gets baked into the model. So if you don't want that, you need to give explicit instructions to move the model away from its default tendencies.

AI can't play jazz

The token-based continuation of transformers used to create LLMs give it another, stranger, macro effect: the sentences it generates are often similar in length.

For instance, here's the first paragraph from ChatGPT 4o's version of this article (Prompt: "Can you please write me an article about AI writing?"): 

Artificial Intelligence (AI) has become a transformative force in modern life, and one of its most fascinating applications is in the realm of writing. (24 words) From drafting concise emails to crafting intricate novels, AI writing tools are reshaping how we communicate and express creativity. (19 words) These tools leverage natural language processing (NLP), a subset of AI that enables computers to understand and generate human-like text. (20 words) Using deep learning models like OpenAI’s GPT (Generative Pre-trained Transformer), these tools analyze massive datasets to learn the structure, tone, and context of language. (24 words) The result is content that is coherent, relevant, and stylistically adaptive, making AI writing an increasingly indispensable resource across various industries. (21 words)

Here's a graph of the sentence lengths for the whole article it wrote.


If you let AI do the writing, sentence length stays fairly consistent, which isn't great.

Although the sentence lengths start varying a little more as you get farther into the article, there are only a couple of short sentences.

Meanwhile, here are the results for the first few paragraphs of the article that won the 2024 Pulitzer Prize for feature writing. The difference is striking. The Pulitzer winner alternates between short sentences, medium-length ones, and the occasional sprawling “monster” sentence.


When a good human writer is writing, you get varied sentence lengths.

Just like great music, good writing is surprising and repetition is deliberate. If you look into the parts of the Pulitzer winner, the author deliberately uses similar sentence lengths to create a rhythm in the piece. So if you are using AI writing, listen to its rhythm. We want momentous, not monotonous.

AI writing is like popcorn

When I think of AI writing, here's the image that comes to mind: an old popcorn machine in a movie theater dramatically shooting kernels all over the place.

Here's what I mean: often the AI will have good points to make in what it writes, but it brings in new ideas seemingly at random, and doesn’t fully thread together the arguments to make a resonant point.

My theory is that this largely stems from LLMs' challenges with logic. Because these models struggle to reason through an argument well, they also struggle to structure written prose in a way that's compelling. They don't take you on a narrative journey like good writing should.


Because writing is linear—you read from the beginning to the end—your ideas have to be smushed into a line. To make that line worth following, you need to be really clear about what goes where on that line. I've found that when I'm getting AI to help me write, I have to do this myself, without AI. (Also note that AI doesn't usually give great criticism on this, either; so if you're using AI as an editor, you probably need to spend more time here.)


I will give you a caveat, though: we're seeing a lot of improvements in the ways AI systems reason, so this might be temporary. But until then, even if you have your argument figured out, the argument structure is going to be up to you.

If you didn't take the time to write it, why should I read it?

I think a lot of the angst about AI writing really boils down to this: often people think that the writers using AI push a button to create meaningless AI slop. And as a writer who uses AI heavily, let me tell you: I hate slop too.


I heard this quote once and I think it speaks volumes: "If you didn't take the time to write it, why should I take the time to read it?" Readers can get frustrated with AI-generated writing because it is no longer a given that there is intentionality with what is communicated between the reader and the writer. The eunoia2 is missing.


So to overcome the growing skepticism around AI writing, it's worth really thinking about what I call your creative core — that is, what is the part of the process that, if you gave it up, you would no longer consider yourself the creator of the output. This is different for everyone and figuring it out for yourself lets you hone in on the parts that matter to you, so you have a good idea of where to spend your time using AI and when not to.


The other aspect is the relatively common belief that AI can only badly copy existing works — see Ben Affleck's comment above. While I fundamentally disagree that it can only be a bad copy of something human-created (for example, my 100% AI generated creative fiction piece Regenerate absolutely does not copy any existing works), I agree that a lot of AI writing does indeed badly copy existing writing. And so developing and honing your own voice is even more essential now that everyone has access to AI.


That takes time and effort. Many of the best writers I know who are using AI got there by spending an enormous amount of time and energy thinking about their voice, and teaching the AI to use it as closely as possible so that they could save time on edits and spend their time on the ideas, the structure, and plot and character development, etc. It's not insurmountable, and here are some tips from a professional ghostwriter.


2Yep, I found “eunoia” on that rare words list too and I thought it was too interesting not to include. It refers to the audience’s goodwill toward the speaker, their trust that they have the audience's best interest at heart. 

How to write well with AI

I think it comes down to three things:


  • Know your voice
  • Know exactly what you want to communicate
  • Figure out the structure to effectively communicate it


Writing is hard, and doing it well takes practice. AI can make that process faster in some ways, but it also forces you to think about your writing in a much more sophisticated way.


Beyond the output, though, in this moment where AI can do so many things for us, we need to be more intentional about what skills we want to build and why they are important. Writing helps you think things through, so it might be worth thinking about whether you want to write something yourself even if AI tools would make it faster.


Ultimately, the question isn’t just how AI can help you write, it’s figuring out why you’re writing in the first place. AI can accelerate the process, push you to think deeper, and even surprise you with creative solutions. But it can’t replace the intention behind your words or the clarity that comes from wrestling with your own ideas.

Featured articles:

Video

How to make a documentary on YouTube that's worth watching

Learn the 4 main types of documentary you'll encounter on YouTube and understand how to craft a narrative that grabs the viewer's attention.

Video

5 ways to boost audience retention on YouTube from real creators

Learn how MrBeast and other high-performing creators boost audience retention on YouTube with 5 tricks to keep viewers watching.

AI for Creators

The best prompt for AI brainstorming, according to a study

AI tools are notorious for coming up with a lot of similar ideas. Researchers have now discovered a prompt to make those ideas more diverse.

Related articles:

Share this article

Get started for free →