Please Stop Using AI for Writing

It’s a statistical model that loves average more than awesome.

Girl with typewriter, 1984 — Christian Boos

Imagine it’s 2002. And you are given a list of music genres:

hip hop | funk | soul | electronica

Which one doesn’t fit?

Or let’s go back to 1891 and look at another set of terms.

woman | modesty | children | radioactivity

Both lists end on a term that didn’t belong. But thankfully, there were people who didn’t care.

When OutKast first released their hit single ‘Hey Ya!’, many people thought: how strange — until everyone loved it.

When Maria Salomea Skłodowska — better known today as Marie Curie — thought to study physics and mathematics, people thought: how wrong — until there was no other option but to award her the Nobel Prize in Physics.

You may think I’m steering a bit off course here — but stay with me.

I think you could be the next OutKast or the next Marie Curie — and by letting AI interfere in your writing, you are putting that in jeopardy.

And it has everything to do with the lists I presented and how AI works fundamentally.

This is not just another polemic that screams ‘AI is bad’ or ‘AI is good’. I look at the very building blocks of how AI works, and from there explain how it subtly undermines your unique voice and ideas. I also identify a few use cases I believe are safe for AI’s help.

Let me start with examining why we should keep AI well clear of anything we write.

The fundamental flaw

AIs like ChatGPT essentially work like this: they take a list of words and calculate what is the most likely word that should come next. Think:

Cesar | was | emperor | of | …

AI will know: Rome!

This makes AI extremely competent at composing fluent prose, but there is a hidden danger, too. AI threatens your original thought each time you use it to write your articles, use it for editing, or even for critiquing your work.

Why you should not use AI to write for you

AI knows so well how to complete sentences because it has been trained on trillions of words of text. More or less, the entirety of human knowledge thus far.

The more text that humanity has produced on a subject, the more competent AI is when asked to write on that subject.

History of the Roman Empire: excellent.

History of your local bowling club: wanting.

To sum it up:

more text already available = better quality

less text already available = worse quality.

Unless you are in high school and given a boring assignment, why would you ever want to write about something that has been written about already ad nauseam?

Your published writing is best when it’s fresh, it’s new. And AI sucks exactly at this kind of content. So why bother?

Why you should not let AI edit your work

Now you may say: but what about when I use AI only to edit my content? What if I give a prompt that contains so much original thought that even AI will be forced to create a genuine article?

Sorry.

It’s still a bad idea.

AI will always write sentences that follow its fundamental rule. It will construct sentences word by word, choosing the most likely word to follow on from the previous one.

It will push your original text and thoughts towards average, yet quotable, prose, and ultimately diminish what is unique.

Let me give you an example.

Here an extract from Nietzsche’s Ecce Home I asked ChatGPT to ‘critique and edit’:

I am by far the most terrible man that has ever existed; but this does not alter the fact that I shall become the most beneficent. I know the joy of annihilation to a degree which is commensurate with my power to annihilate.

The edit proposed by ChatGPT:

I am the most terrible of men — yet I shall be the most beneficent. I know the joy of annihilation in proportion to my power to destroy.

Let’s just look at the first sentence:

I am by far the most terrible man that has ever existed.

Becomes

I am the most terrible of men.

The meaning stays more or less the same, yet ChatGPT took out exactly what makes the first sentence Nietzschean. Don’t let it do the same to you!

Innocent prompt leads to disastrous results — author’s screenshot from my zen-nified Windows

Why you should not let AI critique your work

This is actually the most devious one.

Even if you do not copy a single word from your chat with ChatGPT, your writing will be worse for consulting it.

This is because of the same two factors already discussed: AI pushes you towards average, and what is already known.

Let’s simply ask AI this innocent question:

What should I write about on Medium:

- how gentle productivity revolutionized my workflow

- why true crime has overstayed its welcome

*- how to resist the urge to only listen to music from your generation*

Do you know what it will pick?

Right.

The first one is your strongest bet: “how gentle productivity revolutionized my workflow.”

Do you want to read that?

It will subtly, and seemingly reasonably, push you toward what’s common and boring.

Don’t listen to it.

Write about what comes out of your soul like a rocket.

What to use AI for

The way AI is made fundamentally makes it a poor choice to use as a tool for writing. On the surface, it looks like it’s making our writing better, but ultimately, it is making it so much worse.

However, AI is a very powerful tool, and billions of dollars have been invested into making and improving it.

I still think there are things I can use AI for — safely, without it undermining my voice.

So here's what I still use AI for and why.

Synonyms

The models underlying AI use a few tricks apart from the ‘completing the list’ logic described above.

One of these is the idea to give words coordinates and place them on a map.

Let’s say we’ve got these two words:

mango, grape

Now, to give them coordinates, we cannot use latitude and longitude or x and y. Instead, we make up our own axes.

In this case, let’s use size and sweetness.

Now if we add more words to this graph — such as ‘papaya’ and ‘cherries’ — we’ll see that certain words are closer to one another. Such as ‘papaya’ and ‘mango’ will be neighbors, as will ‘cherries’ and ‘grapes’.

Since AI needs to deal with millions of words and phrases, things would become a bit crowded with just two axes. Therefore, words are commonly arranged not on two axes but on more than 4,000.

This makes AI very competent at figuring out what words are related to one another, and a great use case is asking it for synonyms.

It can even help us put precise words to concepts that we cannot quite name ourselves.

For instance, take this query:

What are synonyms for this feeling that the world doesn’t seem quite right?

AI will help us find such treasures as ‘uncanny’, ‘eerie’, and ‘liminal’ — great, powerful words for what I’m after.

Pattern recognition

AI’s ability to ‘complete the list’ along with the ‘4,000-dimensional map of words’ discussed in the previous section makes it very good at recognizing patterns.

You can provide AI with a list of attributes, and it will find the matching items in its vast store of knowledge.

For instance, if I supply these attributes:

• Humanist
• Female
• 16th century
• Publishing anonymously

AI will find Arcangela Tarabotti and Marie de Gournay.

Just one word of warning: AI has a tendency to ‘hallucinate’ matches that don’t really exist. Especially when there are few results to be found. Thus, the best way to use it is to let it find a set of matches, and then validate them using a web search.

Translation

Translation is what the class of AI that underlies ChatGPT was originally developed for.

This is not limited to only translating text from one language to another. AI can also translate prose into computer code or scientific papers into worksheets for 5th graders.

AI is also pretty decent at translating images and voice recordings into text.

I use AI extensively to transcribe my voice memos and handwritten notes from my Remarkable tablet, and it’s extremely accurate — even when dealing with the challenging environment of my German accent and terrible handwriting.

One of my ideas noted with the Remarkable tablet and filed with AI — image by author

If you have plenty of ideas like me but are too lazy to record them all, check out my workflows for voice memos, ‘Capturing Ideas the Lazy Way’ or how my wife digitizes her journal pages into Notion in ‘What Actually Helped Me Build a Journaling Habit’ — for which she now uses our integration Ink Lift.

Keep AI away from your writing

If there is one thing I would love for you to take away from this article, it is to stop using AI to write for you, edit your writing, or even critique it.

The way AI works makes it appear competent for these tasks, but I hope my arguments and examples have shown that it is anything but.

It will — subtly and cheerily — undermine your original thought and voice.

There are some things I still believe AI to be good and useful for. If you feel the urge to use AI, consider exploring them. But always keep in mind that using AI is a slippery slope.

Just like your Instagram feed, it is designed to keep you hooked. You start with one innocent small question: ‘Suggest synonyms for the feeling of not having eaten pasta in a long time’ (= pastalust) and end up 2 hours discussing your next 10 culinary-themed articles.

Bukowski didn’t need AI. Allow me the indulgence to compose a line in the spirit of his famous poem, so you want to be a writer?

if you have to ask a machine to think your thoughts and move your fingers for you, don’t do it.