Why write, if AI can?


I've been struggling with this big question for a while.

In November 2022, before ChatGPT came out, I was working on a blog post called, "I wrote this post with AI".

I never released it because the output wasn't good enough – I was using the GPT-3 playground so I couldn't ask 'follow ups' like you can today. But one good thing did come out of it: I started thinking about the bigger question of "what's the point of writing anything if AI can write too, sometimes better?"

Then, a month later, ChatGPT went mainstream. Suddenly, everyone was using GPT-generated text for something. Didn't take long for AI generated texts to start flooding the Internet. For me, the question became even more pertinent: why write, if AI can?

After a year of reflection, I've found two good answers that still keep me writing:

1. Write because it makes you think

The process of writing, editing, and rewriting makes you think and form thoughts coherently.

And, improving sentence structures and recalling words will exercise your mind and could push you in a state of flow.

2. Write because it changes how you think

If you're trying to learn about a new topic or change your beliefs about something, writing can be more effective than reading. It reinforces whatever you're writing about, and helps you reason better.

Personally, I also use writing to question my thoughts. You can do this with ChatGPT by asking it to act like a Socractic child, but I find it more fun to play my own devil's advocate.

Predictions on writing in a GPT-full world

Who will we write for?

Less and less for SEO and marketing and things that computers and search engines can crawl. More and more for ourselves and for small niches of readers. Think journals and private Substacks.

I also have a prediction that technologies that keep writings private (which means language models can't train on authors' text) will flourish. Email lists and encrypted text formats could make a resurgence.

What will be worth writing?

Anything that surprises the reader. That is a good heuristic for what distinguishes AI-generated text from human-written text.

In the early months of 2023 I contributed to a leading AI detector app. I learnt that a simple but effective way to detect AI generated text is to pass tiny chunks of your text through an AI model and see whether the next word it would've predicted matches the one that you actually wrote.

Measure this kind of stuff across the entire text and you can get a score for how "surprised" the AI model is with the text. Higher "surprised" scores = "this is not what the AI model would've written" = this is probably human-written.

Kangaroos! An AI model is very unlikely to generate that word given the context above. There's no mention of related words like Australia or marsupials. But humans might. We write using creativity, not probability.

This gives you a nice little insight into "what will be worth writing". Generic texts that read like they came off of Wikipedia will not be worth writing. But authentic writing that surprises the reader will. Both search engines and humans will feel more confident that another human wrote it.


Conclusion by ChatGPT

"creativity that AI can't replicate" – I agree!

My picture

Written by Aryan Bhasin