Writing In the Age of AI

Sometime in the near future, most of what we read online will be written by the machines (LLMs). The question is: who will write for the machines?

Just as we once had to scour the ocean floor for untainted steel from shipwrecks, one day we may have to salvage “unpolluted” (non-AI-generated) text from the digital ocean to train the next generation of large language models.

Titanic Wreck

Pre-war Steel The process of making steel once involved forcing air into pig iron. The detonation of nuclear bombs in the 1940s and '50s contaminated the air on our planet with traces of nuclear fallout. For many years, the level of contamination was sufficient to give the tainted steel a weak radioactive signature. This radioactive signal made the contaminated steel unsuitable for use in radiation-sensitive instruments.
Today the background radiation in our atmosphere has, thankfully, reduced below the natural level, but for many years people had to salvage the so-called *pre-war steel* — steel without the radioactive signature — from shipwrecks for making sensitive equipment like Geiger counters.
Pre-AI TextThis whole idea of having to salvage pre-war steel is very interesting and has a direct analogy to AI and LLMs. Consider this: all large language models that we use today are trained directly from human-generated text on the internet. Today, most people in the industry have started using LLMs to do most of their writing. Soon, most of the internet will be filled with LLM-generated text. This raises an obvious, unsettling question: where will we find clean, unpolluted text to train tomorrow's LLMs?
If we scrape the internet again to collect data for training new LLMs — or even use synthetic data — then we would be feeding LLMs their own echoes. We risk a recursive loop — an ouroboros of thought where the model continually mimics itself until it fades into noise. A mirror reflecting another mirror! A snake eating its own tail!
Writing for AIIn the not-too-distant future, we may need to imitate those steel scavengers of the past, hunting for pre-LLM text in the deepest corners of the internet to train future LLMs.
And that's one of the reasons (not the main reason — see below) why I encourage everyone to write more than they currently do. We'd be doing a service to future LLMs, and hence to the future of humanity!
Writing Is ThinkingIf you are not convinced, then may I offer this great essay by Paul Graham where he argues that the future world will be split between those who write and those who don't? The "writes" vs "write-nots" divide may not sound dangerous at first until you realize that it could lead to a world of "think" and "think-nots."
He uses a great analogy to explain this. In our modern world, the only people who are fit are those who choose to be that way — by exercising. That was not always the case. In the past, when most people did labor work in the farm, everyone was physically fit.
Similarly, in the future, he conjectures that the only people who are mentally sharp are going to be those who make a conscious effort to be that way. Writing is one of the ways that helps you exercise your brain. Writing is the gym for your mind. A mental resistance training, if you will. Each sentence strengthens the foundation and architecture of your thoughts.
Writing with AI
With LLMs, writing has become easier and more fun than ever! No, I am not suggesting we write a small paragraph and ask an LLM to convert that into a full-blown post. That'd be missing the point entirely! Instead, we should use LLMs to:
- - Brainstorm ideas,
- - Overcome writer’s block,
- - Improve grammar,
- - Provide a critique of a written piece, and
- - Copy-edit.
If you are still unconvinced, remember:
If you're thinking without writing, you only think you're thinking.

So we should all write, and write more! Not just for today’s readers, and certainly not just for our future AI overlords, but for our own clarity of thought.

Sources and links:

https://en.wikipedia.org/wiki/Low-background_steel

https://paulgraham.com/writes.html

https://www.nature.com/articles/s44222-025-00323-4

Related Posts

AI Needs More Data Science Thank You Might Think

8 Key Principles for Applied AI

The Coding Evolution of a Data Scientist: 20 Years, 4 Eras