Writing with AI: What's the Best Approach?

To me, AI writing always results in this:

They always look nice at first sight, but over time, they look all the same: complex, unnatural, and busy. Same as DALL-E images.

And if it’s read by a human, it should also be written by a human. Nothing worse than having to read through a stack of elaborated text that has been generated on a small prompt. Naval elaborated it best at the The Differences between Writing and Coding

Based on the The Differences between Writing and Coding that Naval Ravikant shared, writing is consumed by human beings, and they discuss, that I’d rather Read the Prompt, why would I need to read your fleshed-out article by AI, and not the lean prompt with all the insights? That’s what Naval asked Garry Tan, and they didn’t have an answer.

It’s like the below meme, AI making you Exhausted both on the creation time, as it never ends, people have more work than ever, but also on the reading side, as you need to read everyone’s bloated AI-written stuff, rather than just the prompt.

Same as we do in programming: sharing AGENTS.md or Agentic Skills so others can use that insight too, rather than sharing 10 different solutions. Obviously, as explained in the difference between writing in programming, it’s not the same as a computer consumes programming, and if you solve a problem, why not generate the app, the solution, and share that? We can’t do that in writing.

What if we have two kinds of writing:

1. The elaborated but bloated (AI-enriched) one - like what Garry Tan is doing on Twitter or articles as he says
2. Or the raw prompt, with all the ideas and insights lean and easy to read (not the solutions, maybe?). Very much like the philosophy of I’d rather Read the Prompt.

And as I argue in Manual Writing and Will AI replace Humans, the second point has character, a soul and human behind, no AI model will ever have.

I was wondering if AI should suggest to me writings - a question asked by Euwyn Goh . It might for brainstorming. I don’t want AI to write my sentences, but it could help me brainstorm.

Usually, the way I work is: first, I do brainstorming, maybe some research on what happened in history. In that phase, AI might help. But also, most often when I ask what I should add to my articles, AI says to add “real-world examples”. So, most of the time it’s not that creative. And the worst part, after using AI, I usually haven’t done my work, and I’m more exhausted than when I just write it 🤐.

What I used it for just recently, though, was to go through my notes, list MY notes, and add them to the chapter of my book (all in markdown, which’s why it works well). What’s nice is that you provide a way to train my own data for that AI using a local model. At least that’s what I’m trying a little at obsidian-note-taking-assistant with local first Second Brain RAG for any Obsidian Vault (works with any Markdown). Works with DuckDB. Also includes a web-app, see the blog article at Building an Obsidian RAG with DuckDB. Or more examples at Second Brain Assistant with Obsidian (NoteGPT).

I’m all in for Writing Manually and not replacing my thinking with AI, especially if a human being consumes it.

I like Naval’s take, as elaborated above, which says it all depends on who the customer is. E.g., writings are for humans to read mostly, so they should be written by a human. Code is mostly a tool to achieve something, read by a computer, so it’s less relevant to write by hand, unless you **work in a team or need strong maintainability.

Also, I pondered on the question of Should I Change My Writing because of AI?

• Byword
• ChatGPT
• LEX by Dan Shipper
• Claude AI with Claude 3.5 being much better than ChatGPT 4.0
• Spiral by Dan Shipper and Every (Writing Tool) ^114b88
• River Docs: The AI Document Editor
• Wordsmith by Euwyn Goh

See also Use Cases of AI Writing (ChatGPT, Lex, etc.) and Creative Prompt Engineering.

# Pangram Report per 2026-07-09

AI Content Is Everywhere on Social Media, Especially LinkedIn | Pangram Labs - obviously dedecting AI content is also not reliable, again, AI is not deterministic, it might even say once it’s AI written and once not if you asked it twice, but here are some reports to ponder from that article:

And LinkedIn and Medium being top, is what I saw as well, no test needed. Substack is very low makes sense too, as there’s no API or no Social Media Scheduling Tools and Platforms yet to publish notes or substacks newsletters, plus it’s a platform for actual writers.

Here are two more overviews:


Source

Reddit replies are rarely AI written - though I know that is also depenend on the subreddit for sure:

Source

How the data was collected

From the article:
“Collectively, since the Chrome Extension’s launch on April 24th 2026, users who opted into sharing their data for research helped us create a dataset of 1,002,627 posts across several of the largest social media platforms on the internet: LinkedIn, Medium, Substack, X/Twitter, and Reddit.

Each post in our dataset is counted only once, and we only scan items that are longer than 50 words. Every post was analyzed with Pangram 3.3, our latest AI detection model, which achieves a 0.01% false positive rate. This dataset allows us a direct window into what AI-generated content people are seeing on their feeds at this point in time.”

More on How to detect AI Writing.

• Course with Dan Shipper: AI writing gets a bad rap.
• Plagiarism: Check if AI Written: Tools to Check Plagiarism and AI-Generated.
• Counter Arguments to write with AI:
  • Technical writing is too important to leave to language models
  • Will AI replace Humans
• Don’t be a meat proxy
• Opinion | I’m Begging You: Never Write With A.I. - The New York Times

Origin: AOMW Course- The Art of Modern Writing (Foster.co), AI Writing with Dan Shipper - YouTube
References: ChatGPT, Creative Prompt Engineering , My Services and Target Audience (SSP Data),Brainstorming Ideas for Earning Money - Book, SSP Data