Categories
38 Before 38

Can AI replace me as a writer?[38 Before 38]

I “ask” three llms to rewrite my work. My takeaway from the results.

two rows of 19 candle drawings each. Every candle is yellow, except the 9th one on the second row(28th in order) is white, and the last on the first row, the 3rd & 4th in the second row are red. The background is 3 robots in a row typing on laptops.

This blog is part of my 38 35 before 38 series. I write a blog for every single day for the 38 days leading up to my 38th Birthday.

I have written before about how AI will not take engineering and accounting jobs. I have also posted about how awful my workflow is and how it has hindered my ability to post daily. So I asked myself, can AI replace me? Or more appropriately, can it replace me as a writer. Generating text was one of the first tricks that Machine Learning algorithms pulled off. So why can’t it replace writers? Especially amateur ones like me.

To test this theory, I gave Google’s Gemini, Anthropic’s Claude, and OpenAI’s ChatGPT to rewrite two of my blog posts. One about a technical subject(how AI works) and one about a personal subject(how my workflow hinders my writing).

All of the got the same prompt in both cases:

Following is a blog post:

<text of the blog post, including title and subtitle/excerpt>

Rewrite it to make it better.

The prompt is vague on purpose. I only used the free versions of the chatbots as that is what most people are using. I did not want replication of a style, but what the training corpus of these models has calculated as a “good” style. Or at least better than mine. The results I got were not surprising, but quite revealing as to what the direction of the open web is going to be.

AI on AI

Here are the outputs of all three models for the AI explainer:

Gemini is the only one that “got the assignment”. The other two assumed that I was looking for an editor. The output, however, is horrendous. Gemini turned the post into exactly the lifeless SEO blogspam that Google thinks people want to read. The type of useless trash that you will get when you type “chatgpt in layman terms” in their search bar. The type of crap I have been railing against for years.

The editing I got from ChatGPT and Claude, while different in text, are similar in nature. Both seem to be concerned with the tone of my work. Both seem to prefer more “neutral” terms over the my “charged” and poetic language choices. For example, I titled section explaining how LLM companies gobble up data as The Great Maw. Claude renamed it to The Great Data Feast. Similarly, ChatGPT renames the section titled Caveat Emptor, to the rather pedestrian Important Disclaimers.

These LLMs have deemed any expression of a personality as an error. However, it would be arrogant of me to claim that their editing was not helpful at all. Especially Claude’s changes could be readapted to my style and improve the end result.

Are we Dancers?

The other post was more personal. It talked about procedures, but it was about my issues. Here are the results:

Once again, Gemini is the only one that truly rewrote the whole thing. And once again, in the most sanitized SEO friendly way. Search is Dead, and Google seem to want it that way. Claude and ChatGPT on the other hand just provide “helpful” editing.

Both cut down on the word count. ChatGPT pretty much strips it to the bare essentials. Like it was written by a primary schooler that forgot he had to submit an essay the next morning. It also takes out any semblance of a personality as well. Claude fares a lot better. It neuters the language as well. But again it provides some useful punch ups and formatting upgrades.

Convergence

I do have to note that I have never worked with a human editor. I don’t know if writers act like this when faced with changes by a human. The notes at the end provide clear reason for the changes. One common factor is all of them have a tendency towards a more positive tone. A lot of it is due to the controls these corporations have built into to these products. You can’t market your “Word predictor machine” if it spews sadboy emo poems by default.

But all of the outputs also reveal a glaring problem with all three offerings; their training data. The majority of their data is at best glorified copywriting and at worst sterile inoffensive blogspam.

Sapientia Ex Machina

This test sets up the LLMs to fail. I could have improved the initial prompt, and followed up with more prompts to improve the output. I could have purchased the premium subscriptions. Still the tendencies would not have changed. The end result still would not replicate me. It can’t replace me as a writer. They can generate a lot of blogspam for me so I can get to the top of some Google SERP. It can be half-decent editor for my work. But it can’t be me

To do that I would need to train an LLM on my own writings, which is impossible given scarce writings I have. I can use a feature called “Embeddings” in combination with any of these to create something like a “ChapraLLM”. But that is an experiment for another day.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.