Why LLM Writing is Easy to Spot

No, AI/LLM is not used to write this article. 

The way LLM works in Programming

Linus Torvalds once called it autocorrect on steroids. Well, that's what it kind-of actually is. The modals try to figure out what the next most probable word or sentence is. So it is hard for it to correct itself when noticing mistakes. Because correcting itself would mean that the sentence after the correction could become invalid. 

A solution to this problem is to have one or more “thinking” agents to work parallelly, which reduces the errors, but at the cost of tokens, time, and energy.

If you have ever tried to use it for writing programs, you will notice that LLM has the following tendencies:

1. Write to the end of the file when creating new functions without checking if there is a better place for it in the file
2. Ignoring the codebase structure when creating new modules
3. Creating duplicate code in larger codebases without exploring for redundancies
4. Assuming that the codebase use a widely supported framework (especially when it comes to CSS)
5. Ignoring the well detailed instructions you provided in the AGENT_NAME.md or SKILL files.
6. Not being up-to-date with the changes in tools/frameworks
7. A lot of unnecessary comments for obvious code.

Most of these can be solved by careful prompt engineering and not letting it write to file without review, but, that could quickly evolve into a frustrating experience if you are well versed in your field.

While programming is a mess, it is logical, it is easy for a developer to just read or test to see if it actually works. But blogs?

How it writes blogs

What I noticed is that unlike when writing code, the modals do not spin up more thinking agents, probably because unlike programming, there is no way for it to compile or test a sentence. Which makes it write about things that never happened as facts.

Another thing that humans do, is constantly edit their article until they get bored of it before publishing. For example, this sentence was added by me at the end just before publishing this article. 

A way for the modal to test and correct itself is by proofreading, it could be a separate agent, it will need to scour through the internet and other resources to fact-check its claims. But alas, the internet doesn't always tell the facts right, and can be very misleading.

Something that I have seen before the LLM era is that, when a publication or a politician wants to mislead the public, they say a lot of words that means nothing. And when they want people's attention, they hype up an issue with serious words.

People tends to “drift off” when something they hear or read is boring, and also focus too much on a specific thing that catches their eye. Both can be used together to make people ignore real problems and focus on made up ones. Sounds familiar?

If you have done any training on SEO (Search Engine Optimization), you may have also read about increasing the word count so that search can filter and find your article.

LLMs are trained on such internet, so many words it spits out are meaningless, and sometimes intentionally deceptive. So the reasons why you can usually, easily find whether an article is written by LLM or not are because of:

1. Meaningless word-salad
2. deceptive content
3. Inaccurate content
4. Sentences that don't connect very well.

Common patterns that are false flags

1. The em and en-dash, and ellipsis are something people often point out to claim an article is written by LLM, but those existed in articles before LLM; otherwise LLM wouldn't be able to use it. Also, it's very easy to set up in most systems.

Well, I got only one for now. Tell me if you have other false flag, I'll add it here.

Also on:

Interactions

No interactions yet.