I was scrolling through Hacker News on my phone the other evening when an interesting headline caught my attention. It was a technical piece about distributed systems, right up my alley. So I tapped the link and settled in to read, but couldn’t get past the 3rd paragraph.. There was nothing wrong with it per say.. but I just couldn’t shake off this feeling..

It’s hard to pinpoint exactly what felt off, but I knew immediately it was written by an AI.

After reading a few more paragraphs, I gave up and headed to the comments section. Now I know HN is generally very pro-AI, but I thought surely someone else would be just as annoyed by it as I was.. But nope, I had to scroll down so, so far before I found a single comment that pointed this out.

And it kind of bothered me. Are we really getting that used to synthetic text? Or are we just too tired to call it out?

It’s a Slippery Slope

Writing has never been effortless for me.

English is technically my first language, but growing up, my parents spoke so little of it at home that I had to spend part of elementary school in ESL (English as a Second Language) classes. Pair that with my awkwardness and inability to read social cues.. figuring out how to express my thoughts without sounding too blunt, out of place, robotic or tone deaf has taken years of deliberate effort.

This is one of the reasons as to why I started this blog. I wanted to practice writing, and coming up with ideas to write about. I’m hoping the process of putting down whatever unpolished thoughts that come to mind onto a page will slowly help me get better at self expression as a whole. It’ll also force me to be more intentional about what I want to communicate, instead of just rambling on or going on tangents.

Because writing takes real effort, AI writing tools are becoming an unimaginably alluring trap.

Take everyday work communications for example, drafting a email or message can take a surprising amount of effort. You write a draft, delete it, then re-write it, read it multiple times wondering if your tone comes across as too aggressive or if you completely missed some social nuance.

But now, all you need to do is paste your rough bullet points into an LLM. And POOF, an instant polished and socially acceptable response. No need to agonize over word choices or second guess your tone. No need to spend mental energy worrying if your reply will offend someone.

How could anyone resist that?

The issue is, I know consciously that writing gets better with practice, and social skills is like a muscle that needs exercise.. I know that I can only get better at reading people and expressing nuanced ideas by going through the process myself. The same goes for all types of learning. If I rely too heavily on AI, I am denying myself the valuable practice I need to improve, and letting my brain atrophy.

Why AI Prose Feels So Uncanny

How did I immediately know the article I was reading was AI generated?

It comes down to how LLMs predict text. When we ask an LLM to write an article, it doesn’t have an opinion on the subject, or any lived experiences. What it does have is a probability distribution over roughly one hundred thousand tokens in its vocabulary. For every word it generates, the model calculates the likelihood of what comes next based on patterns in its training data.

Most standard decoding setups push models toward high probability paths. They pick the safest, most agreeable phrases. In data science terms, AI generated text typically has very low perplexity, meaning the sequence of words is rarely surprising to a language model.

It also tends to suffer from low burstiness.

Human writing has some inherent rhythm to it. We mix things up naturally. Use slangs, odd metaphors, etc etc. AI writing, by comparison, can sound stiff. The sentences tend to hover around the same length, and paragraphs follow a similar pattern.

On top of that, models undergo Reinforcement Learning from Human Feedback (RLHF). Human evaluators consistently reward polite, structured, and inoffensive responses. All the personal quirks gets sanded down, so everything eventually end up sounding like a corporate memo.

(Now I know people have structured prompts to “de-AI” their writing, mix up sentences, use slang, etc etc. But the results still never sounds quite right..)

Am I Absorbing the Machine?

After reading the HN article, this question kept nagging at me.. How is my own writing being influenced by all the AI generated text I consume on a daily basis?

Our writing is shaped by what we read. If most of what I read all day (technical blogs, documentation, etc.) is now written by LLMs, those sentence structures inevitably creep into my own head.

Lately, when I re-read some of my own writing, I get this feeling of self doubt.. Was my writing style always like this? Do I usually choose these words, use these phrasings..? Or did I reach for them because my reading diet has conditioned me to expect that cadence?

It feels a bit like linguistic drift. When everyone consumes synthetic prose, our own writings begins to mimic this “average”. And when future models scrape the web, they train on synthetic text that humans have started imitating.. Like some sort of echo chamber of homogenized style..?

---

Not sure where I’m going with this, but I guess I just want to make a promise to myself to keep trying. My grammar might not be perfect, my phrasing might be awkward at times, and I might ramble on occassionally.. but I will try my best to make it sound like me. :)