Hacker News Reader

Comment by ikesau | Hacker News Reader

Beyond the OP's AI-written or AI assisted distinction, I'm also noticing people mimicking LLM's speech patterns. I've read blogs from people who I'm quite sure are above pasting AI output directly into their words who nevertheless are sounding more and more like AI as the sum of all their conversations with Claude begins to rub off on them (myself included, probably)

View on HN

I think this is just a sign that AI is now participating in the normal evolution of language over time. Language has always been about imitating... someone or some group comes up with a word/phrase/saying and uses it, other people hear/read it, and if they like it and/or find it useful, they incorporate it into their lexicon. This process is constant, and words and phrases are tweaked and morphed over the years as trends come and go.

Now, AI is participating in that process. It reads human words, and some of those words end up getting used more based on the algorithm, and then people read those words and copy some of them. This will feed back in to the AI as it ingests more content, and the feedback cycle is complete.

So what you're saying is the evolution of language is now being developed by the quirks of a particular floating point architecture?

I think that's kinda wonderful, actually.

I have noticed that sometimes in lists I have had the "The ... Solution: ..." sort of repetition. It is probably pre-existing but now that LLMs overuse it I actually am trying to adapt my speech patterns to not, because patterns LLMs overuse quickly become very grating to me.

Lately I have seem some Reddit posts with bullet points. I know it's a Power Point cliche, but that style almost never showed up on Reddit until about a year ago. I think a lot of those posts are AI-generated.

For a while I worked on defects for a web app at a large corporation. Users would submit walls of text, with a lot of unneeded details. A lot of people need help organizing their thoughts.

I suppose it stands to reason: LLMs were trained on human writing, and overuse certain tropes and patterns because those patterns are commonly represented in human writing. But many people aren't particularly adept writers, and they're going to turn to AI to either do their writing or inform how they write. The trope ends up reinforcing itself as people just start to think that AI output is just what normal writing looks like.

> overuse certain tropes and patterns because those patterns are commonly represented in human writing.

That's factually untrue. Give me a link to a pre-2020 piece of writing that sounds like an LLM.

The actual prose is ham fisted, but the structural bits - “why it matters” headings and bulleted content are traditional clickbait. Compare to “the results may surprise you” and similar stock phrases that we roll our eyes at. This stuff has always hooked people so it’s unsurprising that RLHF got its way.

If you want examples look at popular DIY magazines from a decade ago, they’re full of this sort of material where every article subheading has a catchy sentence.

There's something called RLHF.

> I'm also noticing people mimicking LLM's speech patterns.

I've noticed this happening to myself. I use Claude Code quite a bit at work, and Claude tends to favor certain phrases, like, "the smoking gun", that I didn't use very often before. After encountering phrases like that frequently, I've found myself incorporating some of them into my own speech.

It is the opposite for me. AI loves terms like "you've just hit upon ..." or "it's the holy grail" or "The X Trap" so much that I internally cringe when I see them, and would avoid using them.

“An honest caveat…”

A lot of AIisms in writing are annoying to me, but Claude has completely ruined the words "real", "honest', and "genuine" for me.

> Beyond the OP's AI-written or AI assisted distinction, I'm also noticing people mimicking LLM's speech patterns.

I wonder if long disfavored words like "ain't" might make a comeback as proofs of humanity.

If they do, then AI writing will incorporate them.

> If they do, then AI writing will incorporate them.

Eventually, but at least for something like "ain't" the hill to climb to get there will probably be a lot higher, because you've got decades if not centuries of explicit discouragement and very little presence in existing formal writing (and maybe online writing period).

The more durable proof of humanity will probably be hyper rapid slang evolution, but that's going to have the downside of social fragmentation and making the past less legible. If slang changes rapidly over the course of a year, will anyone still remember the slang from 10 years ago to understand an old post?

One thing I have yet to see on the mainstream AI platforms is a combination of foul language, dirty jokes, rule34 and racial slurs. 4chan-GPT and local LLM's excluded. Perhaps require Rule34 and foul language as proof. There's probably a way to jailbreak that for a little bit but it would get patched.

https://arxiv.org/abs/2409.01754

Skullface from MGS5 predicted this. Hideo Kojima sends his regards.

I'm feeling the same way. People naturally pick up writing habits from what they read, especially if they haven't really developed their own style yet. So it seems only natural that, after spending a lot of time reading AI generated text, some of those patterns start creeping into their own writing.

The less established someone's writing style or vocabulary is, the more susceptible they probably are to that influence. I don't necessarily see it as inherently good or bad. People have always picked up habits from the things they read. AI is just another source of influence, alongside books, blogs, journalists, or other writers. Some influences are great, and some are less so.

I caught myself saying "push back" the other day. I've never said it before, it's a Claude-ism.

That's a corporate phrase that predates the current llm stuff along with things like ping, circle back,table that

Is it? I'd never heard it until Claude (or maybe Gemini was where I first heard it from). Any idea of the time frame when it started being spoken?

At least 5-10 years. Like my whole career.

The llms usually remind me of how corporate people talk

Reach out to tayo42, and they can provide some color.

Learnings, solutioning, synergy

Ideation.

We no longer meet, we jump on a call.

damn forgot about ideation lmao

"what is your ask" instead of "what is your question."

I think we're going to have to take this ... offline

These are different to me. An ask is when you expect some action from the person. A question is for an answer.

Right, but I hadn't really heard it before until I started using AI daily.

This shows that a lot of neologisms that people think of as ‘AI speak’ are often older and common terms but from an unfamiliar domain (like the corporate world or academia).

Reply on news.ycombinator.com