The Way AI writes

[tp4tissue]

Does anyone really enjoy AI writing? or AI simulated human writing, famous authors, etc.

It's just so much smoother, and better.

Maybe not so much about the words, but how they all fit together, in a highly communicative way.


If only the Tp4 could write this gud'.

Does anyone truly find joy in the artifice of AI-driven prose? Or in the mimicry of human expression by those cherished literary figures?

It possesses a certain elegance, a fluidity that captivates the senses.

Perhaps it's less about the individual words themselves, and more about the way they interlace—creating a tapestry of thought that communicates with an almost ethereal clarity.

[tp4tissue]

You can see it's not a perfect transcription, but with very little editing, it's 99% there.  Most democrats would vote for Drump if he spoke like AI Transcriber, instead of his normal moron self.


Drump  >>> The military's job is to dominate any flow and annihilate any threat to America, anywhere, anytime, and anyplace. A big part of that job is to be respected again, and you are, as of right now, respected more than any army anywhere in the world. And when that's happening, I can tell you, you are respected like nobody can believe.

As President, I am assured your focus is on our core national interests. My preference will always be to make peace and seek partnership, even with countries where our differences may be profound. As you know, we're working on a lot of things right now.

When I left office four years ago, we had no wars, we had no problems, we had nothing but success. We had the most incredible economy, the single greatest economy for a president in history. I think we're repeating this guy by a lot growing up, but we had something going on very special. 


The Bolded statements are flagged by AI detector, so they're already using AI broadly.


AI rewrite>>> The task of our military is to assert its presence and eliminate any threats to our nation's integrity, unfailingly, in every corner of the globe. Integral to this mission is the restoration of our esteemed reputation, and at this moment, I stand before you to declare that you possess more respect than any other military force in existence. There is an undeniable aura of reverence about you, one that is scarcely fathomable to those less acquainted with your prowess.

As President, I am confident that your attention aligns with our fundamental national aspirations. My enduring hope is for tranquility and collaboration, even among those whose beliefs and practices diverge sharply from our own. In this very hour, we are engaged in numerous initiatives that seek to bridge those divides.

Reflecting on my departure from the office four years prior, I recall a time unmarred by conflict; a period defined by triumphs and an economy—the finest ever witnessed under a presidential banner. It feels as though we're embarking once more on that remarkable path, for we had cultivated something uniquely precious, a golden age that beckons us anew.

[Coreda]

It'd likely be possible to fine-tune some LLM after a particular writing style, in a way that wouldn't scream it's LLM generated but as it is most people are just using ChatGPT as it's the most well-known, with a minimal ask which produces output that isn't ideal.

Search results are now filled to the brim with clearly LLM generated slop. I say clearly as there are obvious signs of the use of LLMs and especially ChatGPT's 'house style' as some call it. Articles that begin with sentences that include emphatic phrases like 'the key aspect to understand' or 'is important to recognize' and such have sprung up everywhere in the last couple years and have become a go-to heuristic for recognizing it and bouncing. Along with ChatGPT's well-known liberal use of bullet points and repeated summaries that are all watered-down, non-succinct, padded with unimportant adjectives, and variously getting details incorrect.

Not sure if OpenAI's reasoning models at a default level are any better for prose but they're certainly miles better for anything documentation/code related than the non-reasoning models. I know Anthropic's reasoning models are also very well regarded and at a default have a different style than ChatGPT.


Edit: to clarify the reason heuristics for LLM generated content are useful (while there are still signs) is because it's a gauge of how much the author knows/cares about something (ie: is this even worth reading). SEO optimization and clickbait permeates the web and this has lead to very insincere content. Why should those looking to search/read/understand have their time wasted by others who aren't writing based on their own understanding/point of view but merely for ad revenue. The content becomes the least important aspect and is only there to deceive gullible users. It's an incredibly cynical use of such technology.

Just the other day there was an article about how a Hearst Newspapers author got an LLM to generate a book reading list. Only 1 of the 15 books in the list were real. It got published in mainstream newspapers via syndication. Exhibit A: the author does not care about what they're writing so why should anyone else.

OTOH there is a lot of use of LLM for second passes of writing so in that sense LLM-parsed writing is becoming commonplace even outside of the examples I take most issue with.

[Leslieann]

Watch a few of Louis Rossman's videos, he goes into how it's all super easy to spot a lot of the A.I. written stuff because it follows specific patterns and injects very specific punctuation.

It's funny, these websites can't spot it, but once you see what he spotted it's pretty easy to catch.

[tp4tissue]

Just the other day there was an article about how a Hearst Newspapers author got an LLM to generate a book reading list. Only 1 of the 15 books in the list were real. It got published in mainstream newspapers via syndication. Exhibit A: the author does not care about what they're writing so why should anyone else.

Here's the thing.  LLM writes better, in all mechanical ways, that is a fact.

Estimates range from 21-70% of the American population is functionally illiterate.   Something like 50% (mechanically, not literacy) read below a 6-th grade lvl, (12 yr old).

So,  compare an AI LLM to the Majority of people,  you can argue, softly, the AI's output is more worthwhile a "read" than what a 6th grader belches.


Human beings are LLMs, only for now, we're more integrated, with more sensors, and more context trained.

AI surpass humans in most mechanical metrics because they're a distillation of human thought, and procedures of thought.


Tp4 thinks, people tend to poo-poo AI-content primarily out of prejudice. A human can't do better-quality content in the majority of cases, and certainly not for the same cost and speed.


 ____ AI


Here's the essence of it, stripped down to its core. The marvel of a language model is evident; it crafts prose with a precision and clarity that seems to elude many human hands.

Take a moment to ponder the staggering figures: estimates suggest that a sobering 21 to 70 percent of the American populace grapples with functional illiteracy. Disturbingly, about half of them, assessed not by literacy standards but by sheer mechanical ability, engage with text at a level reminiscent of a sixth grader, a mere child of twelve.

When you place an AI language model side-by-side with the vast majority, one might gently contend that the former offers a more engaging intricacy than the hasty utterances of a youthful mind.

As humans, we, too, possess this remarkable capacity to communicate; however, we are entwined within our myriad experiences, sensing the world through an intricate web of emotions and contexts.

Artificial intelligence often outstrips us in what might seem the more sterile realms of mechanical proficiency, for it distills the essence of human thought and the very processes that govern it.

In the realm of critical thought, it seems some hold a bias against AI-generated content, perhaps from a deep-rooted prejudice. Yet, it is an undeniable truth that, more often than not, a human cannot match the caliber of output that these models produce—certainly not at the same pace or with comparable cost.

[tp4tissue]

Watch a few of Louis Rossman's videos, he goes into how it's all super easy to spot a lot of the A.I. written stuff because it follows specific patterns and injects very specific punctuation.

It's funny, these websites can't spot it, but once you see what he spotted it's pretty easy to catch.

You can argue that in the other direction,  Humans are doing it wrong, and we aught to be using those punctuations and AI-writing styles.

For example,  AI playing starcraft, or moba,  gamers have adopted those AI discovered strategies because they're clearly better.

Why isn't the same lens applied to the "creative" fields.  Why is there no objective measure for creativity.

___ AI

One might contend that conversely, humanity is erring, and we ought to embrace those punctuations and the stylistic nuances of AI writing.

To illustrate, consider AI engaging in Starcraft or MOBA; players have embraced the strategies unveiled by AI, as they are evidently superior.

Why, then, is this same perspective not extended to the realms of creativity? Why is there an absence of an objective standard for measuring creativity?
 

[Coreda]

Here's the thing.  LLM writes better, in all mechanical ways, that is a fact.  Estimates range from 21-70% of the American population is functionally illiterate.   Something like 50% (mechanically, not literacy) read below a 6-th grade lvl, (12 yr old). Tp4 thinks, people tend to poo-poo AI-content primarily out of prejudice. A human can't do better-quality content in the majority of cases, and certainly not for the same cost and speed.

The point in the clarifying edit wasn't about being prejudiced against use of LLMs per se.

However, just addressing this point you raise, of course if we're comparing literacy of sixth graders and the cost/speed efficiency of LLMs then obviously they're extremely capable, though at least as a native English speaker I haven't read any LLM written examples possessing an elegance and 'fluidity that captivates the senses', even when instructed to mimic other authors.

Is that a future possibility? Perhaps. There's a constant flow of improvements with neural network research.

To my actual original point though, the critique was that the internet, which already had a problem with insincerity and misinformation due to the the wrong things being incentivized (likes/monetary gain/warped discussions for sensational/fake/insincere content via average people), is now experiencing an even greater wave of insincerity from the use of LLMs as a substitute for people's knowledge/points of view.

Look at LLM use practically. It's strongly correlated with lazy writing and insincerity, regardless of the words used. There's no disclosure either. It's possible that book list I referenced was written decently. Does it matter? No because it was a waste of time and money for everyone involved beside the writer. Could that same author have used an LLM suitably in ways to only improve something they wrote? Yes but they didn't and neither to a lot of people using them.

Thus if there are any identifiers of LLM use that stand out it's a sign the author doesn't care about what they're writing and is a heuristic that neither should you, unless you want to second-guess everything you're reading and spend further time doing your own fact checking about even the lowest stake things which are meant to be the task of the human claiming to have domain knowledge about the subject they're writing about.

This all contributes to the noise that has to be to sifted through by everyone, which as a side-effect is adding to the reduction of attention spans globally.

Given this context, it matters less to me if someone with lower writing proficiency writes straightforwardly or with grammatical errors if it's their actual knowledge/experience/point of view. Or just use LLMs smartly as some actually do, where it's transparent to what the author themselves is communicating and not outsourcing.

[tp4tissue]

To my actual original point though, the critique was that the internet, which already had a problem with insincerity and misinformation due to the the wrong things being incentivized (likes/monetary gain/warped discussions for sensational/fake/insincere content via average people), is now experiencing an even greater wave of insincerity from the use of LLMs as a substitute for people's knowledge/points of view.

I can understand why insincerity and misinformation  is in your "bad things" bucket.

But that is how the core market is constructed.  What product isn't ultimately transforming -Energy- gas/coal/diesel into -Microliters- of Dopamine.    This market can ONLY be insincere, because it has an objective detached from reality, it's extracted a "Thing" a "sensory input" that no longer has any relevance to survival odds.

With the current human biological hardware, we do not recognize anything other than sensational-happiness as desirable, and our market punishes anything outside of the conversion, or anything that may be a reduction in conversion efficiency.


That is to say, everything we produce, sincere or not, serves the INDUSTRY of HIJACKED dopamine pathways. 


Intellectual-Bookish-Luddism might be "the counter" you're looking for,   but you can't SELL THAT, because it's anti-Capitalist.


There's only 1 outcome to our economy as is, the death of all things. With or without LLM.

LLM, assuming AGI is NOT-achieved,  is just a faster furnace that burns what's left of the world more quickly.

[Leslieann]

You can argue that in the other direction,  Humans are doing it wrong, and we aught to be using those punctuations and AI-writing styles.

Considering A.I. seems to have the brain of a 10 year old, no.