My guess is years of RLHF optimized for rater's first impressions of how well-written something was. It uses lists! It has headings! It uses fun phrases like 'smoking gun'! It's long and informative! These things are relatively fast and easy to identify and select as a preference.

But none of this really covers whether the response is _useful for the audience_. If you're not the audience, it can be hard to tell what is and isn't useful.

[0] An example would be something like:

"The user presses a button, and the form submits."

There are valid intransitive uses of "submit," but none that make sense in this context. In this case "submit" either requires an object (like "the form submits data"--plus maybe an indirect object, like "the form submits data to the server") or should be swapped for "the form is submitted" if you absolutely don't want to supply an object.

>We take these results as a proof of concept that a length penalty used during reinforcement learning with GRPO can directly affect grammatical features. We note that the model was not directly instructed to use compressed or dense language, and was not directly given information about the availability of syntactic alternations. Rather, the penalty resulted in an emergent preference or shorter members of syntactic alternations. [...] We speculate that the use of verifiable rewards creates an environment where correctness is decoupled from humanlike communication, leading models to converge toward a novel, potentially less comprehensible distribution of language use.

This sounds like common sense to anyone who tinkered with the models and training, however someone still needs to quantify it so it's good they did. I also want to note that this is only visible in Claude's default register, it's still able to write differently when prompted.

LLM-generated text is hard to understand because while it's perfectly good language that follows valid cognitive and informational patterns, it does not reflect the intent or mental model of another mind.

If we had talking centipedes the results would probably be similar, but for different reasons.

I've been iterating on this writing clarity prompt. The term we came up with was "Asyndetic parataxis with referential ellipsis and repeated rhetorical closure." It doesn't fully describe the exact issue I have with modern AI writing, but it's close.

When you try that with output from an LLM, it doesn't fit because it doesn't come from a mind. (Or at least not good enough to sustain the illusion.) It's subtly but persistently wrong, and all those little warning flags are tiring.

It's like some absurd event where a plastic animal is put out and real ones try to mate with it, because it meets just enough of the instinctual requirements.

Thus it should be possible to fine tune models to produce more human readable text in assistant role while keeping information density in thinking role. Or maybe not. That is interesting to test.

https://role-confusion.github.io/

Prob same reason they are bad at novel system (meaning not copying an existing system or cargo culting patterns) design and organization. No concept of what it's like to deal with bad design.

When communicating verbally, through writing, or through design it's experience and cognitive empathy that leads to "good" communication.

(e.g., "baseline system state transition latency validation") instead of unpacking them with prepositions ("validating the latency of state transitions in the baseline system").

For me, I can't read books at all, because I forget what was said two paragraphs before. I bet people that are good at skimming text have less problems with AI generated text.

Or reading patents.

Terrible ... like someone put in effort to make it unreadable.

AI-generated text is increasingly optimized for other goals like math, code, and reasoning

Yes, this is a problem.

AI-generated text is used internally by the AI to instruct agents, etc. And for feedback loops in reasoning. If you optimize for the result only, the AI will increasingly speak "AI", not human.

Plus using all this made up jargon and the unclear references to everything else (im referring to like the 'it' "XYZ is implemented using ABC. 123 referrs to it" you get in claudeslop) is very hard to understand but also kind of impossible to be wrong no? If it said "123 refers to ABC" but meant XYZ instead of ABC it's SOL but instead if it says "it" future tokens can ramble themselves out of being wrong thanks to the annoying slop ambiguity.

the problem with those hypothesis is that models don't in fact all talk the same (unless they're distilling claude). however if I unburden myself from having to be correct and allow myself to make up some more bullshit - what if "its not x its y" helps compute the right answer. Attending to that 'not x' could help point in the right direction in whatever dimension in latent space it needs to find the right y — and what if emdashes are proletariat filler tokens.

probably tho they're just RLing them to hell and back on code stuff and not paying attention to it and none of that is real lol. IIRC they rled opus 5.5 on talking more normal and it does and is also better than opus 5 so like with most things here, just another product decision on where to burn GPUs

Usually I’m getting better results by telling it to be more concise.

[0] https://github.com/cursor/plugins/blob/main/pstack/skills/un... [1] https://github.com/cursor/plugins/blob/main/pstack/skills/te...

I seems to remember a distinct time during which Claude for example became incomprehensible and weird with its writing, whereas before it was reasonably OK, even if still not that good

Em dashes are perfectly acceptable — nay, desirable — linguistic constructs, however, and I'll stand with Claude and die on this hill.

So there's the usual structure of "it's not x it's actually y" signs.

People who are knowledgeable in a particular obscure area tend to do the same thing.

The text is AI models produce is incredibly dense and the way it structures sentences just makes long form reading excessively hard.

I basically wanted something that could expand copy into components of "meaning" and then dump the output back out in more normal language.

What I came up with is still "AI generated" but the readability is something passable.

Will definitely be taking a look through this and the recommendations.

It’s almost surprising that LLMs generate sense quite as much as they do, given how they work.

It uses a lot of metaphors. It references stuff implicitly - a LOT - and assumes pretty much all the time that the thing it did is what you saw, which you didn't. But if you just.. read the stuff it says, and know a few old-school sayings and thesaurusize your reading style, it usually says things quite meaningfully.

I think a lot of people just don't read what it writes. So many folks I've watched used LLMs don't write in complete sentences, don't read the sentences that come back, don't watch the turns and tool calls in the middle, and then just respond to the message as if it says something it doesn't. I'm saying all this before I read the paper, but I've done a lot of personal and team experiments with this, so I'm quite curious to see what they say. I'll edit this after I go through it.

--

Fascinating! I did actually run into a lot of this with my personal 'engineering and communication styles' experiments. Most of the papers out there talk about models moving toward human language inside of a context window, and not the other way around (much smaller effect for human->model). I had a few years of conversations I ran on local and 'frontier' Claude and found that, yes, in general, information density is what each response turned out to be geared toward, whether it was everyday health discussions or deeper engineering topics and research. The component tests they ran I've been exposed to already, but I think their conclusion nailed it: ```This discrepancy may help explain why many frontier LLMs use linguistic constructions which are technically allowed by the grammar of English, yet are out-of- distribution for human language and are associated with comprehension difficulty.```

The best approach is just to write the outline and initial text in a very messy fashion yourself. And then have it clean it up and add detail. And then edit again by hand.

A third team mate uses a voice dictation tool to get down the basic ideas and then massages it with AI, but these two guys rib each other about it.

Another my other team mates (team mate 4 in this story) sheepishly turns around and asks if people can tell when he does it? (It's EXTREMELY OBVIOUS that he isn't writing anything) I confirmed it, in a friendly way. He defensively said he tries to get the AI to use his own voice.

I casually mentioned that "writing is thinking" or something like that and how I would prefer to think the thoughts. He didn't say much.

I hate AI writing. So much.

I think the difficulty in comprehension was mostly an Anthropic problem… partly because Anthropic employees stopped reading Opus output around the time Fable came out and stopped noticing RLVR going off the rails with compression.

The paper was a good read, in any case. Maybe I gotta start sending this article back to LLMs when they start compressing text too much :)

But I also wonder why AI loves to add slogans, catch phrases, and witty sentence fragments. It’s grating and a dead giveaway of slop.