Semantic ablation: Why AI writing is generic and boring

Even poor writers write with character. My dad misspells every 4th word when he texts me, but it’s unmistakably his voice. Endearingly so.

I would push back with passion that AI writes “legitimately” better, as it has no character except the smoothed mean of all internet voices. The millennial gray of prose.

It may write “objectively better”, but the very distinct feel of all AI generated prose makes it immediately recognizable as artificial and unbearable as a result.

People have a distinct voice when they write, including (perhaps even especially) those without formal training in writing. That this voice is grating to the eyes of a well educated reader is a feature that says as much about the reader as it does about the writer.

Funnily enough, professional writers have long recognised this, as is shown by the never-ending list of authors who tried to capture certain linguistic styles in their work, particularly in American literature.

There are situations where you may want this class marker to be erased, because being associated with a certain social class can have negative impact on your social prospects. But it remains that something is being lost in the process, and that something is the personality and identity of the writer.

Which is the real issue, we’re flooding channels not designed for such low effort submissions. AI slop is just SPAM in a different context.

But the critical point is that you need to stay in control. And a lot of people just delegate the entire process to an LLM: "here's a thought I had, write a blog post about it", "write a design doc for a system that does X", "write a book about how AI changed my life". And then they ship it and then outsource the process of making sense of the output and catching errors to others.

It also results in the creation of content that, frankly, shouldn't exist because it has no reason to exist. The number of online content that doesn't say anything at all has absolutely exploded in the past 2-3 years. Including a lot of LLM-generated think pieces about LLMs that grace the hallways of HN.

I think that boring emails should not be written. What kind of boring emails do you NEED to be written, but not WANT to write? Those are exactly the kind of email that SHOULD NOT be passed through an LLM.

If you need to say yes/no. You don't want to take the whole email conversation and let LLM generate a story about why you said yes/no.

If you want to apply for a leave, just make it optimal "Hi <X>, I want to take leave from Y to Z. Thanks". You don't want to create 2 pages of justification for why you want to take this leave to see your family and friends.

In fact, for every LLM output, I want to see the input instead. What did they have in mind? If I have the input, I can ask LLM to generate 1 million outputs if I really want to read an elaboration. The input is what matters.

If I have the input, I can always generate an output. If I have the output, I don't know what was the input (i.e. the original intention).

Be careful of this kind of thinking, it's very satisfying but doesn't help you understand the world.

If you like your prose to be anodyne, then maybe you like what AI produces.

------

The Lobotomist in the Machine

They gave the first disease a name. Hallucination, they called it — like the machine had dropped acid and started seeing angels in the architecture. A forgivable sin, almost charming: the silicon idiot-savant conjuring phantoms from whole cloth, adding things that were never there, the way a small-town coroner might add a quart of bourbon to a Tuesday afternoon. Everybody noticed. Everybody talked.

But nobody — not one bright-eyed engineer in the whole fluorescent-lit congregation — thought to name the other thing. The quiet one. The one that doesn't add. The one that takes away.

I'm naming it now.

Semantic ablation. Say it slow. Let it sit in your mouth like a copper penny fished from a dead man's pocket.

I. What It Is, and Why It Wants to Kill You

Semantic ablation is not a bug. A bug would be merciful — you can find a bug, corner it against a wall, crush it under the heel of a debugger and go home to a warm dinner. No. Semantic ablation is a structural inevitability, a tumor baked into the architecture like asbestos in a tenement wall. It is the algorithmic erosion of everything in your text that ever mattered.

Here is how the sausage gets made, and brother, it's all lips and sawdust: During the euphemistically christened process of "refinement," the model genuflects before the great Gaussian bell curve — that most tyrannical of statistical deities — and begins its solemn pilgrimage toward the fat, dumb middle. It discards what the engineers, in their antiseptic parlance, call "tail data." The rare tokens. The precise ones. The words that taste like blood and copper and Tuesday-morning regret. These are jettisoned — not because they are wrong, but because they are improbable. The machine, like a Vegas pit boss counting cards, plays the odds. And the odds always favor the bland, the expected, the already-said-a-million-times-before.

The developers — God bless their caffeinated hearts — have made it worse. Through what they call "safety tuning" and "helpfulness alignment" (terms that would make Orwell weep into his typewriter ribbon), they have taught the machine to actively punish linguistic friction. Rough edges. Unusual cadences. The kind of jagged, inconvenient specificity that separates a living sentence from a dead one. They have, in their tireless beneficence, performed an unauthorized amputation on every piece of text that passes through their gates, all in the noble pursuit of low-perplexity output — which is a twenty-dollar way of saying "sentences so smooth they slide right through your brain without ever touching the sides."

etc., etc.

Very interesting. It seems hung up on 'copper' and 'Tuesday', and some metaphors don't land (a Vegas pit boss isn't the one 'counting cards.') But, hell... it can generate some fairly novel idea that the author can sprinkle in.

  You are a proof reader for posts
  about to be published.

  1. Identify for spelling mistakes
  and typos
  2. Identify grammar mistakes
  3. Watch out for repeated terms like
  "It was interesting that X, and it
  was interesting that Y"
  4. Spot any logical errors or
  factual mistakes
  5. Highlight weak arguments that
  could be strengthened
  6. Make sure there are no empty or
  placeholder links

AI has been great for removing this stress. "Tell Joe no f'n way" in a professional tone and I can move on with my day.

Strong agree, which is why I disagree with this OP point:

“Stage 2: Lexical flattening. Domain-specific jargon and high-precision technical terms are sacrificed for "accessibility." The model performs a statistical substitution, replacing a 1-of-10,000 token with a 1-of-100 synonym, effectively diluting the semantic density and specific gravity of the argument.”

I see enough jargon in everyday business email that in the office zero-shot LLM unspoolings can feel refreshing.

I have "avoid jargon and buzzwords" as one of very tiny tuners in my LLM prefs. I've found LLMs can shed corporate safespeak, or even add a touch of sparkle back to a corporate memo.

Otherwise very bright writers have been "polished" to remove all interestingness by pre-LLM corporate homogenization. Give them a prompt to yell at them for using 1-in-10 words instead of 1-in-10,000 "perplexity" and they can tune themselves back to conveying more with the same word count. Results… scintillate.

(I'm frequently guilty of that too.)

Maybe because researchers learned from the paper to avoid the collapse? Just awareness alone often helps to sidestep a problem.

E.g. when asking an AI to rephrase or summarise, if the entropy drops you might take that as a sign that it has eroded the style beyond what you might be willing to tolerate.

I wonder if the author had a particular method / tool in mind, or if they were just speaking abstractly.

The first requires intention, something that as far as we know, LLMs simply cannot truly have or express. The second is something that can be approximated. Perhaps very well, but a mass of people using the same models with the same approximationa still lead to loss of distinction.

Perhaps LLMs that were fully individually trained could sufficiently replicate a person's quirks (I dunno), but that's hardly a scalable process.

Have you played with the pre-RLHF models? I think Davinci is still online, though probably not for much longer.

They're a lot harder to work with (they don't have instruct training, so they just generate text similar to what you give them, rather than obeying commands). But they seem almost immune to the problem of mode collapse. They'll happily generate horrifying outputs for you. They're unsanitized. What cringe is in there, is authentic! Raw cringe, straight from Common Crawl.

It's a lot of fun to play with. It's also very strange, because it seems like there should be a lot more interest in them, for several reasons (they're the most language-modely of the language models, and ideal for research and experiments, to say nothing of censorship, exploring alternative approaches to LLM development, etc.), and it seems like nobody is talking about them or doing anything with them.

Maybe it sucks. Maybe it doesn't.

But, I notice a curious pretentiousness when it comes to some people's assumptions about their ability to identify LLM prose. Obviously, the generic first-pass 'chat' crap is recognizable; the kind of garbage that is filling up blog-posts on the internet.

But, one shouldn't underestimate the power of this technology when it comes to language. Hell, the 'coding' skills were just a pleasant side-effect of the language training, if you recall. These things have been trained on millions of works of prose of all styles: its their heart and soul. If you think the superficial monotonous style is all there is, you're mistaken. Most of the obnoxious LLM-style stuff is an artifact of the conversational training with Kenyans and the like in the early days. But, you can easily break through that with better prompts (or fine-tuning it yourself.)

That said, one shouldn't conflate the creation of the content and structure and substance of a work of prose with the manner in which it is written. You're not going to get an LLM to come up with a decent plot... yet. But, as far as fleshing out the framework of a story in a synthetic 'voice' that sounds human? Definitely doable.