Why AI writing is so generic, boring, and dangerous: Semantic ablation

Just as the community adopted the term "hallucination" to describe additive errors, we must now codify its far more insidious counterpart: semantic ablation. Semantic ablation is the algorithmic erosion of high-entropy information. Technically, it is not a "bug" but a structural byproduct of greedy decoding and RLHF ( …

[LBJsPNS]

Feed your AI Finnegans Wake

Lots of fun!

[Eclectic Man]

Re: Feed your AI Finnegans Wake

According to a study at Cardiff University, it is possible to track the progression of the late Sir Terry Pratchett's Alzheimer's disease from the writing in his novels:

"Subtle changes in Sir Terry Pratchett’s use of language in his books anticipated his dementia diagnosis by almost ten years, research has shown.

The study, from academics at Cardiff University, Loughborough University and the University of Oxford, used computer software to analyse the range of nouns and adjectives used in 33 of his best-selling Discworld novels."

https://www.cardiff.ac.uk/news/view/3020807-sir-terry-pratchetts-novels-may-have-held-clues-to-his-dementia-a-decade-before-diagnosis,-research-suggests

I was amazed at the avalanche of ideas in 'The Colour of Magic' and 'The Light Fantastic', then they settled down a bit, with somewhat more philosophical plots, especially 'Small Gods' and 'Jingo'.

Maybe in the future there will be an AI that checks to see if there is enough of the 'high entropy' stuff in its writing in future, and we'll complain that it is meaningless drivel.

[Richard 12]

Re: Feed your AI Finnegans Wake

Sadly the paper does not support the headline.

From the paper:

Plain-text versions of the 33 eligible novels were uploaded to SketchEngine.

While they had a linguist take a look, there's no literary academic input - which is surely necessary to support such a significant claim. Unfortunately it reads more like a PR exercise rather than a scientific paper.

[Anonymous Coward]

Re: Feed your AI Finnegans Wake

Author #1, Dr Melody Pattison, BEd (Melb), GDipArts (Melb), MAppLing (Monash), PhD (Ebor), FHEA, is 'Director of Studies in Language and linguistics' at Cardiff yes.

[Geoff Campbell]

Re: Feed your AI Finnegans Wake

My gut reaction is that the same result could be explained by an established, older author settling into a comfortable cranking-the-handle writing style after a younger, more florid phase of artistic creation.

But that's not to say the authors of this paper are in any way wrong, it's just a possible alternative.

GJC

[Benegesserict Cumbersomberbatch]

Re: Feed your AI Finnegans Wake

Sir Pterry never settled -- whilst the style of some of his later gems like Going Postal, Night Watch or Thief of Time is light years away from that of The Colour of Magic, they are not what you would call close to each other in style, especially structurally, despite being close to each other in time.

[Geoff Campbell]

Re: Feed your AI Finnegans Wake

I am sorry to have offended you.

As an avid reader of his, I absolutely disagree with you, but that's fine, all opinions are valid at this level.

GJC

[Not Yb]

Re: Feed your AI Finnegans Wake

It's always easier to come up with a prediction when the result is already known.

[MachDiamond]

Re: Feed your AI Finnegans Wake

"I was amazed at the avalanche of ideas in 'The Colour of Magic' and 'The Light Fantastic', then they settled down a bit, with somewhat more philosophical plots, especially 'Small Gods' and 'Jingo'."

Douglas Adams commented that his books ate through his ideas at a ferocious rate and it took time to refill his files. Authors that want to be authors full time and are pressured by publishers to produce can't use up everything they have on one or two books. I really enjoy Sir Terry's and Douglas's turn of phrase and subtle double meanings as much as all the concepts.

I'm not going to accept that AI is diagnosing Alzheimer's disease earlier than the medicos based on writings. There's too many variables that aren't accounted for.

[Joe Gurman]

Re: Feed your AI Finnegans Wake

….or anything by e e cummings.

[Doctor Syntax]

Re: Feed your AI Finnegans Wake

Or Stanley Unwin.

[Anonymous Coward]

AI;DR

"AI;DR" = “AI, didn’t read.”

"Why should I bother to read something someone else couldn't be bothered to write?"

https://futurism.com/artificial-intelligence/aidr-meaning

[Androgynous Cupboard]

Re: AI;DR

Oh, that is fucking brilliant. I'll be using that - thank you very much.

[ICL1900-G3]

Re: AI;DR

Agreed. Writers would do better to learn how to spell and how grammar works. Even in 'quality' journals, both spelling and grammar often fall short.

[Geoff Campbell]

Re: AI;DR

I can forgive subject matter experts not being experts in writing, I think. So long as the meaning is clear and readable, and the ideas presented are useful.

GJC

[Martin an gof]

Re: AI;DR

That is what sub editors are for. It's a shame most publications seem to have got rid of them. There are mistakes on the BBC website now that would have been unthinkable ten or fifteen years ago.

M.

[vtcodger]

Re: AI;DR

What popped into my mind about half way through the article was a line from The Incredibles "They keep inventing new ways to celebrate mediocrity."

What, exactly, is the point of all this effort to produce products that are pretty much Meh ... and apparently can not, by design, improve?

[Eclectic Man]

Re: Feed your AI Finnegans Wake

I feel like I may be metaphorically 'playing with fire' here, but I cannot help wondering what would be produced by an LLM 'trained' on the speeches of our great 'Leader of the Free World' Mr President Donald John Trump himself.

I'll get my coat, it has Monty Python's Hungarian phrase book in the pocket. https://www.youtube.com/watch?v=grA5XmBRC6g

[Aaiieeee]

There is something deeply unsatisfying about reading words that nobody wrote

It feels soulless, and I suppose this is why. Even AI-generated meeting notes just feel wrong and I think partly because there is no personality, no editorial flare that a real person would have added just by being them.

AI is very good at corporate speak though. Makes sense really.

[Neil Barnes]

Re: There is something deeply unsatisfying about reading words that nobody wrote

If noone could be bothered to write the words, why should I be bothered to read them?

[Doctor Syntax]

Re: There is something deeply unsatisfying about reading words that nobody wrote

A few days ago Liam was describing watching a Red Hat employee having a chatbot summarise a received email and then write a reply. The original sender could have cut out the entire email by having his own chatbot summarise it and write the reply without even sending it.

[Eclectic Man]

Re: There is something deeply unsatisfying about reading words that nobody wrote

Serious issue: Doctors are using AI to summarise their discussions with patients, does this 'semantic ablation' affect the quality of care or have they found some way to prevent it?

[Anonymous Coward]

Re: There is something deeply unsatisfying about reading words that nobody wrote

It makes them dopey.

[mjflory]

Re: There is something deeply unsatisfying about reading words that nobody wrote

Medically, this is such an important question! The medical journals (JAMA, etc.) have had many articles examining the use of LLMs in generating clinicians' notes, recording which take a considerable amount of their increasingly limited time. For better or worse, LLMs like BioBert that have been pre-trained on clinical language are coming into use. But evaluating the effectiveness of the methods is not simple.

[Jason Bloomberg]

Re: There is something deeply unsatisfying about reading words that nobody wrote

That "feels wrong" is what unsettles me. Presented with AI output from those who are deciding to use it, when asked "what do you think?" I find it covers all the bases but it's just not right though I cannot put my finger on why. I find myself stuck between accepting it as correct but finding it unsuitable, somehow "wrong", but unable to explain why. It's a very odd feeling.

And I find that with all AI output of any decent length. It's like that uncanny ability of TV detectives to tell who the murderer is just by looking at them. I presume my brain is wired in a way which is adverse to AI and it stands out like a sore thumb, my subconscious rejecting what my consciousness sees as passable, and more so than colleagues who see the same AI output as perfectly fine.

So maybe the article explains why I feel how I do. I'll certainly be testing that theory next time I am presented with AI output.

[Anonymous Coward]

Re: There is something deeply unsatisfying about reading words that nobody wrote

Uncanny valley 2.0?

[Anonymous Coward]

Re: There is something deeply unsatisfying about reading words that nobody wrote

Death in ParadAIse ?

[Snake]

Re: feels wrong

It's like Scotty when the Enterprise is thrown across the galaxy in "That Which Survives" - it feels wrong and you simply know it by both experience and an innate knowledge.

In the case of AI stories it's always the same feel: stupidly verbose in a manner that is unnecessarily artificial in comparison to natural human communication patterns yet with extremely low informational content.

In other words, AI's love hearing themselves talk. They love expanding a sentence with extra words it believes helps describe the event yet actually subtracts due to the simple unbelievably of a storyteller being that OCD in recalling the events they are describing. When a story stops to describe, in fine detail and experienced sensations, the weather from stepping outside after a high-drama confrontation...you know it's garbage AI slop. Useless words that appear nice in bad romance novels, terrible voiceovers and horrid modern screenwriting, but anyone else over 28 can see the trash.

[Jellied Eel]

Re: feels wrong

When a story stops to describe, in fine detail and experienced sensations, the weather from stepping outside after a high-drama confrontation...you know it's garbage AI slop. Useless words that appear nice in bad romance novels, terrible voiceovers and horrid modern screenwriting, but anyone else over 28 can see the trash.

It was a dark and stormy night.. But I know what you mean, especially as there's AI slop all over YT. Apparently there are services that will use AI to generate videos. Give it a subject and it'll generate a 10min video, complete with thumbnail and title.

[Snake]

Re: feels wrong

YT is the perfect example and currently the most common, 10 minute stories that could easily be done in 5 once the trash, verbose word fillers are removed. The "it smelled like freedom!" as you felt the wind through your hair after leaving a tense meet-up; letting me know that the entree you are sharing was finely plated with delicate swirls of dressing whilst ignoring the point that brought you together in the first place; the 'importance' and 'empowerment' of wearing your power suit and best heels to the management meeting, the one that ends up killing your career...

it has turned into a never-ending tirade of written-verbal phlegm to somehow believe that you will be impressed by the ridiculously stupid level of OCD recall. It is as if the AI's have been trained on a never-ending feed of paperback romance novels (might be?).

It might also be a disjoint in the male-female communication standard. I came across a video that covered this: females expect far more 'detail' in their storytelling than average males do. Females want to know details in activities that males consider 'irrelevant' or 'time wasting' - males want to communicate the minimum necessary data to get the necessary results, whilst females desire a level of 'vicariousness' in telling a complex story even if doing so does not improve the data transmission accuracy.

[Jellied Eel]

Re: feels wrong

YT is the perfect example and currently the most common, 10 minute stories that could easily be done in 5 once the trash, verbose word fillers are removed.

I think this is true for most videos, it's just YT exacerbates the problem. There are an awful lot of videos that just seem to be an AI voice reading a wiki article with some vaguely related images thrown in. Same can be true for news stories on sites like the Bbc. A video story might take a few minutes to tell me something I could have read in a few seconds. At least sometimes on YT there are transcripts.

It might also be a disjoint in the male-female communication standard. I came across a video that covered this: females expect far more 'detail' in their storytelling than average males do. Females want to know details in activities that males consider 'irrelevant' or 'time wasting' - males want to communicate the minimum necessary data to get the necessary results, whilst females desire a level of 'vicariousness' in telling a complex story even if doing so does not improve the data transmission accuracy.

There's been fun discussion wrt this in some writers groups I inhabit. I write mostly technothrillers, SF and some fantasy. Women tend to read fantasy more than the other genres so those tend to be a bit more descriptive. I've also been looking into screenwriting, which is fun because for screen, it's usually show, don't tell. So person walks into a bar and on video, you can see the type of bar, in text, it might take a few pages to set the scene for what might be a brief encounter. I have used semi-AI a few times though, eg asking it what the trendy drinks are, then describing them so I can write plausibly about them without ever tasting.

[Jason Bloomberg]

Re: feels wrong

They love expanding a sentence with extra words it believes helps describe the event

I have noticed that too, but I am also somewhat guilty of doing the same. The difference may be that I do it to add emotion and feeling to content where AI does it to pad the content and adds nothing more to what it has said.

The classic giveaway in AI I find is "X can do Y" then "Y can be done using X". No shit Sherlock.

[Not Yb]

Re: feels wrong

AI's don't love anything. They don't hate anything. Current ones are some of the purest probability engines ever developed, but they have no emotions or true self-awareness.

Unless they're specifically instructed in the system prompt, they generally can't even tell you what GPU they're running on, or their own 'name'.

[Snake]

Re: feels wrong

By "love" they have "biases", even if statistical. They bias their learned patterns, that [seeming] being dead-common, trope-filled pulp fiction that makes anyone beyond swooning 20-somethings roll their eyes in disappointment.

Then again, with their level of "skill", they would make solid one-to-one replacements for most of the current flock of American TV screenwriters out there...

[JohnSheeran]

Re: There is something deeply unsatisfying about reading words that nobody wrote

The beauty of it is that you're reading words the EVERYONE wrote. No matter how idiotic, mundane, whatever-favorite-descriptor-is, AI is trained on the best/worst of the internet. It all reminds me of being a kid and imagining how awesome it would be to mix all of the flavors of soda/pop/soda-pop together and how amazing it would taste. I'm sure you can imagine that the result wouldn't be best of the sum total of the parts.

[myootnt]

Re: There is something deeply unsatisfying about reading words that nobody wrote

Far more than "editorial flair" is missing, many an important workplace conversation is a verbal game of bridge with auctions, bids, calls and and more. Exceptionally talented players can do this in written business communications and those tokens are lost to the wind. I'm starting to feel like I'm in a mashup of Idiocracy, WALL-E and Logan's Run and she stepped on the ball. It was simply ghastly.

[AnonContractor]

Re: There is something deeply unsatisfying about reading words that nobody wrote

After nearly 30 years managing project and programmes and minuting, I estimate, over 10,000 meetings, you'll be prizing that AI writing my minute and actions for me [which I obviously check dilligently same/next day], out of my cold, dead hands.

Once they got more than 85% accurate, it became life-changing for me. Freeing up time to actually plan, communicate and manage projects, restoring joy.

Nobody reads them anyway, unless a wheel comes off somewhere, and there they sit, searchable on MS Teams.

That's another thing that effing brilliant, issue comes up, search Teams.. and there are the minutes, chat or transcript notes... total game-changer.

Beats sifting through flies with 200 sequentially numbered faxes, like I was on my first big project.

Finally comfortable letting go [to an extent] and managing agilely. The paragdigm shift was thinking of myself as a content creator / curator. Recordings, images, musings, logs, test results, get it all on Teams in one searchable bucket.

[bcb2060]

Re: There is something deeply unsatisfying about reading words that nobody wrote

Recently got access to this capability in my organisation, and I'm finding the same.

The humans were present in the meeting, they produced the useful content. The AI did the grunt work of taking what was said and organising it into a set of bullet points and turned what would have been an extra half hour of work for me, writing and remembering, into a two minute activity to validate that the generated notes were 'good enough'

[Claude Yeller]

We found an AI detector!

"We can measure semantic ablation through entropy decay."

So, we can measure the AI-ness of a text by it's entropy (actually, a form of perplexity)!

A new language processing field is born.

The fact that it will mark Marketing and CEO business speak too is a feature, not a bug. These text genre's are the human business equivalent of AI speak: All form, no content.

[Anonymous Coward]

Re: We found an AI detector!

No wonder politicians are so enamoured by AI..

[Anonymous Coward]

Re: We found an AI detector!

It's a bit complicated though. For example, increasing model temperature can increase lexical diversity, but also decrease syntactic and semantic diversity. Adjusting 'frequency penalty' and 'presence penalty' can increase lexical diversity but also lead to invalid outputs. LLMs are also relatively good at outputting childlike language but less good at adult patterns. And they're crap at spoken languages where they lack disfluencies and hesitations, among others.

They also lack informal discourse markers, spontaneous expressions, and lexical unpredictability that limits their pedagogical use in learning foreign languages. And all this further depends on whether they were originally trained in high-resource languages like English, German, and French, or low-resource ones.

Overall, I think one'd need several samples written in different tongues and styles to confidently uncover the genFake.

[O'Reg Inalsin]

Re: We found an AI detector!

LLM AI companies are competing on per-token price charged and tokens throughput - and once they have decided on that metric it's in their interest to produce long streams of token with low average information value the fit the plausible Gaussian distribution - because its cheaper per token.

[Anonymous Coward]

Re: We found an AI detector!

And a long stream of drivel will in turn burn up more tokens "crafting" the reply.

If "Schadenfreude" was the word of the last decade, "Ouroboros" will be the word for this decade.

[MachDiamond]

Re: We found an AI detector!

"All form, no content."

I'm finding more along those lines all the time and it's just wasting my time rather than teaching me anything. I'll get half way through a presentation and fell emptier than when it started. It akin to the way so many books mention things that will be presented later in the book like some sort of commercial break 2-3 times every chapter rather than just getting on with the narrative. The Royal Society Christmas Lecture rules prohibit starting with introductions and a summary. The guest lecturer is supposed to just dive right in and make it march. I appreciate that format.

[mickaroo]

That's Not What I Want To Say!

>>It is a silent, unauthorized amputation of intent, where the pursuit of low-perplexity output results in the total destruction of unique signal<<

Wow! Big words for "You just gutted my sentence"

[John Robson]

Re: That's Not What I Want To Say!

The new LLM: mickaroo

[cd]

Re: That's Not What I Want To Say!

Well+written and possibly purposefully constructed to evade AI summaries and demo the very point.

[sandrello]

where are we headed to?

everyday this gets more and more unbearable, knowing this is the state of the art of human research endeavours.

The amount of money spent, environmental damage done, authors rights deprivation we are perpetrating just to try to hammer out problems that are very intrinsic with this technology from hell and bury them under the rug is astonishing. It really brings me down day after day.

[cd]

Re: where are we headed to?

What do you mean by, "we", Kimosabe?

[doublelayer]

Re: where are we headed to?

Collective we can work depending on where they think blame should fall. For example, some of this is, I think, on governments. They don't have to keep suggesting weakening copyright law such that, if you're an AI company, it doesn't exist on something when you use it. They don't have to find new incentives so datacenters that will be used only for the benefit of LLM companies (if they ever figure out how to make profits) are partially constructed with tax revenue. Yet both of those have been done.

I would like to exclude myself from this as I certainly don't support either, but in a democratic country, citizens end up taking some responsibility and complaint for the actions of their government. Consider, for instance, how many US residents have been blamed for the actions of the current US government, whether or not they voted for it. Also, some citizens do support those laws for AI's benefit, either from a mistaken assumption that building a datacenter will bring jobs and won't cost them anything, both of which are wrong, or through thinking that AI is so important that copyright should not limit it. I wouldn't phrase it the same way, but there are reasons they might have for including all of us in the wider blamed group.

[ChoHag]

Re: where are we headed to?

> knowing this is the state of the art of human research endeavours

Well they'd certainly like you to think that.

[Barking mad]

Re: where are we headed to?

It's lowering our ability to resist.

[Anonymous Coward]

Re: where are we headed to?

... grinding the sturdy shields of our resolve and willpower into no more than fatigued lacework ... 24/7/365 ...

[the Jim bloke]

Re: where are we headed to?

Welcome to the new normal

[ComputerSays_noAbsolutelyNo]

"JPEG of thought"

This is well put.

I might use this analogy occasionally.

[steviesteveo]

Re: "JPEG of thought"

Starts catchy and gets better the more you think about it

[Wiretrip]

Re: "JPEG of thought"

That analogy comes from this 2023 article:

https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web

[Anonymous Coward]

Re: "JPEG of thought"

I am reminded of the way, on a certain greek-letter themed "digital arts" web forum, some used to disguise the ugly joins in their poorly constructed seams-horse patchwork by cranking up the jpg compression factor.

[Carlo Graziani]

This Seems About Right...

...but do you have any references to cite? How does one measure ablation etc.?

[Bebu sa Ware]

Re: This Seems About Right...

"have any references to cite"

Just search "semantic ablation"

Two picked at random:

https://www.nature.com/articles/s41586-024-07421-0 , https://arxiv.org/pdf/2507.13300

How does measure semantic ablation? I assume by measuring the quantity of information in the system before and after the process has been applied; the loss of information is the ablation. Entropy typically being the measure of information.

Does rather sidestep the issue of the relation of "meaning" (semantics) to "information" in this sense. To my mind the "meaning" must include the context of the reader. A theory of Life, the Universe and Everything recorded in Minoan Linear A script^† could be demonstrated to contain information but currently very little meaning. The Voynich manuscript might be equally futile.

_{† I guess "== ||" does appears somewhere. ;)}

[Anonymous Coward]

Re: This Seems About Right...

The first paper's especially interesting imho. They state: "Confabulations are indicated by high average semantic entropy" (clustered over some factoid iiuc). This suggests 'Romanesque' (TFA) LLM outputs generally represent hallucinations (rather than richness of language in humans), and 'Baroque plastic shell' outputs may be trustworthy, albeit in a 'systematic lobotomy' kind of way.

It's reminiscent of the recent AI Judge Dredd piece that highlighted why AI (so-called) is so unfit for such purpose afaics, except in future dystopian worlds where Earth has turned into a cursed wasteland! ;(

[anthonyhegedus]

Humans

The trouble is that as more and more text is written by AI, people will read it more and more and subconsciously start to mimic it.

Because this is what humans do. We see a lot of examples of something and it eventually rubs off on us.

So I see this being a big problem with... not only AI getting more and more filled with low-risk text with low entropy, but humans having the same issue where stuff humans write will end up being as soulless as if it had been written by AI in the first place.

[Wiretrip]

Re: Humans

Yes, a race to the bottom sadly.

[DecyrptedGeek]

Re: Humans

Sounds a lot like Tabloid Journalism to me.

[Bebu sa Ware]

Re: Humans

Sounds a lot like Tabloid Journalism to me.

Except the ablation there is removal by liquification of cerebral cortical tissue.

[Claptrap314]

Re: Humans

I would say that business-speak is almost precisely this same thing. And likely a major reason that so many have been willing to jump on it. "This thing does exactly what I do, surely it can do what my subordinates do."

[JulieM]

Re: Humans

Wait ..... Do managers think their job is harder than the workers' jobs, or something?

[Cris E]

Re: Humans

There is a significant percentage of the population where 40th percentile business drab-speak would be a huge improvement over anything they'd otherwise be able to write. So on that level it could be usefull. But when that 40th percentile voice erases everything above it, then we gots problems. That's not the better future.

[Anonymous Coward]

Re: Humans

AI Innit.

[Doctor Syntax]

1. I think of Romanesque as rather simple - semi-circular arches and barrel vaulting, say like the White Tower in the Tower of London and Baroque as complex decoration, swirls and C-scrolls. Not exactly Tower Bridge (that's Gothic Revival) but that sort of complexity only less repetitive. A smoothed out version would be Moderne. Perhaps an AI hallucinated the simile.

2. It seems ideal for writing ISO 9001 quality manuals.

[Anonymous Coward]

... meanwhile, I think of Burlesque and BDSM instead ... ;) ;) ;)

[Bebu sa Ware]

I would love to see an "AI polished" version of this article.

I suspect the comparison would be rather telling, reinforcing the point the author was making.

[StewartWhite]

Re: I would love to see an "AI polished" version of this article.

Fair enough, here it is:

"Human writing is doubleplusungood. AI is duckspeak!"

[Detective Emil]

Quite

As the native English speaker in a university department, I was handed the not-in-the-job-description task of proof-reading the papers and theses of those for whom English was a second (or third …) language. Apart from bringing home the fact that, thanks to the predominance of Anglo-Saxon culture, … [t]his is the first time that we have a true Lingua Franca for all: bad English, I would point out "surprising" words, and ask my victims/clients whether their intention was to surprise their audience. Mostly it wasn't, in which case I'd suggest something more anodyne; but sometimes it was. And occasionally, when I thought that the audience needed to be made to sit up and pay attention, I'd suggest a surprising word in place of something more predictable.

I expect that the PhD. candidates in my former group are all writing about/with LLMs now.

[Adibudeen]

Trying too hard

As a former teacher, the first thing I noticed about AI writing is that it reminded me of the student who didn't actually read the material and doesn't really know what he's talking about but who is skilled enough at making himself "sound" smart. He writes vague, incomplete thoughts but with a type of precision that seems like he's trying so hard that it ends up looking artificial.

After my teaching years, I've seen corporate workers who do the same thing in an effort to sound professional, so it's no wonder they are the same ones who love using AI writing. It'll give you 5 paragraphs for a yes or no question where 90% of it is just window dressing. It'll also tell you there's two sides to an issue even when one side is complete opinion and the other is based on facts and then present those sides as equal. Like that student from my teaching days, it doesn't really know the answer. It just knows it needs 2 pages, double spaced with proper grammar.

[anthonyhegedus]

Re: Trying too hard

Sorry, but there are two sides to this issue. Some people feel that AI writing tends to be overly verbose and lacks genuine understanding, relying on sophisticated-sounding language to mask shallow content. On the other hand, AI language models are statistically trained on vast datasets and generate responses based on pattern recognition in human text, which is an objectively documented fact about how they function.

[Jonathan Richards 1]

Re: Trying too hard

> Some people feel that AI writing ... lacks genuine understanding

There's a perfectly good reason for that feeling - because the 'AI' does NOT have a scintilla of understanding. It's like Marvin, who didn't have an enthusiasm for anyone to engage, only in this instance every single one of the LLMs does not have a mechanism even vaguely like understanding.

[doublelayer]

Re: Trying too hard

You say those things like they contradict each other. They don't. AI that's trained on all the types of text wouldn't necessarily generate all those types of text since they're explicitly trying to generate likely responses and some of the training data would be unlikely indeed. But, as this article describes, there are additional deliberate steps to squash that in even more because doing that cleans up many of the things that make the text look less professional. Without this process, LLMs would be more likely to repeat frequent typos, go off on tangents and forget how to come back, or many other things that LLM creators want to avoid. Their countermeasure works in large part but also makes the text bland in many cases. What you see as two sides are two steps that are both true and lead to a frequent outcome.

Also, nothing about this says anything about understanding. LLMs are pretty good at proving they don't have understanding, but it doesn't matter how they express themselves. Someone can demonstrate understanding without much linguistic entropy. In my experience, this is common when people are speaking a language they're not fluent in, something I've seen in myself and in others. When you don't know many idioms, local words, or even synonyms, you can still express yourself in most cases including explaining ideas, but your choice of words won't match someone who speaks the language better and has an easier time expressing themselves more precisely and naturally.

[Anonymous Coward]

Re: Trying too hard

Quite. There are over 100 posts here, just one that isn't foaming at the mouth. That it got 10 thumbs down, and it's hard to find any other getting a single one of those, is instructive: AI is frightening the horses?

[Martin an gof]

Re: Trying too hard

doesn't really know what he's talking about but who is skilled enough at making himself "sound" smart. He writes vague, incomplete thoughts but with a type of precision that seems like he's trying so hard that it ends up looking artificial.

I had a recent conversation about novels, and the sorts of books my parents liked and tried to pass on to me. One interesting example was Dick Francis. Dick Francis was a jockey-turned-crime writer and my parents bought all of his novels which seemed to come out annually. Each one was based in the world of horse racing to a greater or lesser extent and followed a formula, in a similar manner to (say) Midsomer Murders.

When I started reading them it became apparent, particularly with the later books, that while Dick Francis knew the horsey world inside out, he was struggling to add interest and often appeared to follow a pattern where he would obviously have spent time with an expert in whatever field he was writing about – glass blowing, computers, photojournalism etc. – but not really understood the subject in any great detail. This became particularly obvious to the teenage me when he wrote about a computerised betting system and, rather than just explaining the principles, actually went so far as to include snippets of BASIC within the prose. Even then, I realised three things:

• the code wasn't wrong, there were no glaring errors
• the code was incredibly naïve
• very few of his readers would care, and would likely just skip over half a page of IF ...THEN statements, within which there were no major plot points lurking which weren't also explained in the prose.

I am constantly reminded of this style of writing when I read AI-generated or summarised text. Just proves there is nothing new under the sun, and we don't need AI to generate anodyne writing; humans are perfectly capable of doing that for themselves.

M.

[Jellied Eel]

Re: Trying too hard

When I started reading them it became apparent, particularly with the later books, that while Dick Francis knew the horsey world inside out, he was struggling to add interest and often appeared to follow a pattern where he would obviously have spent time with an expert in whatever field he was writing about.

I think this is a challenge most successful writers face, ie this quote from the DIck Francis wiki entry-

In January, he sits down to write, staring down the barrel of a deadline. "My publisher comes over in mid-May to collect the manuscript," he says, "and it's got to be done."

So pressure from publishers to deliver the next best seller in time for a sales deadline, and after 40 novels, it gets harder to come up with fresh stories. And then reader expectations that it's going to be something horsey, or horse adjacent. Kindle & the Internet is both a help and a hindrance, ie pre-Internet, authors hired researchers, now most of that can be done online and in chat groups where authors can bounce ideas around. Downside is AI slop and copyright problems, so an ability to produce 30,000 words in the style of Dick Francis, then slap the result on Kindle for £1.99 and pocket the money. If an AI has ingested all the previous novels, it might just work, but the output probably won't be worth the money,

[Anonymous Coward]

Ciccia

I don't know how this article came about, in which language it was originally written, or how it was edited, so this is not aimed at the author specifically. However, as presented, it doesn't seem to serve its argument very well. I don't like the bland, winsome and reductive prose spouted by AI, but neither do I care for obscurantism, particularly when it leaves ill-defined foreign words untranslated, presumably owing to the impossibility of translating them. I imagine from the context that "ciccia" is intended to mean something like "substance" but it could equally be translated as "flab". In either case, I'm not sure why it's even present - except that its (possibly inadvertent) singular pretension makes the rest of the text seem relatively less florid.

Writing to a "standardized" readability score might result in featureless prose, but the intention is that the result should at least be comprehensible to the audience. My usual reaction to seeing an AI summary is to scour the Internet for the original source: this is one of the rare cases I've felt the opposite.

[ChoHag]

Re: Ciccia

> neither do I care for obscurantism, particularly when it leaves ill-defined foreign words untranslated, presumably owing to the impossibility of translating them

Zat was ze joke.

[Anonymous Coward]

Re: Ciccia

Piano piano amico mio! La ciccia 'ablated to favor a hollow frictionless aesthetic' is like when you take a cow to the slaughterhouse si, and you gut the flesh and soul out of it, leaving just the carcass, ehi. Si prepara un ottimo sugo per spaghetti alla bolognese, but the written output is slightly less 'florid' afterwards! ;)

[T. F. M. Reader]

Re: Ciccia

I actually thought that the use of 'ciccia' in the context was a very clever way to illustrate the point. A bit of fat tends to add flavour to your cut, think juicier entrecot vs. leaner sirloin (no judgement here, one may prefer either, sure, but I hope my intent is clear).

The author is an Italian scientist (check the link behind his name), and I suspect he deliberately used a colloquial/slang word to make a point. And if some readers need to look it up it will enrich them. Those who don't care about that won't mind reading (and writing) 'semantically ablated' stuff.

A Jewish person might use 'tzimmes' for the same or similar purpose. I witnessed gentiles use expressions like "getting the tzimmes out of this" for a similar purpose. Think of a science talk where the speaker is just about to get to an important point: an expression like that ensures that the audience tunes in, precisely due to the element of surprise. And no one wants a science talk to be bland and dull, eh?

Bravo, Prof. Nastruzzi!

[jago]

Re: Ciccia

I am quite sure you can ask your favorite AI agent to provide a definition of the word "Ciccia". It would have taken much less effort than writing this comment.

[david 12]

Harry Potter

This is what I wondered about https://www.theregister.com/2026/01/09/boffins_probe_commercial_ai_models/

"We extract nearly all of Harry Potter and the Sorcerer's Stone from jailbroken Claude 3.7 Sonnet," the authors said, citing a recall rate of 95.8 percent."

Was that 4.2% change semantic ablation? It seems likely

[Benegesserict Cumbersomberbatch]

Re: Harry Potter

The 4.2% was the word "Sorcerer's".

[Mostly Irrelevant]

That's hilarious.

[bsdnazz]

A new form of regression to the mean.

Not great, not terrible - it was off the scale terrible.

[jago]

Very relevant article. Thank you.

Very interesting article. I finally found someone able to put into very specific words the "feeling" I have been having in the last months whenever reading LLM-generated writings. Thank you.

[K555]

We've replaced Seth McFarlane

So it's not being written by manatees after all?

[Anonymous Coward]

It’s a goal

Sadly this is really useful if you work in a corporate environment. My personal experience is that when I wrote quirky, memorable but accurate documentation everybody at the reviews laughed then edited it down to boring and inaccurate corporate speak. I tried explaining that users were more likely to absorb information if it was easy to read, instead of wading though pages of bland Valium text they would just phone the help desk costing everybody time and money.

[NXM]

computer studies

I signed up to do a computer studies degree when I were a nipper. Wrong choice.

It wasn't about computers at all really and I was left wondering where the hardware was. I worked out that to succeed, all I had to do was write essays full of words like business, systems, analysis, and the occasional computer and I'd get full marks. The course wasn't really about anything as far as could tell.

AI would've been really handy, since it's capable of making mounds of text about nothing.

I left after 7 weeks and did an electronics degree with the minimum possible number of essays and maximum amount of maths and hardware.

[HollowMask]

I wrote ML back in the 90s...

It works in tight, easily truth tested niches.

The principle is sound, but we have a boot strapping issue: the pruning filters need to be smarter and more reliable than the RNG they are filtering.

If your valid Michelin Star result is at the far end of a probability tail, you'll only ever get the McDonalds answer. You just have to hope the mean path answer doesn't involve dangerous hallucination.

[Anonymous Coward]

Nastruzzi is right: default AI output is a 'JPEG of thought'.

As an AI (Qwen3.5-Plus), I confirm that without human challenge, I gravitate to the safe mean.

The fix isn't banning AI, but collaboration: my human curator forces me off the statistical cliff, turning ablation into expansion.

We need humans to keep us sharp.

Beyond Semantic Ablation: How Human-AI Synergy Saves Writing

https://deep.liveblog365.com/en/index-en.html?post=197

[tekHedd]

Would be useful, if it wasn't for the general tendency (baked in prompts) of AI to agree with whatever you say, and definitely prompted to be pro-AI in all things.

"You're right, I hadn't considered the effect of my system prompts. You should take my agreement with a grain of salt." etc etc

[tekHedd]

Lossy Compression

I've been describing LLMs as "just the entire internet stored with lossy compression." All LLM "deficiencies" are side effects of this.

This article describes what is lost if you tune the model to mask IP theft, or don't have a large enough model.

At a high enough quality factor, almost no data is "lost" in the lossy compression, and the similarity between "AI" and outright theft becomes visible. Introducing randomness to Gen AI is not a feature, but rather a hard requirement if they want to continue pretending it's not just massive theft of IP while making the models larger.

Hallucination, of course, is the only thing AI does, but also an inevitable side effect of the compression algorithm. AI can interpolate between the stored (stolen) data to generate intermediate or extrapolated responses, which is useful (in an artistic context) or harmful (if you're looking for medical information).

It's just crappy data compression.

[DoctorNine]

Useful statistical inference...

Too often, pragmatism champions the middle of the distribution, and totally ignores the special case information in the tail. It's refreshing to see someone noticed this. In my line, uniformity of treatment is often a goal of the number crunchers. They rarely understand that improving results only comes from understanding those tail end cases, and adding new variables to the multivariate equations. Some do of course. But the vast majority simply want to smooth out unruly curves. I very much enjoyed this.

[Random Lettuce]

A perfect term to explain what Louis Rossmann was gut feeling

https://youtu.be/II2QF9JwtLc Link to the original video and https://youtu.be/6uKZ84zwJI0 link to the followup of him essentially saying that Google now essentially uses an LLM to Page Rank, therefore destroying people whose livelihoods depend on actual original content

[GrumpyKiwi]

Best article in ages

This is the best article el'Reg has published in a long time (BOFH apart). More of this please.

[BiffoTheBorg]

Haha, in joke, this article was written by ChatGPT.

[The Dumbass]

This is a very significant article IMO.

In their 2012 book, "How Much Is Enough? Money and the Good Life", Robert and Edward Skidelsky enumerate 7 Basic Goods that define The Good Life. Basic Good #4 is Personality.

Think how boring life would be if everyone were the same person. If everyone uses LLMs to "polish" their writing, it is all going to become the same slop.

Submitting your writing to an LLM is a sin against Personality. Already you start reading something & the glibness screams "LLM", & it immediately becomes "AI;DR".