There's some psychological mechanism by which my brain immediately recognizes AI generated text and just short-circuits to "there is no information here".

And when I force myself to read AI-generated text I realize I'm making my brain do creative work to impart meaning to the words. It is exhausting because my brain is literally trying to do a just-in-time rewrite of the text into something valuable.

Something is deeply wrong with AI generated output, and I say this as someone who is typically very impressed by AI.

My son is currently learning Romanian and I was trying to help him with verbs. I don’t know Romanian but recalled when learning a foreign language for the first time it really helped me to break down how a verb form or tense worked in English, then learn the equivalent in the new language. So I wanted to make some charts and pages that he could use as learning resources.

I used Claude to help. I don’t know how to quite describe it, but because the text was polished and well constructed my brain was giving me the the signal “if you aren’t getting this it’s because you’re not focusing” so I’d read it again and then again and it still was not landing. It sorta felt like when you read something technical or heavy when very tired - you are reading but not processing.

Only after wrestling with this for a few days did I realize that it wasn’t me. As I started going through, sentence by sentence, forcing it to re-write things to be more clear the concepts became easy to understand.

I wish there was a name for this situation. It’s almost like a pseudo-language where it has the correct form and presentation but is missing critical components.

The more complex the topic, the more I sense this.

I've been stuck in loop all week where a coworker asked me to review some methodology documents that are clearly generated by Claude (which is fine in this case) but every time I sit down my brain can't get through 2 sentences without feeling like I'm reading nothing. It actually feels like the content, which isn't empty, has no meaning. So I go on to another assignment and come back later and the same thing happens. And now I have anxiety about opening it so it has compounded in difficulty haha. Glad other people are experiencing this.

I also find it impossible to parse half the comments that Claude tries to sneak into our pull requests. I’ve never had an issue understanding code comments written by humans like this before. The structure of the information is like a waterfall that leaves me unable to swim to the surface and grab the air of comprehension.

So, “Please write a one-liner comment manually to replace these 5 lines of AI generated comment” is a common refrain in my PR reviews to colleagues.

That last image is bizarre. The quiche, cream and even the salad look like they've been given the trypophobia treatment.

Which might even make sense, because there were always (still are?) those horrible ads in the chumbox area of news sites that used trypophobia and other creepy body-horror stuff to get you to click. [1] So maybe the hope is that you don't really look closely at the quiche, but some reptilian party of the brain gets oddly activated and drives you towards the restaurant?

1. https://medium.com/the-awl/a-complete-taxonomy-of-internet-c...

  "There's an ongoing discussion of whether humans are good at recognizing AI-generated text. While most research claims that humans don't really do a good job there, I disagree. "

I wonder if humans that spend all day working in tech are good at recognizing AI-generated text, but people who spend all day doing jobs that don't involve computers aren't as good.

And I wonder if those of us in tech are the only ones who really care?

Claude has become noticeably, painfully worse at writing in the last six months. At this point it’s practically useless for anything except code.

In high school, a teacher gave me a copy of "How To Read Better And Faster" which teaches you speed reading. This came in very handy in college.

I find that when I try to speed read modern human writing, there are often errors (like missing or misused words) or awkward expressions that I do have to slow down and think harder a lot to really parse it.

With AI writing, it's sort of self redundant and the information density of each sentence seems to have more even information density. This makes it very easy to do a very high level speed read and get the full gist.

There are also what I'm assuming are bots on hugging face (or maybe non-native english speakers who are using ai for translation) that interact with me where I have no idea what they are saying until I read it very slowly.

I'm starting to have the same problem with code. It's unbearable for some reason. It's like someone took the LinkedIn writing style and translated into a code style for code that has no business existing in the first place.

AI is missing the part of intelligence which involves insight. LLMs generate language blobs with highly sophisticated correlations. The governing insight in human cognition is essential to intelligence. It prunes the tree. It puts things in context. Everything is important to an LLM because insight intelligence is missing from the technology. The usefulness of LLMs doesn't tell us they are intelligent. It tells us how brilliant an invention language was.

At least when it comes to code comments and error messages, I have found that Sol produces substantially better prose than Opus 5 (sometimes literally 10x more terse for the same prompt). When I use Opus 5 for coding, I will often give Sol a prompt like this, and it works quite well:

Review changes in the current working tree for comments and error messages that are verbose and excessively detailed, and make them more concise while still informative and accurate. Do not hesitate to delete comments altogether if they are not helpful.

I find this struggle with coding too. When Claude spits out a plan I have to really focus to not skim over the details. It's also difficult to parse the plan because I have to work backwards to determine if the high level context matches my understanding of what Claude will do.

I've been given AI generated documents or presentations and you can always tell when they've just one-shotted the output. Claude has a way of adding so much jargon and unnecessary text to a document that it makes it very hard to read or even understand what the original intent of the writing even was.

I'm not sure why this issue is so prevalent, it's not hard to point Claude at the Wikipedia article on signs of AI writing or ask Claude to write content anyone of the average American reading level could understand.

To me it just gives off a sense of laziness, that you cared so little of your content that you did not take the time to read it yourself and edit it to effectively communicate the message you wanted to communicate. To that point, it's just not worth my time reading, otherwise my eyes glaze over trying to read between the lines of machine written language for other machines.

I found that it's not just text. It's also design and overall user experience. If you have an eye for it, your brain picks up on it immediately with the little pills, busy landing pages, the use of serifs everywhere, bronze accents, and lots of dark mode.

I wrote about this recently on one of our company's IT channels. I called it "slop-fatigue". I - and others - have started to get noticeably dismissive of content that appears to be written by an LLM. The whole thing was a piece about the ironies of software development in the world of LLMs. This part spoke about how easy it is to generate docs now, but it's also even easier to dismiss it. This contributed to cognitive-debt, because humans didn't write the docs, and other humans tend to dismiss the content. I saw a recent reaction to this on HN as `ai;dr`. That's almost exactly how I feel about it.

Margaret Storey's recent piece on Cognitive Debt: https://spawn-queue.acm.org/doi/10.1145/3807966

I'm struggling to find the right balance of AI generated text vs. my own words. When I filter something through AI it definitely sounds grammatically better, but ya, I'm scared people will start to just ignore it.

I've taken to asking the engineers on my team to just share the original prompt, no judgement because its frustrating having to read through pages of AI generated technical documentation. Another approach is to feed the AI generated content into AI... who knows where this will end.

Part of me likes the cliche Claude voice. Not because it's good, but because I can immediately recognize it. When I see it in the Claude app/code then it's fine. In the wild it's a sign to me that I shouldn't keep reading.

I need to support integrators for a mobile app SDK. Bank stuff.

We have one integrator which consistently uses a truly bad AI. My brain skips after two sentences already. There are always +9 questions what should just be 2 max. Redundant info is requested (e.g. please provide a change history of this function and if it will be deprecated) and everything is just unbelievably verbose.

90% of questions we get from them are already answered by the web documentation (which even has a search function) or are just non sense.

I'm truly considering to build an MCP just for them...

It may be nostalgia but I feel like we reached peak humanness with GPT-4o and since then it's been getting more and more alien. Particularly Fable.

I don't like AI-generated text at work, at all. It feels lifeless and unfocused.

But what I hate the most is that it is objectively better than what I had before. No typos, clear structure, and, regrettably, the verbosity and autistic obsession with detail of the LLM is more actionable and useful than the human guy who wrote lists of commands and URLs as documentation, without explaining anything. Or the colleague who writes in uppercase and with question marks and who doesn't make any sense and forces me to engage in an interrogation effort to get to the bottom of what they are trying to say. Or the colleague who simply hates writing--despite being decent at it--and will call you to give you a meandering verbal explanation that lasts two hours of what they want from you. The cynic in me bemoans that we brought this upon ourselves, in more than one way.

OT, but I feel the title is wrong, or at least ambiguous. For the author, "AI blind" means they can't process text produced with AI. For me, "AI blind" rather means one can't see when something has been AI-generated.

It can be helpful to read LLM text the same way you read technical fluff - start from the assumption its mostly garbage with a core technical point. I find skim reading the output to find the useful bit reduces the friction.

When I ask ChatGPT questions I usually only read paragraphs 2 and 3. The first paragraph is glazing me, anything after paragraph 3 is repeating what was said earlier.

I just throw it in AI; if it’s tech docs, I tell it to build whatever is there, if it’s something else I tell it to give me max 5 points to tell me the meat of it. Either way, I will never read it myself. This works well for tech docs/apis really; it usually gets it to work fine without me checking, even though the docs are usually subtly wrong in many places; after that I can read working code and that works a lot better for my brain.

LLM text and images have crept into so many places, and I've experienced quite a lot of existential dread this year from it. I've definitely felt what the author wrote about here.

Another trend, starting sometime last year, has been witnessing people reading scripts on camera: YouTube, public speeches, advertisements, and podcasts where a real human is reading from a script that is LLM-generated.

In the case of smaller video creators and podcasters, I then start wondering if they really think they can pretend that this is organic human speech? Or do they think I'm stupid and will continue to consume this slop without any critical thought..? Or do they think that their audience must know, and therefore consent to it, and so it's okay? In some cases, the person reading the LLM script is literally unaware that they're reading LLM text because a human scriptwriter handed them the page. I see this with advertising copy and a few YouTube channels.

Either way, it has led to quite a few instances where I got the LLM "ick", like I'm witnessing an LLM wearing a human skin, and it grosses me out so much.

As one final note: some of us assumed in ~2023 that LLMs would help to improve technical documentation. Thus far, 3 years after that prediction, I can say definitively that we've not ushered in an era of great documentation. We've got a steadily increasing percentage of documents that cause my eyes to glaze over because I don't know if a human expert ever once vetted it.

> A technical requirements document that describes a rather simple concept in a very verbose way.

This is not AI specific. I have come across many humans who describe a simple concept in a very complex and verbose manner.

There is something liminal about AI-generated text indeed. First few sentences are promising, then my expectations of reading something tight evaporates eventually. I feel let down, (almost) being tricked into believing that the intro and the perfect grammar should've led to a revelation written by a deep thinker. Instead, it's just... a lump of text, stitched together because the training corpus said these words go together.

My current theory is that this reflects a weakness in Claude’s ability to see the “big picture”.

When writing code, I have to explicitly tell it how to structure things at a high level, or the result is sort of a flattened spaghetti. Similarly, when it’s explaining things, it’s not good at pulling out unifying concepts and explaining top-down as a smart human would do. It groups little things together but often doesn’t generalize or synthesize explanatory connections from them.

I’ve been experimenting with explicitly working through a sequence of outputs at different levels of detail, but I haven’t found a consistently successful method.

Recognizing AI generated text is incredibly easy if you're someone who communicates with AI all day.

I've had the same trouble with AI tech docs, and I struggled to articulate it. There is something difficult about trying to point to the specific "problem" with a given document

I've been a big reader, and many AI outputs nowadays reads polished similar to published books.

The reason I brought it up is because, people who learn English normaly start with a book. It's heavily polished.

When you speak English as you learned from the books, it does not sound very conversational.

If you are native/fluent English speaker, you can feel the impedance mismatch and feel something's off.

The AI-blindness stems from the fact that those polished edits are so common in publishing field, they all sound the same, and unable to recognize the diffs between AI-generated and human-generated.

There is no real human conversational vibe to them and well. i will stop now.

These days all I see is, people on the sending side produce huge amount of AI generated text with zero understanding and the people on the recieveing side feed that same text into some other (or perhaps same) AI to scavenge meaning from it. And they do this so much so, that I some time wonder, if we could invent a high density wire/binary format for AI outputs and enable direct agent to agent communication to save some energy and bandwidth.

AI blindness is making my reviews of others' PRs harder and harder.

There's an amortization aspect as well.

If I'm going to be iterating on one document or idea in an extended manner with feedback from the chatbot, I will make an effort to setup the decorum it should follow, because it's a small proportion of the time in that chat.

But the guy just trying to get a report or email out quickly? Not so much.

Something that bothers me about AI generated content more broadly is how unmemorable it is. I don't mean as in bad. I mean literally, as in hard to remember or recall.

Despite seeing a lot of them, I cannot think of one AI-generated photo that I can picture clearly in my mind; a few are partial but elusive. Whereas I can recall (visualise) a whole bunch of traditional photographs.

The same is true of AI generated text. Only the annoyances stick. I cannot recall real details of text I have generated, until I commit it to memory some other way.

I don't think this is about ephemerality either. If we assume it's about celebrated/famous/infamous images, there are definitely non-ephemeral, cultural moments in AI generated images in particular, like Boris Eldagsen's Sony Prize winner:

https://petapixel.com/2023/04/14/artist-refuses-prize-after-...

This really should be memorable, but isn't. I had forgotten the second person is in the image.

Or Jason Allen's fake painting:

https://petapixel.com/2022/09/01/ai-generated-artwork-wins-f...

I had already forgotten there's more than one figure in it, and I only looked at it a few weeks back. I remember the colour, the bright circle, some vague hints of texture; one figure. And that is it. Only the crudest shape elements.

For me, something about AI-generated text and images confounds recall. It is really peculiar.

I used some check marks and x-es in a work chat, because it was easy to do, and so I could highlight the good and bad outcomes, and someone immediately asked me if I was using AI. I was caught totally flat-footed, because I hadn't used AI, but it looked VERY MUCH as though I had.

I don't care if it is AI is involved or not, but I care how much attention and thought there was behind the final text (or code) from the human behind it, and it shows.

Ironically, the “summary of the situation” linked within this article seems clearly written with AI.

To me AI text is only for the consumption of other models. Reading it is like is like tapping into the circuitry on a motherboard to read what the PC is doing.

Interesting. One needs to be exposed to a certain level of AI flows to become numb to them. Wonder how they read to someone who isn’t exposed yet.

Idk I’m starting to have have trouble with the frontier models and a couple how to write like a human skills…

There is another aspect to this that I find even worse than LLM writing style: badly-human-written documents and posts.

You know this: you read something and it's like 'here's a mass of poorly structured text, it's not clear what the points are, it reads ugly, I have to read to the end to discover that there are no points! Or if there are points i now have to reread it. What a lot of effort.'

LLMs have annoying grammatical tics and can have other problems too, like the author says, but sometimes they can help to take bad writing and make it clear and more digestible.

Why read more word when less word do trick?

If you ever meet someone you knew from an online community it's a blast. Now reverse that: you know somebody IRL and find their twitter account, and it's a fucking nightmare. Like you are looking inside their brain's content and it just collapses the image you had in mind of their personality.

AI is surfacing layers you never had a chance to see, makes you rethink your career path and who you are willing to work with very fast.

ITT still, among the highly technical, a surprising incomprehension of the fact that describing LLM in terms of "token prediction" and dismissing this as statistical,

neatly dodges literally everything that is interesting about what they do.

You to, friend, consume inputs and generate outputs.

What's interesting is how you do that, and, if you prefer to look at it through a technician's lens, whether or not what is done is reducible.

I don't mean quantizing the model, great, now you have a crankshaft with no oil. But the motion of the pistons and wheels is roughly the same.

What I mean is, to put it in plain terms, the only way you find out what a model is going to "predict" from a given input is to ask it.

If your mental model is still that LLM are "glorified overhyped giant markov chains" performing "parroting" you need to improve your understanding.

I think AIs are writing less and less for humans. The fact that reasoning trace is using English and we require them to be performant means the models need to squeeze in more of information into it. More information that is perfectly fine for AI to ingest, but it makes the text less and less comprehensible to humans. It's almost like steganography where model encodes crucial information in phrases meaningless or empty to a human. Human directed messages for LLMs become a rounding error they don't pay that much attention to. Efficacy of reasoning is what dictates their style.

We can no longer assume that the models will create a passable English in their user targeted messages. So in the short term, we need to have a tight guidelines in the harness, but in the long term, we'll probably need separate languages for thinking traces and for human communication. Possibly even not going through tokens at all when doing thinking part.

Alternatively, we could evaluate thinking traces on human readibility, but that's probably infeasible due to sheer volume.

I find that typical human content has about 5% information density, while typical AI content is about 1%. AI is much worse, but the same extreme skimming techniques apply to both. In both cases I'm like "I have no idea why you had to say all those words to make this point, but let's get on with it."

I can relate. I noticed this exact phenomenon when encountering NotebookLM-generated diagrams recently. Even if I know there's some intelligent thought behind one, it's like some slop detection circuit-breaker is tripped.

People love to inflate their jobs if there is no cost.

I saw my ex-coworker tried to inflate text that can be expressed in couple of bullet points text. It doesn't add anything. It just fill the fancy flavour text that looks like professional. Then the reader of that inflated text is also another coworker in other team.

I said "Forget it and just send these bullet points" but he refused.

Bullshit job at it's finest I guess.

AI;DR

Having just a little background in traditional linguistic discourse analysis help understand why this is.

The best basic explanation I've found is this video.

https://www.youtube.com/watch?v=ORgKY9AlybA

The real answer is to take a class in discourse analysis and learn how to pick apart sentences. Once you start understanding how language works in the brain you can start to see better how LLMs fail.

Oh man I can so relate to this.

read: "I am becoming an AI"

if you go on x.com you will see that, most of the video/photos are likely AI, and is fatiguing to trying to filter it. Probably why Nikita left, he know that sht is going down

[dead]

[flagged]

[dead]

As people rely more on AI they experience cognitive atrophy. This is measurable in IQ loss, and other symptoms we might otherwise associate with early onset dementia or Chronic traumatic encephalopathy.

Prompt an AI with:

  What is the best chicken noodle soup I can make?

And you will get at least one recipe which surely is tasty.

Ask a person the same question and you might be told:

  One you make for someone you love.

This is the difference between a statistically probable response and understanding.