Reacting to LLM writing

When I groggily get out of bed in the morning, the first thing I do is go to my phone and see if Substack has emailed me any notifications. On one recent morning, I saw that I’d received a comment on my post about measuring pitch, from an account called ‘Neural Foundry’:

The autocorrelation approach is smarter than FFT for this because of the finer resolution, but what's interesting is where it fails. The missing fundamental illusion shows pitch perception isn't just mechanical signal processing but involves psychoacoustic interpretation. I've debuged similar octave errors in my own audio work, and the fact that periods can double wihout changing the underlying harmonics makes these edge cases inevitable. The comparison between Carpenter's accuracy and modern pitch-corrected vocals is a nice empirical demonstration of how tech changed studio norms.

This read as LLM to me, but was there half an insight in there? After more careful study with a clearer head after breakfast, I blocked the account. If I am wrong, then they will not be able to complain in the comments to this post.

The first sentence is very bland. The second sentence caused me some initial hesitation – I had knowingly omitted some commentary on missing fundamentals, and I wondered if it had filled in a gap I’d left. But on a closer read it was purely regurgitative. The third sentence is in first-person and contains two typos; in isolation, I would assume it was human. The final sentence is a dubious remark on a topic that doesn’t cohere with the rest of the comment.

While the first three sentences seem loosely connected, the sudden jump to Karen Carpenter at the end makes me feel that a more harmonious interpretation is that each sentence’s subject matter is independent of the others.

Comments left by the account on other Substacks share the same bland style. I judged it to be LLM spam, the first-person anecdote deceitful, the only purpose of the whole being to drive clicks to the spammer’s profile. Pangram, by popular acclaim the only good AI-detection tool for writing,1 also flags the text as AI.

One day (soon?) I will have to figure out what to do with insightful spam comments that deepen my understanding of the world, but we are not quite there yet.

Sarah Labrie, 21 November 2025:

there’s an op-ed in the NYT today that was very clearly written by (or at least with a lot of help from) chatgpt and I know no one cares, but this just feels like the end of something.

The piece referred to is a Modern Love column, headlined ‘I Was Deemed Unfit to Be a Mother’:

By then I was fully sober but trying too hard. I smiled too big. Overexplained. Looked at him like I was begging him to see how much I loved him. But love that comes with an apology in every glance is a heavy thing to hand a child.

The real unraveling came quietly, in slow motion.

He grew more distant. His smile faded quickly. Every affectionate moment seemed to cost him something when we parted. And eventually, the cost got too high.

He stopped answering my calls. Stopped smiling. Stopped looking back. And I couldn’t steer him away from his own self-protective impulses.

I don’t think I would have picked this as AI, though I sort-of see the tells. Perhaps I don’t read enough of this genre, or enough AI-generated long-form prose, to recognise the rhythms; I’m used to bolded emphasis and emoji bullet points. Pangram’s assessment agrees with Labrie: most of the article is AI-written. (I fed it a few other recent Modern Love columns, each of which it deemed human.)

While several commenters on Labrie’s note expressed general agreement with her, one wrote, “What!!! I read this and was in near tears. Are you sure? Like how can you know? So pissed that I can be manipulated by AI…” At least I’m not the only person with non-expert AI recognition skills!

John Scalzi recently tweeted on Bluesky:

I think a lot of people who don't often write (or who just don't care) [don’t] understand how easy it is to spot "AI" text once you've been exposed to it. It's gotten to the point that I can look at a piece of "AI" writing and tell if it was written by ChatGPT or some other LLM. They have "house styles."

I certainly can’t tell the difference between ChatGPT and Claude. Perhaps my most common reaction to LLM writing is non-recognition.2

Sometimes it’s really super-obvious though. In my roles in local chess administration, I sometimes receive emails from prospective players, and sometimes these emailers are teenagers. They might write tersely and informally, or perhaps overly formally; that’s all fine. Last year one wrote to me with introductory words that sounded like a politician too-smoothly hitting a rehearsed talking point, but:

• the lists of questions in the emails were very clear;

• I knew exactly what information was being asked of me;

• the communication was efficient.

It was an unusual but generally positive experience. There’s a joke that people are getting AI to draft emails based on some bullet points, and recipients are getting AI to summarise the emails back into bullet points, but honestly ChatGPT can write a good, succinct email. I ramble a lot by comparison. I never ask for AI summaries.

The Facebook algorithm has recently showed me some posts from ‘North Perth Views’, North Perth being a suburb that I don’t live in. This is extremely ChatGPT:

So who should be in control of cycling infrastructure and promotion?

This series has shown a clear pattern:

🚲 Cycling treated as optional
🚗 Cars embedded as the default
🛣️ Major rebuilds that lock out safe riding for decades

Whether cycling stays with the City of Perth or shifts toward State control is an open question. The State already funds most major infrastructure — and has taken over other local government functions before. So it’s not impossible.

But here’s the part that doesn’t depend on who’s in charge:

Hahaha I love those emoji bullets; I am one with the RLHF raters. Maybe you feel like clawing your eyes out, but I don’t mind this sort of thing. I’m sure that behind the LLM text there’s a real human cyclist feeling real human cyclist frustration at living in the most car-centric capital city in Australia. Using an LLM to try to build community support for a genuinely-felt cause in local politics is fine with me. The neutrally-competent writing style means that readers can argue about the content rather than getting distracted by spelling mistakes or whatever. Hopefully Russian bot nets don’t get involved.

Elsewhere on Facebook, the algorithm knows that I’ll read old cricket stories. Lankan Lions generates a lot of these; AI. I don’t think I was ever reading the nostalgia-bait too closely. I’d see the picture, read the first sentence and chuckle at the memory, then skim the rest. After seeing the Pangram verdict, I read some of the posts more carefully, and found occasional strange or inaccurate descriptions. I haven’t blocked the account.

(Pangram assesses Cricket Superstars in the Leagues as human; it’s written by Scott Oliver, sharing stories from his in-progress book series. It’s great stuff, highly recommended for cricket fans. Most of my readership is American, and a rough equivalent would be like if there were major league baseballers who sometimes played a season of single-A ball and crushed their opponents.)

Philip Trammell and Dwarkesh Patel recently kicked off a round of discourse with the Substack post ‘Capital in the 22nd Century’, arguing that once AI can fully and cheaply substitute for human labour, income will go to the owners of the AI capital. Since inequality of wealth is much greater than inequality of income, the labour share of income dropping to zero would lead to a drastic increase in inequality, with plenty of associated political implications.

The arguments are being contested, perhaps most notably by Brian Albrecht on Twitter, and I am interested in following the debate.3 Rob Wiblin chimed in:

Brian makes a good point: markets don't automatically eliminate less productive producers. Efficiency doesn't win out instantly.

But while the industry dispersion data is useful I don't think it supports the conclusion as strongly as presented:

The semiconductor figure is suspect. NAICS 3344 is "Semiconductor and Other Electronic Component Manufacturing" — which lumps together cutting-edge logic fabs, legacy automotive chip producers, PCB assembly shops, capacitor manufacturers, and connector makers. These aren't competing products. A 30-40x "productivity gap" between a small PCB shop and a leading-edge fab isn't evidence that inefficient competitors persist despite massive disadvantages — it's comparing apples to industrial machinery.

…

No-one writes like this, especially on Twitter. I think that if I had typed something with my own two hands, and someone replied to me with an LLM essay, then I would be irritated. “It's co-written with Claude because you'd be crazy to write stuff with[out] AI assistance these days,” Wiblin wrote in reply to someone chiding him for the “Claude voice”.

Eh. The transparency (there was no attempt at hiding the style; people in a discussion about AI would recognise it as AI) is a point in its favour. As long as the poster is engaging in good faith, then I can grudgingly accept that this could be a productive future of written debate. Two AI instances arguing back and forth, maybe eventually reaching agreement. Even if it works though, I don’t like it, either in isolation or as part of a broader pattern of behaviour.

“As long as the poster is engaging in good faith” is a strong condition. There will be plenty of bad actors eager to join in political debates, and I worry that normalising LLM essays in this context will tilt the balance of discourse towards the bad actors.

(There is much worse AI content on Twitter, but Grok is out of scope for this post, since I’m focusing on LLM writing presented as coming from a human.)

Peter Wildeford has been an online acquaintance since the early days of Effective Altruism; we’ve been Facebook friends for a long time, though he has gone onto bigger things than I have and we rarely interact. Peter is an extremely good forecaster, in the sense of estimating percentage probabilities of future events; basically trying to set betting odds for non-sports topics, though there is not necessarily any money being wagered. (This old post of mine, looking at my performance in a forecasting tournament, discusses how the quality of such probabilistic forecasts can be measured; Peter finished near the top of that tournament, a couple of deciles ahead of me.)

I subscribe to his Substack in part for these contingent historical reasons and in part for his analysis of AI developments and assessment of future trajectories. I follow him on Twitter and know that he uses AI because he tweets about it sometimes; I don’t expect that on bread and water alone he can really punch out so many hundreds of words of background about the situation in Venezuela in so few hours. Pangram says it’s AI, of course it does. I don’t have to read it if I don’t want to; it’s soulless information transfer, but I’m subscribed for percentages.

“Great writing!” says a commenter, author of a Substack called ‘AI Pathways’.

Elsewhere on Substack, the post ‘How to Spot AI Hallucinations Like a Reference Librarian’ was recently in my Notes feed. I read it and was not overly impressed – a couple of mildly interesting possible insights about what hallucinated references look like, against a background that does not feel relevant to me. Usually the paid version of ChatGPT 5.x will give me links that I can click on to verify, not requiring any special skills.

Then the other day, Saloni Dattani (author of the excellent Scientific Discovery Substack) tweeted on Bluesky, “Most confusing genre these days - viral blogposts critical of AI which are themselves clearly written by AI.” She mentioned the ‘Like a Reference Librarian’ post as an example in a follow-up.

I hadn’t noticed it was AI! But Pangram agrees with Saloni. I need to train myself to recognise the style.

Still I must disagree that it is confusing for this AI-critical blog post to be AI-written: the blog post said that AI was bad at reliably generating references; it did not say that AI was bad at generating blog posts. There are lots of successful AI-written blog posts!

A recent post on Medium was shared on a Teams channel at work.

AI Took My Friend’s Job — But Tripled His Salary 6 Months Later (Here’s What Nobody’s Telling You)

Close tab. I didn’t read this garbage when it was artisanally hand-crafted for LinkedIn virality, and I’m not going to start now. But let me get some more practice at picking those rapid-fire short sentences:

“Dude. ChatGPT just wrote better code than me in 30 seconds. Am I screwed?”

Jake’s a software engineer at a mid-sized tech company. Makes $140K. Has a mortgage. Two kids. He’d just spent three weeks on a feature that Claude finished in minutes.

I get it. The headlines are terrifying. Every week there’s a new story about AI “coming for your job.” Anthropic’s CEO warned that AI could replace half of all entry-level office jobs within five years. Goldman Sachs economists predict 6–7% of the US workforce could be displaced.

I’m sorry. I’m not even on LinkedIn.

“What If Readers Like A.I.-Generated Fiction?” asks the headline of a recent New Yorker piece by Vauhini Vara. ChatGPT-4o was prompted to write passages in the style of various literary authors; lay readers tended to prefer the AI output to the human output, but graduate students in creative writing preferred the humans.

After 4o was fine-tuned on each author’s work though, even the writing students started preferring the AI passages. Vara presents four passages – two hers from an unpublished novel, two from 4o fine-tuned on her earlier work – and asks us to guess the creator of each. I scored zero out of four, no better or worse than “a professor of literary arts at Brown University” who’s been reading Vara’s writing since they were at college together a couple of decades ago.

Insofar as AI writing annoys me because of its annoying voice, the possibility of more varied and human-like stylistic output hints at a more pleasant online reading future. But I think that – when I spot it! – my reactions to LLM writing stem mostly from whether I want there to be human intention behind a text, and whether I feel that such intention has been communicated. If I sense that someone’s sending me LLM text outside of a work environment, then the bar for the intention feeling communicated will be very high, even if the words flow nicely.

As for my own writing, I never use more than a very rare isolated word or phrase from an LLM; certainly never a whole sentence. ChatGPT is for computer code and background research, and only occasionally for a reverse dictionary lookup, “What’s the word for…?” When I write about Fourier transforms, it is my authentically expressed self, full of soul.