Pangram Blog

I was surprised not to see any discussion on whether the author used AI to help write this post. As many people say, writing is thinking.

I started getting that "I'm reading another AI-written blog post" feeling around 1/3 of the way through, but I don't consider myself super calibrated on this.

Pangram seems pretty confident it's AI (https://www.pangram.com/history/e9f6eb77-86f9-46d0-a6c1-e57ce418c6d7). But I know these tools aren't perfect. I'd love to hear from the author what their process was in writing this piece!

Related question (I'm trying to work this out for myself):

If you believe using AI to write an email or blog post for you isn't okay, but using AI to write code for you is... what's the difference?

Right now my instinct is something like:

- Code can be verifiably correct (especially w/ good tests) so it's less of a purely-creative act than writing.

- But always, always double-check the tests!

- I still wouldn't submit a PR where I can't vouch for every line of code.

- AI-written documentation and specs are mostly still bad and should be looked down upon. But mostly because the quality, at least today, is poor. (Lots of duplication, lack of a clear understanding of the reader's intent and needs, no thoughtful curation, etc.)

- Be psychologically ready to update these priors as models change.

I'd love to hear from anyone who's thought more about this.

Look, I appreciate the article, I believe these are important news worth discussing, but was there really no way to write it without LLMs?

I'm sure there are enough proponents of "what's important is the content, not who and how exactly typed it," but my brain has developed something like ad banner blindness but for LLM writing - it just shuts off as soon as it sees "this distinction matters" or "does not automatically guarantee." This is not to mention that LLM generated texts are just really confusing and difficult to read, take for example this sentence:

> Linux is not explicitly banned. Desktop Linux users could access a website and scan a QR code using a supported mobile wallet.

At least two people here agreed it's a really weird way to put this, but at this point I'm honestly inclined to believe that this is just the product of AI putting together some words that kind of sound on-topic even though they are actually meaningless in the context (a.k.a. AI slop), and not someone's actual thought.
---
100% AI pangram: https://www.pangram.com/history/8c60b6ec-cd89-41fb-b8c8-abc096973f47

This product does not even have a plausible theory of how it could work.

LLM-generated text does not carry a watermark or other identifying marks. The "theory" is that an LLM trained on human writing, to mimic human writing, can be distinguished from actual human writing in under 100 words.

Notably the first diagram on the research overview page (https://www.pangram.com/research/how-it-works) shows feedback for "misclassified human examples." This is a category error; Pangram will not find out when it has misclassified text in the wild, except in rare cases. Only the "licensed human-written text" in its training data can be used as feedback.

Scams like Pangram also cause real harms, mostly because laypeople do not understand that what is being offered is not possible. Pangram advertises 99.98% accuracy, and they pitch it as a tool for teachers and universities. Translated: if a college like University of Alabama rolled this out, you could expect ~40 students to have their lives upended by this snake oil, every year. And this is the best case, using the number on Pangram's homepage.

I'm not affiliated with it in any way, but you might want to check out the leading model https://arxiv.org/abs/2402.14873 and the mixed authorship detection https://www.pangram.com/blog/pangram-3-0-technical. LLM detection has come a long way since the first GPTZero.

Human heuristics - I've prompted millions of tokens across every frontier model iteration for all manner of writing styles and purposes - also helps greatly.

Concerning to me are long-time posters who (perhaps unknowingly) advance the decline of this human community by encouraging the people breaking HN guidelines. Perhaps spending a few hours on Moltbook might help develop such a heuristic, since "someone who knows how to write" is just a Claude model with a link to the blogpost.

https://hn.algolia.com/?dateRange=all&page=0&prefix=false&query=by%3Adang%20%22own%20voice%22&sort=byDate&type=comment

Thanks for your comment!

https://www.pangram.com/history/f119a8ee-7e60-442a-8d0c-3fa0fe96d4b7/

I show this more of as a discussion point than as a definitive answer. There is more research on this tool than others: https://www.nature.com/articles/d41586-025-02936-6

And I think it's interesting that it flags it as confidently AI generated. I also got a whiff of AI and I'm never sure how to take confirmation bias from AI detectors - though that said, I've gotten a whiff of AI before and had this detector say it's confidently human.

Reading this I kept waiting for the... point? I feel like the whole thing was more like saying "UDP is like a convertible because you can strap a tarp on top when it rains". Like... sure? But that tarp is going to be crappy compared to a real roof. And the idea that that tiny layer is "the best of both worlds" is frankly ridiculous to me.

It's very clear to me on its face that it's AI, but not "obvious as the sky is blue" others seem to be implying. I would dislike the writing style even if it weren't AI.

For the record, an AI detector that appears to have put work into reliability and that I trust very much from my own testing, Pangram (https://www.pangram.com), says this is 100% AI generated. I've used it plenty before when experimenting with AI-collab writing, both fiction and non-fiction, and it's frustratingly accurate in identifying what is and isn't my contribution. I have since largely given up trying to do AI-collab writing, because no matter how nice the writing looks in the moment, it always reeks when read closely, or on later days.