Don't Use AI to Write

Reading through docs and doing research: of course. They're like much better search engines in that regard. It would be silly not to use them. Obviously, you need to review their search results against their sources.

Structuring arguments: bad idea. Here you're having the model do your thinking for you. The problems that occur with having it write text recur at a macro level; too-obvious patterns. Readers just want to know what the prompt was.

Writing the prose: obviously not.

Proofreading: I think this works pretty great, as long as you have a structure to how you're doing it. "How do I make this better or sharper": bad, you're having the model structure your arguments for you. But for a predefined rubric of stylistic things --- writing you know how to do, that's just tedious to check manually? Hard to see a downside.

I didn't see in the article where he recommended AI to structure an argument? He did suggest it can be helpful for structuring data.

Giving the compiling evidence of how addictive generative AI is actually, it actually makes perfect sense that people are looking for strategies to excuse their AI-usage. I think no matter the strategy (and given AI is actually as addictive as evidence are suggesting), chances are the usage patterns will all converge to “Just use AI dummy“.

This reminds me a lot of smokers who all have a great strategy beat the nicotine addiction. Only smoke on the weekends when out partying with friends, only smoke to use the extra sense of energy, only smoke to suppress my appetite, go to the gym 5 hours a week to offset the negative health impacts of smoking etc. etc.

In reality most smokers will converge to about a pack a day.

The likes of Sir Walter Scott, Peter Ackroyd, Stephen King, Tim Ferris, Robert Greene, Tucker Max, Catherine Asaro, Neil Gaiman, Charlaine Harris, Nancy Holder etc... all rely heavily on research assistants in various capacities including proofreading.

Honest question - where does an Author like the above place in your authenticity framework?

there's a growing contingent of anti-ai extremists shouting down anything that looks or feels like the thing they hate: actual slop. but there's a missing understanding of where slop really comes from - "you used ai instead of thinking" is just another guess, and mostly wrong since you literally cannot use ai without thinking - the prompts don't write themselves.

slop has been around a lot longer than ai - the contingent needs to realize what it is they hate. instead of "don't use ai to write" it's "don't use ai to write slop" or better: "don't write slop"

The statement should be: never use AI to draft text that is to be presented as your own authorial voice.

Never, never, do that!

I draft all my own text that anyone will ever read as “that’s James Bach talking to me.” I want no one to be able to credibly question that.

But of course, writing is more than merely drafting the text. As I draft, I write about ideas that other people or AI may have helped me come to.

Structuring your argument is probably something you should do, not an AI. This is the core of what you're trying to convey, and is exactly the type of thing you should be thinking through.

Of course, this is different than having the argument ready and using an AI to make it read a little better, which at that point, is just details.

My take is slightly different.

- Brainstorming with AI is fine, but it takes skill to use it as food for thought, rather than as a source of answers.

- Do NOT structure your blog post with AI. Once the structure is there, it is almost impossible to humanize it.

- Do NOT use it for writing a draft. Same as above. A few times I did it, and ended up rewriting from scratch. Unsloppification is hard.

- AI for proofreading is wonderful. But I always prompt with "only fix grammar mistakes, do not change the style" and actually look at all changes. If not kept in check, it too eagerly homogenizes style.

> Of course if you wanted to make a better impression, you'd type it on a typewriter, or better yet, hand-calligraph your writing with a fountain pen.

Why do people feel the need to invent the past that never existed?

>Once upon a time the word processor was invented, and it made it so much easier to arrange your words on paper.

It’s easier to rearrange and rethink and tinker. Hard to imagine now but it was more efficient in the days when you would have to pencil out a memo and have someone type it up, you would make corrections and the final copy would be done in a couple passes. This has been studied, efficiency has actually gone down, in favor of immediacy and direct control.

What it frees you from is organizing your thoughts ahead of time, which takes practice and discipline. Using an LLM is just taking that further, you can avoid organizing your thoughts and ideas a bit more.

You nailed it on what I've always wanted to say.

If someone uses LLM to help wording their idea in a way that's easier to understand, I am fine with it as long as they don't copy-paste it verbatim into their blogpost.

I am getting a little ranty here, but the usual LLM-isms like "not this, but that" or "that's the real footnote here" simply add zero value to the text, and I can't imagine anyone who has proofread the LLM-generated text even once will opt to keep those sloppy paragraphs. If I see them, I can't help but to think that the author had oneshotted the text. And it strongly suggests to me that the author has put zero effort into their work.

Well said, the issue being, the form factor of many AI products disincentivises this behaviour. To give an example, using a CLI tool with a frontier model to write a blog (actually fairly common), you have to re-open it in another editor, potentially reformat it or convert MD>DOCX, etc etc. There's a lot of noise and mess, and people naturally tend towards "not bother" and 1-shot some "amazing" prompt (because the value of their elbow grease is disproportionately low, unless they actually have something to say- many do not).

> Start a chapter in 1PP and just have it rewritten to 3PP

I write differently with pen and paper than I do in a word processor. Sometimes one is better than the other. (Word processors are generally better, for me, at editing. One of my favorite uses for LLMs in this way is taking a picture of some text and sketches I’ve made and OCR’ing and rendering them into editable documents.)

Switching from first to third person is in part a grammatical shift. But it’s also a real change in perspective. If you’re only making the motions but not trying to see where the new perspective opens up (or closes) views into what you’re saying, you’re doing copy editing more than writing.

At least right now, this hollowness is part of what makes AI-generated text stand out.

We have copious evidence for behaviors like note taking and writing creates deeper thinking and retention that passively watching / reading

https://www.scientificamerican.com/article/why-writing-by-ha...

> Okay, but reading also requires substantial thinking.

I don't think anyone would argue reading a book takes as much thinking as writing one…

I've been watching chess videos for many years but my own chess playing is just recreational.

Writing/programming as a specialization of just going through the damn thing is actually sacrosanct IMO. Spectatorship as a generalization of reading is second-class. You do become a worse programmer as you move to "higher order thinking", it just matters less in a world where LLMs exist.

> And reading and adjusting AI output requires substantial thinking.

It does, but less of that thinking translates to skills transferred. What amount of the creativity and magic of Tolkien transferred to his editor? I'm sure it took substantial thinking to edit Tolkien's work.

AI can’t write for shit because it doesn’t have your experience for context and it doesn’t know what you actually want to say.

Hell, most people barely have a rough idea of what they are trying to say when they sit down to write, and they have all of the context.

Writing and coding do have many similarities, at a high level. But the things that make AI good at coding don’t really apply to writing. The analogy doesn’t work well here.

This is a neat workflow. Incentivises you to proactively think about your writing, even as you cut it.

"The marks are suggestions to weigh, not corrections."

"that could go without changing what you mean. Your text stays exactly as you wrote it"

"Marking works when this page is open in Claude and you are signed in. It isn't available in this view"

The last one may be why I even gained the ire to interrogate the copy of the page. I was signed in g'damnit!

edit: I've seen you get quite unfairly treated in the comments section of this place, you always seemed so kind in response. Thanks for all the writing over the years!

If the point of the code is to be read, maybe.

I don't write code for my colleague's pleasure, I do it to solve a problem as fast and efficiently as possible.

> Writing code requires substantial amount of thinking as well. Shouldn't we write code without AI as well?

You absolutely should. Using an LLM to write your code for you is not a good idea. It produces bad code, and it also robs you of understanding which will be valuable down the line.

I don't know what this is supposed to mean. An LLM can certainly understand the structure of your writing. It can accurately diagram sentences, identify the grammatical agent, note transitions between sentences and paragraphs, determine the position of modifiers, evaluate topic continuity, spot hedges and qualifiers, count word and construction use and identify overuse, and evaluate openings and stress positions of sentences.

this SPECIALLY for arabic the shit I see everyday

I want LLM translations to die.

We used to have machine translation that worked well enough for readers to know the overall meaning most of the time, but when it failed, it failed spectacularly and the reader would know not to trust the output.

Now with LLM translations we can’t trust anything. Even simple translations which used to work perfectly now shoves extra LLM-isms into the translation making the prose ugly.

Proofreading yes, automatic "corrections" no

> Does coding not require thinking or what?

Code is deterministic, and language is not. If I write:

total = 0 for x in [1, 2, 3]: total += x

…nobody is going to argue with me that “total” is anything but 7.

But language is different. What you perceive and what I perceive from the written word can be completely different (e.g https://youtube.com/watch?v=naleynXS7yo&ra=m )

> Fixing grammer mistakes wont change the soul.

Pangram disagrees, if your grammar was fixed by AI it's "100% Written by AI"

What is a soul, how do you measure and quantify it? If you can’t, how would you know whether the grammar mistakes change the soul? I’ve heard people say they keep the grammar mistakes in to prove the existence of a soul, so presumably correcting them would alter the soul in some measurable way.