Almost Never Use AI to Write Anything Substantive

From TFA: "Eric Schwitzgebel writes that . . . There’s a huge cognitive difference between nodding along while reading something and actually productively generating a text. Two reasons: First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo. Second, as I suggested above, I doubt that human beings, even experts, have a good sense of all the factors that shape word choice -- everything they’re being sensitive to. You would have phrased it slightly differently, and even if you don’t know that, or why, a different signal is sent and received."

This is the best articulation I've seen of why simply reviewing and copy-editing does not provide remotely the same value as writing from scratch. I spent a considerable amount of time over the past two weeks reviewing and improving a work document that was the output of an LLM. Given the number of people involved and the final level of effort, I'm firmly convinced that writing it manually would have been faster and resulted in a higher quality product. Getting the wording right matters.

Use AI to write things for you to read that you wish someone else had written. ‘Give me a summary of the research on this topic’; ‘Write a report on this data to help me make a decision’. ‘Take this transcript of a meeting and write me the email it could have been’.

Don’t use AI to write things that you are producing for someone else to consume.

Instead of asking AI to write something substansive for you, ask it to critique your writing. Then use your own judgement as to which bits of advice to follow and which to ignore.

The LLM will always give you a full rewrite — don't use it. It always does too much, and persuading it to tone it down is a constant battle.

"AI writing is vague and wrong in hard-to-notice ways" This one felt to me, at work I lost a lot of time re-reading and correcting AI prose that was launched to "summarize" a collective white paper, and where subtlety and nuance disappeared in ways that harmed the initial ideas/arguments I was writing about, but in sneaky ways.

If I may propose a broader rule, that I think this post presents a special case of;

Only ever use AI to make yourself think harder, and more.

I don't know what everyone else is doing but I've certainly been writing a lot more often because of my interactions with AI. Maybe some people used to write more but have been leaning on AI so they can write less but I don't think that's universally the case.

I agree that being lazy about writing and simply using a short prompt or list is detrimental if you are replacing your own output but you don't have to do that and you don't have to accept any of the output either. The best chats I've had usually start with a large amount of my own writing up front, thinking through the idea, listing several alternative ideas, asking a lot of questions, jotting down related topics, etc. and then reading the resulting output and critiquing it, asking for clarification, doing my own research on it, even just discussing it with the AI and doing this for a number of turns until I feel like I've exhausted the topic in the chat. I then often take what I've shaped in my own mind from that process and write something myself, either that or take the best parts of the output and edit, rearrange, reinterpret, and add to it in order to produce something.

I guess it all comes down to whether you are actively engaging with the material, regardless if that material is the result of a Google search, pulled from a book or generated by AI. If you just read it or copy paste it and don't engage with it and think about it then it doesn't do much good.

This advice is too coarse. I believe LLMs have made me a better writer of technical reports. I have a keen ear for language, and I can usually tell when a sentence or a paragraph is orbiting around a point rather than landing on it firmly. Some writers insist that unclear prose is dispositive proof of unclear ideas -- a sign the writer has not thought things through. My experience is different: I often have a clear idea in mind that I cannot quite find the right idiom to express. Sometimes I can find it if I keep revising, but staring at a sentence too long can entangle you in a particular linguistic frame that can be hard to escape.

LLMs are helpful because their idiom-list is vast and they are indefatigable and infinitely patient. You can keep iterating on a sentence and it will keep giving you fresh takes. Eventually, whether through careful steering or brute-force iteration, it will come up with something that has the right resonance. Like a word on the tip of your tongue, you recognize it when you hear it.

For me, the end result of this writing-process is often a document where much of the language first came from the agent, but the voice is recognizably mine, and I feel a clear sense of authorship. It is still my work because of the microscopic attention I paid to every word. The marble is the agent's, but the chisel and mallet are in my hands.

I can't help but think that some of the arguments here apply equally well to writing _code_ as they to to other substantive writings.

Writing code is thinking, AI code is often vague and wrong in hard-to-notice ways, and the (human) reader of code is the one that pays the cost for this later.

The cost/benefit analysis may still work out differently for code though...

I don't find AI helpful for truly constructive or creative prose, it devolves to the median output very quickly and it's almost painful to hear in my inner voice when I read out an AI-ified idea. The "not this but that" formulation is especially bad or cloying attempt at pretending to refine an idea or lightly correct something to start from a position of authority, very annoying.

That said, I have a lot of situations at work where I am asked to simplify something I am an expert in or convey something for a different audience, particularly as I prepare for presentations and I do find it helpful in helping me step down my writing or work. YMMV.

I also skip words a lot and can miss that even after 2-3 editorial passes and it's very helpful at that vs. say normal spellcheck.

> I think so because (1) the writing process is an essential part of the thinking process, (2) AI writing is vague and wrong in hard-to-notice ways, and (3) writing with AI (and not labeling it as such) is rude and misleading. I’ll explain these points in more detail below, but first, a few throat clearings.

1. Agree writing process is essential, but just getting your thoughts down is the start.

2. Human writing is very often vague and wrong in hard to notice ways, but I would say it's more likely to be wrong in easy to notice ways, which is ... better?

3. I think it's rude to not use the best tools to convey the message most clearly and most respectfully of my time. I often use AI to help me write emails and nearly every time it provides more concise well structured versions of what I have to say. 99% of the time I'm just trying to convey a message clearly, that's all. When I write my draft there is often words I can delete or phrases that I can shorten. It's hard to spot but when you have AI point them out it's obvious. This is the role of an editor (check out Stephen King book On Writing which makes this point). Unfortunately I don't have an editor, but I have AI that helps trim the fat, get to the point and save my readers time.

If these tools are available to you and you don't use them, then I think that would be disrespectful. I don't take any solace thinking how much time and effort went into what someone wrote me. If I could save them time and have AI write or edit, and save me time reading it since it would be more concise and better structured, it's a win-win.

People were self-conscious about their writing (quality, speed) before LLMs. AI seems like a way out of that weakness. That's just the nature of a tool: it seems to extend our ability. People are going to use it if they don't have a better option.

The only cure: be the person people can turn to instead of the machine, whether that's helping them write or helping them become better writers. And, yeah, you'll have to do that cheaper than the AI (netted against the social cost of being the guy/gal who got caught slopping).

Anti-AI people want things to go back to the way they were before; I want a 600-pound bench. Neither are happening without expending a massive amount of effort and dealing with no small amount of sociopathic socieconomic upheaval. Decide what you want.

I think about this a lot. I started my teenage-to-young-adult life in the literary world as a poet who loved programming, and made a living as a ghostwriter. At some point, I fell in love with mathematics and wound up in ML in research/engineering for the last 8 years or so. So, I've thought a lot about writing and ML and their intersection.

I think one of the underappreciated things about writing "substantive" work is that the work you see at the end isn't the first attempt. And I don't mean the first draft of the piece. I mean that almost always, writers iterate on the same topic many times, either with complete published pieces or abandoned drafts or even just conversations and sessions of unproductive daydreaming. It's a cliche that your best work typically also comes out fastest, but it's not because of divine inspiration, it's because you've whittled the big idea you actually care about down so much in your mind that you instinctively know exactly how to write it.

My experience has been that for people who don't work this way or don't write a lot, LLMs can give them this incredible feeling of leaping straight from inkling to "substantive" writing. And because they haven't built up those muscles or "taste", they don't immediately recognize that it's imprecise and hard to follow.

That's not to say they're not brilliant in their own right, just that they haven't spent a lot of time on this particular thing. Sort of like a very gifted programmer who doesn't have a ton of experience yet (speaking as someone who is gifted at nothing and frequently has to do things they're inexperienced at).

So I don't think the problem is that LLMs just write bad. It's that LLMs are so wonderfully powerful that they allow you to confidently leap forward to create something that is a little beyond your experience.

And that's why I have different reactions to heavily AI generated writing. When it feels like marketing at scale, it grosses me out. But when I feel like it's just someone who is excited to write an idea and maybe doesn't have a lot of experience doing it, I'm not judgemental. My hope is that it makes them more excited about writing, and that trying to make their next piece better will lead them inevitably to start thinking about where the last piece fell short. And if there's some slop along the way, eh, I'm not compelled to read it.

Communication between two people is hard enough. Involving another actor as an intermediary rarely helps. It quickly becomes a game of telephone. Doubly so if the intermediary is an LLM.

So, I think I agree with all of this, but want to use this post to host my take on AI writing, a take I don't think I can get across on the big thread here on my post about AI writing:

https://news.ycombinator.com/item?id=49747070

I think people are stuck in the idea that the way you "use AI to write" is to let it generate alternative words for your thoughts. That's a terrible way to write. I advocate for a rule: "any word an LLM suggests to you is disqualified, even if it's better than the word you've chosen". Nobody is vigilant enough to keep LLM word selection from bleaching out their personal style.

But LLMs can do things for language (natch) that deterministic programs can't. Those things are helpful and you should consider taking advantage of them. Here's a short list of ways an LLM can potentially improve your writing:

* It can instantly spot overused words and turns of phrase, or, better still, worthless filler and throat-clearing like "just" and "very" and "it's important to note".

* It can rescue active verbs trapped inside nouns, where sentences are wrapped in zombie verbs like "make" or "reach" or "start" or "have", carbonite-frozen verbs like "decision" or "agreement" or "distortion".

* It can match the subjects of your sentences with actual characters in the action of your story or argument, flagging all the times you accidentally nail the subject of a sentence down to some random part of the scenery. It can check to see whether the new detail each sentence adds (if it adds any at all --- something else it can check) is in the stress position of the sentence.

* It can check your transitions and flag places where the openings of sentences and paragraphs are abrupt. For that matter: it can check the topic sentences of paragraphs and the flow from graf to graf.

* It can check for passive voice, but also note the (many) instances where passive is the right call for what you're trying to say.

The model never gets tired. It generally never forgets the rules. It can instantaneously diagram out a sentence and work out an accurate model of the semantics of your writing. It's doing things you cannot do with a grammar checker.

You're not going to want to act on everything an LLM flags. LLMs have an idiosyncratic sense of style (I keep saying they write every sentence like it's the headline of a magazine article). Some of these quirks of writing will be part of your voice, and you'll need to keep them.

You're not going to want the LLM to give you alternate words and sentences. That way lies Velveeta. And many times, the LLM will flag a usage issue and your response won't be to act on the suggestion, but just to rewrite the sentence or paragraph entirely --- or, better yet: just delete it, which is an awesome feeling.

This is using an LLM as a copyeditor. I've worked with professional copyeditors, and the feeling of working with an LLM copyeditor is comparable, except that the LLM is much more thorough, and I don't feel bad about ignoring it when we disagree. This style of working doesn't allow the model to infect your writing; it's just making you better informed about the words you're choosing. I think more people should try it.

I find it so weird that people use AI to write things. Writing is one of the easiest things you can do. Just put one word after the other.

Who cares? AI gives me lots of text and it's not like it's ever going to go away. Any time in my life in the future when I need text, I can just have AI generate it. Thinking is solved.