The social contract of writing
161 points by joladev
161 points by joladev
This quote from the Oxide RFD
Finally, LLM-generated prose undermines a social contract of sorts: absent LLMs, it is presumed that of the reader and the writer, it is the writer that has undertaken the greater intellectual exertion.
is a less pithy, less quotable but much more explicit phrasing of the common "ai;dr" meme:
If you couldn't be bothered to write it, I can't be bothered to read it.
What I’m missing is a sense of professionalism in the writers who integrate LLMs into their work. To me, all these LLM smells shout loudly “I am a hack who can’t be bothered to do my work”. For software, you could (but I won’t) argue that consistent, predictable and boring low-variance stuff is actually good, but the opposite is true for writing (mostly).
Wonder what part of it is desperation and trying to keep up when everyone else is mass producing content, and what part is giving in to the temptation of just offloading all the work and responsibility.
Offloading responsibility is just wishful thinking, and at least some of these hacks will find out the hard way when editors and publishers start waking up to the fact their readership is bailing on them.
For software, you could (but I won’t) argue that consistent, predictable and boring low-variance stuff is actually good, but the opposite is true for writing (mostly).
I strongly disagree. I think we need more consistent, predictable, and boring low-variance writing. Most guides to writing prioritize style over clarity and truth.
I write this not to defend AI writing (I don't use it for writing, and don't like the smell of it). I just wish more human writers had the courage to be boring.
Clarity and truth, eh? Not exactly what LLMs are known for.
Besides that, most good writing guides teach you brevity. Ruthless cutting of unnecessary verbiage to the point where students ask how their voice can shine though. That’s also the opposite of what LLMs offer. There it’s one sentence in, many bland and repetitive paragraphs out.
You may note I said I wasn't defending AI writing and don't like it, so it doesn't make sense to tell me that LLMs don't write in a way that matches my values.
As for brevity, it's ok, perhaps a little overrated. Being clear often means taking longer to draw attention to and resolve potential ambiguities. I think popular writing advice focuses on superficial smoothness, not actual clarity of ideas.
We were discussing smoothness of writing in the context of LLMs, so I wanted to point out that's the wrong kind of smoothness. But point taken.
Brevity is so hammered on because people need to unlearn their deeply ingrained text padding habits from high school (as in "write a 5000-word essay about your hamster"). But you have a point here too, sometimes it can be valuable to spend a bit longer to explain things in multiple ways and even a modicum of repetition can be helpful to drive home a point or to jog the reader's memory. Other than that, I think especially in the current age of short attention spans, many professional writers for popular publications are still quite over-indulgent with their meandering writing.
This is a good observation. I think stressing brevity and focus is a good antidote to school habits. My opinion is really directed at mature writers who've already overcome those particular bad habits.
brevity
Just like with code, one often has to start out with a long, complicated mess, and then reword and reshape it into something more concise, chipping away at the unnecessary stuff until only the essentials are left. As the old quote goes, I would have written a shorter letter, but I did not have the time.
I just wish more human writers had the courage to be boring.
What do you mean by boring? Even to come to a relatively boring conclusion, it takes significant amounts of work to write content that is worth reading...and boring? If you use the term "boring" in the sense I understand, I think of the endless "hamburger essays" on world war 2 you read from children.
Most writing aims to come to some conclusion or at least add signal to something. Boring would be no value-add?
Of course I don't mean that, but I can say more.
By boring, I mean taking the time to be precise, even if it means providing technical details. I mean carefully thinking about every claim you make and ensuring that it is defensible, even if you must write "defensively" or add qualifications that don't flow as nicely as an unqualified assertion. I mean taking time to quantify your uncertainty, and describe the holes in your evidence. I mean prioritizing the content of what you write above the style.
You're correct that to some readers, my ideal won't be boring, and they'll prefer it. My sense is that to most, the kind of writing I prefer will be boring, and it will lose readers compared to more easygoing work.
that’s a very narrow definition of what “good writing” is that does not include most of what most people think is good writing. There’s not one “good writing” style, since different writing has different goals and different audiences.
For software, you could (but I won’t) argue that consistent, predictable and boring low-variance stuff is actually good, but the opposite is true for writing (mostly).
What if you anticipate the most likely readers are LLMs themselves?
So in fact it doesn’t matter whether the content is good, or even that the writing is fine, it’s the action of using an LLM to write instead of writing yourself. The very fact that the author reduced the effort they made to product the content is a violation of the social contract.
Same with pull requests. Most of the effort used to be in thinking through the problem and writing the code. Reviewing can also be a lot of work, but at least it used to be balanced, or when it was unbalanced it was because reviewing was at the same time onboarding/training someone.
If you care about the quality of your project, the equation has changed completely. People who couldn't care less about the internals of your project, can dump PRs on you orders of magnitudes faster than you can review them.
It's exhausting. At least code review has the benefit that you can use vouch or some system like it to disallow PRs from new contributors. But with blog posts, etc. you always end up second guessing if it was written by someone who actually understood or cared or whether you are bumping into another wall of slop. At least the slop is still easily recognizable, that will probably change as slop producers will instruct the model to block some of the typical patterns.
The only slop I celebrate is LinkedIn getting flooded by slop comments.
The only slop I celebrate is LinkedIn getting flooded by slop comments.
Which actually had no effect at all the average post quality. (\s. Self-aggrandizing shit before, self-aggrandizing shit after. Plus cartoon AI image, I guess.)
blog posts, etc
I've started throwing URLs into the Wayback machine in order to check when they were first created: anything before 2022 is substantially less untrustworthy by default. People I know are putting before:2022 in image searches.
I would be careful with that. Content farms have existed for far longer than that.
This is what tells me that the outrage at AI-generated texts is about more than just the text. Didn't anyone notice the SEO slop that came before? The internet was already flooded with garbage. It has been for many, many years.
Content farms have existed for far longer than that.
In my experience, most content farms these days are tuned to artificially future-dating their content, not back-dating it. "Evergreen" by virtue of updating the dates to always look recent. It would be interesting if people intentionally started trying to backdate to pre-LLM content, but I don't think both can simultaneously be true, and one of them is significantly more valuable (looking updated) than the other.
Content farms have existed for far longer than that.
Yeah and we didn't like them either.
It's not the only signal but it is unfortunately a good one.
I think people honestly where less critical of marketing garbage blogs (tinly veiled company promotions, blogs only serving to promote one selves) before. In the past I have often been amazed how such content was highly upvoted on websites like this.
What has changed is that LLM prose has given people a fairly easy tell without critically having to look at the content.
Having said that, I also do think there is more of it now due to LLMs.
For me, the most important social contract when reading blog posts, is that the author will use the experience gained to improve their writing going forward. And that's also broken by usage of LLMs.
I can have a debate about using LLMs to write code, but LLM usage when writing blog posts especially, is just revolting.
This put to words thoughts and feelings that I've had for a while. Especially the desire for the imperfection of a humans touch. I want it from others, and I also want my humanity to be present in my own work.
I'm a notoriously bad speller, its something I get a lot of crap about from my friends and colleges. However I've seriously considered leaving in some of my mistakes while writing blog posts just to show I'm still human..
maybe that's going two far though
I'm a terrible speller, as well, and have certainly thought about this for my own blog.
One thing I've noticed in the Age of AI is the growing untrustworthiness of information in general, to the point where a number of people who have trouble telling the difference between human content and generative AI content just call everything "AI slop," without any real regard about how that might impact the human who actually did write the work themselves.
As an aspiring writer, I constantly worry about being accused of using an LLM to do my work for me. I'm new, trying to find my writing voice, so some experimentation is expected which may change how I "sound" to some readers, now and then. Adding the fear of being accused for letting an LLM do my work for me is just piling on top of that anxiety. This is something we don't talk about enough, the psychological impact on actual writers who are out there struggling to get going and improve their skill when it only takes one public comment to plant the seed that some writer is secretly using an AI. It's a bit maddening, to be honest, and most days it makes me wonder why I try.
What keeps me going is that I do it for me. I don't care for metrics and I'm not trying to sell anything except myself (hoping that someone hires me for a writing gig, eventually). My blog is, essentially, me trying to write content that I want to read, while at the same time giving me a space to express myself in an unobtrusive inconsequential way where readers can choose to read or just move on. I enjoy writing too much to give up, but with the amount of times I have been fooled by LLM content despite being aware of it, the future of writing is looking rather bleak, to me.
Perhaps you are right, and leaving the imperfections in would lend itself as part of the art as it evolves to account for this cheapened landscape we seem to be moving toward.
Well done; it's not going too ( ; ) ) far - I got stuck on it, then re-read your comment to re-appreciate it, then re-read your comment again to check for other typos :D
I'm in the middle of writing a (fairly complex) technical blog post right now, where the prose is entirely written by hand. It took a couple attempts to nail down what I wanted to say, but I think I have a pretty good structure now.
Where I did use Claude for assistance is turning some textual descriptions into Mermaid diagrams. I appreciate all the work that's gone into Mermaid but I find some of the syntax (particularly once you move off of the standard structures like flowcharts and sequence diagrams) to be a bit too esoteric.
(I work at Oxide and had some input into RFD 576.)
Have a look at d2. I find it quite pleasant to read and write alike: https://d2lang.com/
I've started using d2 in my blog. Much simpler to write. I take the markdown fenced code, give it to d2, and inject the svg directly, no js on the post at the end. I'm quite happy with it
Personally I think that is entirely fine. I also think that "writing" is such a broad spectrum that unless you are principally against LLMs (which I respect) you can't make blanket statements about the usage of them.
Writing for the art of writing, yeah I 100% agree that people shouldn't use LLMs. But once you move to "functional" writing I personally think it becomes less clear cut. As you demonstrated for diagrams or getting started with diagrams in my opinion is a fine use case. Other use cases might be prompting the LLM for an overal structure of a document. Sometimes when documenting something technical I find it helpful to give an outline of what I want to document and prompt for suggestions on chapters, paragraphs, etc. I most likely will find myself reordering sections, dropping others but as a initial scaffolding tool it can be useful.
As far as the contents of those sections go, I rather write them myself. Because brevity and clear language is not something LLMs are all that good at.
I enjoy writing and I’ve been doing it all my life [...]. I’m going to try to make this more of a routine thing now. It feels meaningful. Worth doing.
I feel exactly the same, though I have two fears:
For #2, I don't really see a way to kill the many Goliaths. For #1, I hope that we'll develop some sort of "proof of human work" that can attest that a piece of content actually has a sufficient amount of human attention behind it. I thought about streaming on Twitch while I write a post, but it's very impractical, and I'd actually be "ashamed" of making my writing process public, with all the stupid ideas, the realizations that what I'm writing is completely wrong, the endless rephrasings to make it match my taste, etc.
I don't have any other ideas for a proof, though.
What we can also do is to put high value on human writers and human writing, on truth and the institutions that work to spread truth. Refer to and edit encyclopedias, support archives, use libraries, read and subscribe to good reporters, local news papers, investigative journalists, show support for independent jurors, researchers and scientists, and speak up against those who spread lies.
We’ve put a premier on good grammar, vast vocabulary, good use of expressions and metaphors, and general text composition. LLMs do all of that just fine.
No, they don't. I'm assuming that we're talking about post-trained chatbot products, since raw LLMs routinely have terrible grammar, degenerate word choice, a lack of abstraction, and overall roughly the quality of an AIM conversation at best. Even when considering ChatGPT and Claude, their grammar is incongruent with reading their words aloud (which is part of why their text-to-speech products routinely stutter, leave awkward pauses, and pack clauses incorrectly); their vocabulary is a sycophantic veneer, their metaphors are insipid when not borrowed from humans, and their overall command of text is pathetic compared to a 1990s text editor. They can't spell or do basic arithmetic; they can occasionally simulate it (discussed previously, on Lobsters) but they don't have the genuine learned capability because they don't manipulate letters or numerals like humans.
Sure, they just won’t stop repeating the same patterns, the expressions are tired, the metaphors are a bit out there, and they’ve given the em-dash a bad name.
So, not "just fine." The chatbots are bad writers and we should stop minimizing the badness.
I broadly agree with the content here, AI produced prose has a 'writing smell' about it that seems to sap the feeling from the writing.
I've asked LLMs to review writing before to try and provide rebuttals to arguments I may have overlooked. I would say I only act on one in three suggestions, many end up being generic or not really in line with the point I am trying to make. In that sense AI tools have been handy to broaden my perspective when I've written a post with a specific argument or focus in mind.
I blanket refuse to let it make changes to a draft though.
What if you use the LLM to redouble your rigor (deep research passes, etc), rather than replace it?
I imagine there are some use cases out there for using LLMs to write English. I just don't seem to run into any of them. For everything that I write, I'm trying to communicate ideas to other people, and that means I need to tailor the message to them specifically.
For example, writing an email to a coworker or group of coworkers, or writing documentation, or writing a commit message, or writing a PR description. In all of those, I don't want to go quickly, I want to make sure that I'm understood. LLMs don't help with that.
Don't even get me started about blogging with LLMs. I don't see the purpose there at all.