LLMs and the dwindling number of people who appreciate and value language
-
-
I just watched David Kippling's (astronomer, professor) latest video. It's timely to this discussion.
Link to video -
-
-
I've been working in publishing for twenty-three years, which surprises even me. In print, I've published novels with two publishing houses and academic books with two. I've published with three audiobook publishers. My short stories and essays add at least seven book publishers to the list and dozens of magazines and newspapers. Since most things pass from the editor who acquired the piece to copyeditors, I've worked with a heckuva lot of people whose job it is to make my prose better. I appreciate these people very much. They keep me from going out into the world with my metaphorical pants down.
These days, some of them are doing a much poorer job of it. I believe that this is because they are not themselves doing it. They are outsourcing the job to algorithms that are deeply flawed. I am trying to hold the line on this, but there isn't anything I can do about changes that happen after text leaves my hands for the last time. This isn't supposed to happen, but it does.
I have a book, Agatha Christie and the Twentieth Century Woman, coming out in January from a major academic press, Bloomsbury Academic. It is based on my doctoral research and it is very important to me. Bloomsbury has published me before, in 2022. The copyediting and proofreading process has degraded markedly since then.
I have spent more time than I like to think about with the final proof of this book. The first round of errors that I've had to fix were related to a five-page sequence of material that I knew would be hard to lay out in a way that was accurate and looked good, so I included notes in the manuscript that were intended to help. The notes were ignored, which resulted in a mess. If a human had looked at the mess or (please God) at my notes, it wouldn't have gone so far, but as best I can tell, somebody fed it to an LLM and sent the resulting disaster back for my review. This would not have happened in 2022.
On the most recent go-round, it would have been logical for me to only review the passages where I'd asked for changes, but I'm neurotic. I read the whole thing yet again. I found that making the above changes had caused the kind of errors that no human or machine should have missed. A stray comma. A missing quotation mark. An extra quotation mark. But the flabbergasting thing was that there was an inexplicable change in a totally different part of the book: I had discussed the work of Mary Evans and, for clarity, added "(no relation to the author)" right in the text. On this pass (and not any of the previous passes), this was changed to "Mary Anna Evans (no relation to the author)." So an error had been introduced and, again, a human being would not have made it, because they would have realized by the parenthetical note what was going on.
This was so out-of-the-blue that I asked for a redline so that I could check each and every change that had been made. How else could I know if something weird had been done somewhere in those 250 pages at the last minute? Well, the redline of the editor's changes doesn't show the Mary/Mary Anna debacle. It's like it appeared out of the ether. The only conclusion to be made is that somebody ran it through an LLM after they finished the job and simply trusted the machine. Like an idiot.
And this wasn't the only snafu. Earlier in the process, my editor's assistant sent me a draft book description that will go out to retailers and reviewers and will be printed on the book jacket and the publisher's website. It was terrible in a way that's hard to describe. I had to rewrite it from front to back. For one thing, it had three sentences in two paragraphs that began with participial phrases that didn't quite modify the subject of the sentence. They were like the classic example--Walking down the street, the trees' leaves were colorful--but weirdly phrased when compared to the years of student errors I've seen. They were so weird and inhuman that they had a chalk-on-the-blackboard aspect. I looked it up and, yes, this is an AI tell. Also, this draft book description used my original title of the book instead of the one the publisher gave it. This was the final giveaway. Somebody had fed the book description I wrote for my proposal three years ago into an LLM, and this is what it spit out.
Within the year, this kind of bastardized language will be ubiquitous. People will stop noticing. Languages change, so is this a problem? Well, it is if the resulting text no longer says what the writer intended to say. And it's really a problem if the resulting text doesn't make any sense and the reader is forced to ignore it or live with the cognitive dissonance.
Languages change, so is this a problem? Well, it is if the resulting text no longer says what the writer intended to say
Why do people not care about language?
I suspect most of the time it is because people feel it's not worth their time.In text/instant messaging conversations, people get straight to the point.
E.g., "lunch at 12:30, meet at Panera?" "can't do 12:30, how about 1?"
No "beautiful language," no real need for AI, most people just type or dictate it, the words say exactly what you mean.Then there are emails ... you have to write more words than are absolutely necessary to say what you really want to say. Instead of simply saying "review the attached report, check numbers on page X, send revision or confirm the numbers are good by COB," most people feel compelled to provide context about about the report, explain why they want the numbers on page X checked, explain why the COB deadline is important, etc. Now that you have to write more words than you believe are necessary, the urge to delegate to AI to generate those extraneous words for you gets a little stronger.
Then there are letters, academic papers, presentation slides, formal reports, books ... you need to put in more and more extraneous words than those necessary to convey what you believe is essential.
Take academic research papers, for example. Instead of just writing your method, results, and conclusions, most journals/conferences expect an "introduction/background" section, an abstract, a conclusion section, maybe a section on "why you think your stuff is new/important" ... Now you have to repeat your key contents a few times (e.g., your key results may have to appear three times -- in the abstract, in the result section, and in the conclusion). The introduction/background section is mostly you recounting someone else's stuff, little to no originality. With expectations for all these extra words that do not speak to what you really want to convey, the urge to delegate to AI gets stronger.
Similar incentives structure for corporate reports/papers/slides. Often lots of boilerplates, lots of fillers because the template says so, because of policy, lots of hedging and disclaimer language to CYA, etc. The urge to delegate to AI gets even stronger.
It just gets worse as the documents get more formal (e.g., regulatory filings).
I don't have a solution for this. I hope people learn to focus their energy on manually drafting/tweaking the parts that are most important, the parts that convey novel/original things, and that people learn to identify and focus these parts when they look at written communications. The rest will continue to exist because systems/institutions require it or because people expect it, and I am resigned to accept those formal/performatory bits to be overtaken by AI slop.
-
I've read recent articles where companies that embraced AI for things like emails and internal reports are bemoaning the time being wasted because AI writing is so verbose. It seems like the algorithms could be tweaked for that, but I really don't know if it's possible.
Absent something in the code that privileges succinctness, an LLM has no incentive not to write long. The text is generated almost instantly. Why waste resources slimming it down for the convenience of human beings? It's more efficient to just let slop be slop.
-
-
-
-
I still haven't had time to look for the article, but your point triggered a question I can't answer. I'm told that the algorithms don't have the kind of memory that we use for writing, They predict the next word, but they don't plan the whole piece and they don't recall what's already been written when they get further into the piece. This means that a fully LLM-written novel will not and cannot have the kind of coherent throughline that a (competently) human-written novel would. (It also calls into question whether it's actually a novel, since a novel has narrative constraints that differentiate it from a book-shaped pile of pages. I have not done a lot of reading in narratology and I haven't read a fully LLM-written novel, so I can't say much more than that.)
Anyway, when I think through how I'd go about writing, say, an email when I've been told to keep it under 200 words, it's a multi-step process. I plan what I'm going to say, I say it, I check the word-count constraint, I check to see if I said what I'd intended, and I edit accordingly. I'm not sure current models do that kind of thing very well. It may be that it's a sufficiently recursive process that such a process would multiply time and energy demands to an extent that would make writing succinctly difficult.
-
-
-
I’ve asked it for one paragraph answers before. I don’t use it for writing, I use it for learning. It’s fantastic as a subject matter expert for learning new things.
But it would certainly understand and follow such instructions like ‘keep it under 350 words’ or ‘no more than 2 pages’.
-
-
I'm reading this thread carefully when I have time to concentrate.
I will say that oral communication among humans, later recorded by symbols, is at the root of all of human society and culture.
The fact that the educational system is being destroyed- by design- should make any rational person pause, IMO.
-
@Bernard I agree with everything you wrote, Bernard.
They have indeed swiped a good chunk of human effort. I think the Anthropic settlement which, although it was a billion-and-a-half dollars, feels like they're throwing us a bone. When it came through, I thought that perhaps all the AI companies were going to at least have to throw us a bone, but I see no sign of it.
The way the settlement was written, books were thrown out because of restrictions on the pub date and because our publishers sometimes neglected to register our copyrights. We've always been told that registration is just a formality and it's not necessary to defend our work in court, but the settlement was written in such a way that it matters. Our contracts are written in such a way that the publisher is responsible for registering the copyright, but has no liability if they didn't. Because of details like this, the database of stolen work includes 24 of my titles, but only eight of them qualified for the settlement.
It isn't known yet how much the lawyers will take, but the judge sent their request back and told them they couldn't have the huge percentage they wanted. Whatever makes it through that gauntlet will be split (50/50, in most cases) betwen the writer and the publisher. (Yes, the same publisher who cost me many thousands by neglecting to register some of my copyrights.)
I think my settlement check will be in the low five figures, but it really should be approaching six.
@Bernard I agree with everything you wrote, Bernard.
They have indeed swiped a good chunk of human effort. I think the Anthropic settlement which, although it was a billion-and-a-half dollars, feels like they're throwing us a bone. When it came through, I thought that perhaps all the AI companies were going to at least have to throw us a bone, but I see no sign of it.
Did the other big AI models train on pirated copies too?
-
I believe they all trained on pirated work, because there's no other way to get enough material (other than obeying the law and paying for it), but I can't say I'm absolutely certain about that. It's hard to be sure of your sources these days.
I do know there are other lawsuits, but they seem to turn on different fine points of copyright law, so winning the Anthropic one doesn't mean that they will all go the same way. There was a big news story a few months ago saying that Meta didn't just train on pirated databases, but it was revealed in discovery that they had discussed it by email (or text, but anyway, it was in writing), and went so far as to talk about destroying the databases, thus confirming that they did it and knew it was illegal. But then the news cycle moved on and I've heard nothing else. And, of course, I have to trust the reporting that this was true, which is a problem these days.
We've been told that the Anthropic payout is coming in installments and the first checks will go out in November. I don't even know yet how much mine is supposed to be, but I guess I'll know a little more then.
-
-
I believe they all trained on pirated work, because there's no other way to get enough material (other than obeying the law and paying for it), but I can't say I'm absolutely certain about that. It's hard to be sure of your sources these days.
I do know there are other lawsuits, but they seem to turn on different fine points of copyright law, so winning the Anthropic one doesn't mean that they will all go the same way. There was a big news story a few months ago saying that Meta didn't just train on pirated databases, but it was revealed in discovery that they had discussed it by email (or text, but anyway, it was in writing), and went so far as to talk about destroying the databases, thus confirming that they did it and knew it was illegal. But then the news cycle moved on and I've heard nothing else. And, of course, I have to trust the reporting that this was true, which is a problem these days.
We've been told that the Anthropic payout is coming in installments and the first checks will go out in November. I don't even know yet how much mine is supposed to be, but I guess I'll know a little more then.
We've been told that the Anthropic payout is coming in installments and the first checks will go out in November. I don't even know yet how much mine is supposed to be, but I guess I'll know a little more then.
From what I’ve read it’s 3k per in-scope book but split with the publisher. So 1500? I don’t remember if the 3k number is net of attorney fees or if that has to get taken out of the 3k before splitting it.
-
The presumption in the writing community is that attorney's fees will be taken out of the $3000, but I haven't seen that confirmed. I also have not seen anything confirming how much the attorney's fees will be.
Splitting the payout with the publisher 50/50 seems to be the default. I looked over my contracts to see whether this was true for me. My subsidiary rights clauses specify a 50/50 split for most rights sales, but I didn't see anything on the list of specified rights that I thought definitely applied in this case. My agent, however, said that the 50/50 split applied in my case, so that's what I put on my claim, since she negotiated the contracts and presumably knows. (Interestingly, I know of no route for agents to be paid. My agent has gotten her 15% for things like audiobook sales, but I don't think this is the case with the Anthropic settlement.)
Some publishers, including big ones like Wiley, are trying to take 100%, which is unconscionable. Those writers' claims, and those of writers who didn't go with the 50/50 default, are finding that their claims are being disputed by their publishers. Will this result in lawyers' fees eating the claims up? Nobody knows, but it will certainly delay their checks. My 50/50 claims were not disputed, so I am supposed to receive my first payment in December, but I will believe it when the checks are in my hand.
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login