Problems with Publishers Moving to AI-Based Production
Straive is a firm that uses AI to, among other things, help publishers with various tasks “across the publishing value chain”.
One of its clients is Springer Nature, the publisher of many philosophy journals.
Recently, Harry R. Lloyd, assistant professor of philosophy at UNC Chapel Hill, had an article accepted for publication in one of those journals, Synthese.
How’s Straive doing with its aims of helping publishers with “accelerating time-to-publish,” “empowering authors,” and “driving cost and operational efficiency”?
Well, in an email describing the publishing process he recently went through, Professor Lloyd used the phrase “nightmare experiences.” That doesn’t sound good.
He says:
Of course, one expects some common-or-garden incompetence from production companies, but unfortunately Straive takes this to a whole new level, introducing dozens of bizarre errors. For instance, in my articles they have: removed entire paragraphs; replaced several instances of the pronoun “one” with the numeral “1”; deleted countless possessive apostrophes; and removed numerous grammatically necessary connectives. (Oh, and they are also completely incapable of typesetting even basic mathematics.) One of my articles with Synthese had to go through six rounds of proofs, and even then Straive left errors uncorrected in the published version.
Professor Lloyd has a couple of questions:
1. How widespread are these problems? Have other authors had similar experiences?
2. Are journal editors aware of these problems?
To these we might add:
3. Are journal editors given a choice in whether to use Straive?
4. Have journal editors voiced concerns to Straive or to the publishers using it to, and if so, what have their responses been?
5. What other firms are publishers using in this capacity, and which publishers?
Discussion welcome.
Meanwhile, grad students need income. Welcome to progress, folks.
I had an article published recently at a Springer journal. Straive turned the latex file into a pdf that completely mangled the citations and messed up several symbols. The errors took 3 rounds of page proofs to fix, but at different stages, additional errors were introduced. The final published version has errors that they introduced, and the html version online does not display correctly.
People who write for the academic presses could really benefit from some sort of description of copy-editing practices at different journals, book publishers, etc. I doubt Springer Nature is the only group outsourcing production to machines. My impression is that the amount of crazy errors that would never have passed a check by a literate human. has been increasing rapidly over the years. Copy editors are not introducing these errors. They are produced during automated conversion to XML templates (or even MS Word files) and things like that. Copy editors are also not approving them. The author’s eyes are, in many cases, the first ones to see the wreckage. Journals and (esp.) book publishers who have in house copy editing teams who work directly with author sourcecode and review compiled products would be good to know about. It is amazing to me that there is no easy way to learn about this.
I have recently had two papers accepted in Springer Nature journals and both processes have been absolutely horrendous. The proofs I’ve received had enough errors that it’s taken me several workdays just to catalogue them and the online proof-editor keeps crashing and losing all progress (despite regularly pressing the ‘save’-button). It’s gotten to the point where I am seriously considering refusing to submit papers to Springer Nature because the production process is just such a frustrating hassle.
Sadly it isn’t just Springer — Cambridge University Press also uses Straive, and their performance working with CUP seems to be no better than with Springer
I had a similar incident with a Taylor and Francis journal (many rounds of proofs due to changes introduced by copy editors, including the dreaded “paragraphs being deleted and needing to be reinstated.” The executive editor said that it had happened before but whatever was happening with the copy editing it was something the publisher was doing, and that the process was essentially a black box.
Answers. I typically handle about a hundred submissions at a time, and direct a team of about twenty associate editors, for AI & Society.
Professor Lloyd has a couple of questions:
1. How widespread are these problems? Have other authors had similar experiences?
Apparently, widespread. We were surprised to see problems recently; took me off-guard, I assumed it was the authors using LLMs to make their reference lists, do their research, write their papers, which is increasingly common. When we learned that it was Straive introducing these errors, I immediately suspected that it was LLMs. Thank you for the confirmation.
2. Are journal editors aware of these problems?
Can’t speak for anyone else, but, this story naming Straive as a source of problems is VERY helpful. We started seeing many mistakes during communications, requesting extensions to deadline for instance, which had been smoothly processed for all of the years with which I had been active with the journal and with Straive handling the processing, until this past few months. Our regular contact was out of the office for a long while this past autumn, and had assumed that substitutes had been poorly trained, so had sent some feedback about this to the people who are nominally responsible for such training and the supervisions of said staff. Now, my suspicion is that Straive may have been testing a bot to do this work, as we began seeing problems with errors introduced during production, as described by Prof Lloyd, also during that period
To these we might add:
3. Are journal editors given a choice in whether to use Straive?
No.
4. Have journal editors voiced concerns to Straive or to the publishers using it to, and if so, what have their responses been?
Yes, as above. I had received no response. But, I didn’t really expect one. Our managing editor also had communicated concerns, specifically about errors introduced during processing, to who I am not sure but I expect that it was with Springer people, directly, and about these communications, I have no information. And, trying to do my job well takes enough time to try to do everyone else’s too.
In the end, I suppose that the take-away is that Springer-Nature is aware that their roll-out of LLM managed production has been problematic. But, my suspicion is that they anticipated some collateral damage. And, I further suspect that the end-game is to cut human editors and reviewers out of the loop for most publishing, completely.
The cumulative case for abandoning for-profit publishers like Springer and moving to a model like Phil. Imprint or Ergo is getting pretty strong. The copyediting and typesetting is supposed to be one of the big things to be said in favor of these journals. If that’s gone out the window, it’s not clear why we should jump out with it….
The question is if there can be a community run journal like Ergo or Phil Imprint that is also able to process the vast amount of submissions and correspondingly publish a lager amount of papers, in the way that certain Springer journals like Phil studies or Synthese do.
I don’t see why that’s the question.
What if there can’t be, but there can be lots of smaller ones?
I think the question is how we can get enough independent funding for free online journals to do the job the for-profit ones are doing now. It won’t matter if that amounts to funding thirty big ones or ninety small ones.
I have two forthcoming papers in Springer journals. Straive’s conversions of my .tex files introduced countless incomprehensible errors. I am currently on rounds 3 and 4 of proofing, and I expect several more. Some of the most blatant issues have not yet been fixed, despite being flagged multiple times. I fear that no human at Straive has looked at my proofs. It is as if they are being fed back into their AI along with my list of requested changes, with only the first few being registered each cycle, and with new errors being introduced.
Moreover, Springer’s in-browser Proof Editor is borderline unusable. Unintuitive, unstable, slow, and inexplicably limited.
“I am currently on rounds 3 and 4 of proofing.”
One of the de facto functions of computer technology is to shift work and costs from one person to another. E.g., back in the old days, if I wanted you to read a paper by me, I would print it out and mail it to you. Now, I just send you an attachment, and YOU print it out; the work and the cost is now mostly yours.
In this case, Straive is making you do the work. They use an error-inducing computer program, and you fix all its errors. This is labor- and cost-saving… for them. But really it isn’t cost-saving, it’s cost-creating and cost-shifting.
Update: Two weeks ago, I submitted a list of 47 necessary changes that need to be made to Proof 3. I’ve received Proof 4. Straive says, “Hope corrections done are all up to your satisfaction”. But… Proof 4 is indiscernible from Proof 3.
I had similar problems with a recent paper at a Springer journal. The copy-editing added new, and sometimes surprising, errors to the body of my text. It also missed a few typos that I caught. But the real disaster was my references. It added inconsistency to them, so that sometimes the years of works were put in parens, other times not. It also inverted the editors and titles of collections so that editor names became collection titles and collection titles became editor names. There were also a few references in the body of the text that it didn’t ask me to add, and a few in the references but not in the body that it didn’t ask me to remove.
I’ve now sworn off Synthese/Springer after a recent disaster of this kind. Among other errors clearly only producible by LLM were a large number of author queries saying “this reference does not appear in the bibliography”, inevitably complemented by more queries saying “this bibliography item is never referenced in the text”. Any human would spot the one-to-one relationship.
I’ve made a new year’s pack to really try to only consider submitting to university/non-profit presses, even if it ultimately means I may wind up with fewer papers accepted and more forever preprints posted to the archive.
This all sounds really bad. However, I’m reminded by this point in Ethan Mollick’s *Co-Intelligence*: whatever AI you’re using now is the worst version you’ll ever use again. Because AI keeps getting better.
So, granting that Straive is terrible now, what will it be like in six months? Is the issue here that it was just rolled out too early? Or is there some other worry (i.e., if it worked fine, would we still care?)?
the word “better” is doing a lot of work here. in some basic sense, the quote is accurate. at the same time, here we are 2.5/3 years into llms and i would argue that the issues we saw initially are still prevalent. so the implication of the quote, as it appears here, that these issues will be resolved seems optimistic at best and unconcerned with facts at worst.
it’s true that outrageous hallucinations are less common, and it’s true that the phrasing and tonality of llm produced writing is less stilted than it was two years ago. at the same time, hallucinations are still a regular occurrence (see my students’ ai-produced final papers this quarter) and actually more troubling because they’re less obvious than before. additionally, there seems to be a rising vigilance online and in classes concerned ai-produced work, meaning that folks are getting better at spotting it even as it “improves.”
now, an in-house system that isn’t a chatbot and uses algorithmic ai to do proofing *might* show a great deal of improvement with use in a way that internet connected llm chatbots haven’t, but the idea that unpaid beta testing is to be done by scholars trying to get their work out there is pretty stunning really.
> an in-house system that isn’t a chatbot and uses algorithmic ai to do proofing *might* show a great deal of improvement with use in a way that internet connected llm chatbots haven’t
I actually suspect the reverse here. There’s plenty of reason to think that the general purpose AI chatbot systems used by the public will have pressure to continually improve, even beyond the evidence we have of them already improving. But as Brian Weatherson says, these proprietary single-purpose systems may not improve as much.
Still, one other caveat on all this is that even as one system (or person) improves, they also often get new responsibilities. A system or person that is better at its old responsibilities, but is now in charge of additional new responsibilities, often does a worse job. (For humans, this has often been known as “the Peter principle”, which states that people who do a good job often keep getting promoted until they reach a level of responsibilities at which they no longer do a good job. But I suspect the AI version will be even more obvious.)
But why should we think Straive in particular will be better in six months? Is there market pressure to get better? It sounds like it was really bad before the widespread use of LLMs, and it is still really bad. The examples everyone describes here are much worse than you’d get if you simply said “Dear mass market LLM, please copyedit this paper and output it in some kind of markdown that I can feed into a typesetting program”. They are so so much worse than you’d get if you added “Also, turn on advanced research/extra thinking/whatever fancy settings are available, even if takes a little more time and costs a little more money.” Even if they want to use AI for this, the fact that they won’t spend the equivalent of $5 to $10/paper to use the best available AI makes me sceptical of their good intentions and good practices.
Put another way, you only improve when the cutting edge improves if you are already somewhat tied to the cutting edge. And there seems to be a lot of evidence that Straive in particular is not.
whatever AI you’re using now is the worst version you’ll ever use again. Because AI keeps getting better.
I would not be so sure. For one, there may well be limits we’ll hit quickly, that can’t be over-come without lots of money, and most people won’t or can’t pay lots of money. But also, lots of tech stuff has become worse, not better – google, lots of features on Amazon, etc. The stuff will get better if it’s to someone’s clear advantage, and it may not be. So, I think it’s overly-optimistic to think it will just get better and better, at least for most people.
It must have been rolled out too early since I think the paid or even versions of the major LLMs could proofread and format passably. Truly inexplicable
> whatever AI you’re using now is the worst version you’ll ever use again.
Haven’t m a lot of these companies been operating without profits or at a loss, often to try for market capture? If so then they may well worsen the models they offer to save compute as things go on
This is wishful thinking since you can run the same argument for anything, like, say, search engines. You’d think search engines would be better tpday than 5 years ago, but thanks to AI they are markedly worse!
but thanks to AI they are markedly worse!
Importantly, I think, it’s not just AI that has made search engines worse, although of course it has. Rather, it’s because it turns out to be more profitable to have an, in some ways, less good search engine than a better one, at least if there are significant barriers to entry for competitors (who would also like to be making more profits, so have similar incentives.) This impacts what features are available, what sort of customer service is provided, which results get highlighted, etc. You can see this in many fields, not just tech. (It’s clearly true of resturants, for example.) I don’t mean this as an argument against “capitalism”. I’m extremely skeptical that a “public” search engine would be great, either, for well-known reasons. But it seems to me to be overly credulous, even polly-anna-ish, to think this stuff can’t help but get better and better, and keep doing so indefinitely.
Yes, Google deliberately sabotaged PageRank so that we have to run multiple searches to find what we want, thus giving them more opportunities to show us ads. Amazon likewise deliberately introduced slop into its search function, so as to promote either its own content, or knock-offs that paid it a premium. Hell, you often can’t find a book even if you have both the author’s name and the title (and if you search for mine by my last name, you’re redirected to “Chinese”!).
I’m happy to report (I just double checked on this) that The Catholic University of America Press still uses actual humans for copy-editing, book and cover design, the works. They don’t even use AI to write the copy for their catalogs.
While it’s easy to pile on the technology here, the real story of course is the human beings involved and their mentality. The fact that they would even consider risking this absurd situation shows you how deeply unserious they are. They lack any commitment whatsoever to the scholarly process or to the value of human communication.
These are not our friends, they are our enemies. Those of us who can detach from them should do so now. Those who can’t, my heart genuinely breaks for you, that you are forced to feed your life’s work and intellectual passion into this machine. To have it mangled by AI before it is used to train AI. Indignities piled upon indignities.
Anyone want to start a print-only philosophy journal with me?
Do you have any worked out idea about this or just a hunch. I like journals so hit me with ideas 🙂
Btw. I’m also profficient in Indesign as I learned it to lower the cost and increase the speed of publication of a journal I was previously associated with.
One thing I’ve wondered for some time is whether it would at least be possible for journals to send alongside proofs a list of changes that have been made to the author’s copy. This could be a good idea even for copyediting done by human beings with the best of intentions. A good philosophical writer will often use punctuation, word choice and sentence structure to preempt things like scope ambiguities that could confuse the reader — and in my experience even the best copyeditors will sometimes make small edits that completely undo those careful writing choices
I think–but could be misremembering–that my book with Routledge (owned by T&F) a few years ago went throughout the Straive, and they introduced many more errors than they caught, and bungled a diagram for two rounds. (I do some freelance copy editing on the side, so my standards for my own writing are fairly high, and my eye is pretty sharp.) At the time, I attributed it to the copy editor being a non-native English speaker with a good but occasionally shaky grasp of English being rushed. It could also have been an automated system, though likely not an LLM, because it was too early for that (IIRC).
By contrast, two books have gone through OUP’s copy editing in the last couple of years, and (1) they use a real human being (Timothy Beck, though I’m sure there are others), and (2) he is very good and conscientious. Also, (3) he’s quite fast.
Argh, *through Straive. I can’t keep up with the blasted autoincorrect on my phone.
For what it’s worth I’m currently publishing a book with OUP and they’re using Straive for mine at least. I’m still waiting for the proofs but the copy-editing was a mess. I’m not sure about the human:AI involvement ratio, but whoever did it caught a few errors, introduced about as many, and missed c. 90% of those I found myself while reading through. They were inconsistent enough that my hunch is it was mostly a human, but one out of their depth with a complex book using mostly historical, non-English sources. Here’s hoping that the proofs aren’t a complete dumpster fire.
Hmm. I just had the second book go through copy editing last week, and it was a real human being.
Perhaps it’s determined at the series level. Both of mine were part of the BSHP’s New Texts in the the History of Philosophy series.
Michel, are you working with OUP (US) or OUP (UK)? Yob, as i posted separately I had a disastrous experience with OUP (UK) and Straive. Good luck!
Thank you! I’m with OUP UK but assumed they would be uniform between the US and UK, though who knows. Either way, glad yours is going better, Michel! Mine is in a history series, for whatever it’s worth (I would rather not say more – it’s my first book and I don’t want to risk causing trouble this close to the finish line).
OUP UK.
I guess it has to be a series-level difference.
(Un)fortunately, I just recently got a paper accepted into a Springer Nature journal and I’m currently waiting for proofs. The comforting fact is that there no formulas inside, just plain text so it could be ok.
How did things turn out?
I miss the days of dystopian fiction. Dystopian fact just isn’t as fun.
This one is about Phil Studies, which is also a Springer Nature journal. My co-author Matthew Chrisman and I have a paper ready for publication: we submit it as the final draft, and then we wait for the proofs. When we get the proofs, they’re based on the penultimate draft, not the final draft we submitted. Strange error, but easy fix: just generate proofs from the final draft. We ask Springer to do this. We get the proofs back again, and they’re still based on the penultimate draft; this happens again; one day I decide, Fine, I’ll just line-edit the proofs that they keep sending us and be done with it. I get an hour into that process and get annoyed, so we go back to Springer again, saying, Send us the right proofs. There are several more rounds of back and forth, and each time they send us proofs, they’re based on the penultimate draft, which at this point includes the line-edits I spent that hour on. That is, we just keep getting the same partially edited file back over and over again–they’re not replacing the file with anything, at all. It took almost three months to finally get the right set of proofs.
Haha this is insane. They are really taking the piss
to me this is a perfect analogy for working with Ilm chatbots more generally
The currently hyped systems have, in my experience both using them in meeting transcription and pentesting systems based on them (I’m a cyber security professional), the problem of connectives in spades. For example, at least twice I’ve been ‘recorded’ as not using “not” or similar, when I did. So I am not at all surprised that these “editorial systems” have similar problems. I have also had some of the image drawing stuff try some elementary organic chemistry – which they fail utterly at, which might be similar to the mathematical typesetting problems mentioned.
I published a book with OUP (UK) and they also used Straive for type-setting/copy-editing. I was never sure whether I was communicating with a human or an AI but it was an extremely difficult and unpleasant process which wasted dozens of hours of time.
I take the now radical position that philosophy should be written, read, and YES, even *edited* by, human beings.
My first experience in publishing taught me to never trust Springer. All the years since, I always ask them to see a second, or if need, third, forth, or nth proof, before I even start my proofing process.
While this procedure has provided me some protection, it is not perfect. Indeed, I recently had an experience in which they made changes after everything was settled. In this case it was the journal editor that was to be blamed (probably due to communication issues, as I had been promised a standard deviation from another editor).
It is kind of strange how the money-making machinery has standards much lower than diamond open access journals (take, e.g., JESP as an example). Or perhaps it isn’t?
This seems like a useful life hack. Do you mind describing what it is that you actually do? Do you reject the proofs as inadequate, and do they then send you new different proofs in return? (And do you repeat this until you are satisfied with the proofs you get?) Maybe I misunderstood you…
I have reviewed proofs for three papers over the last year or so: two SpringerNature and one with Wiley. Both seemed to use automated typesetting from my LaTeX files. I saw many of the problems recounted in the post and comments from SpringerNature, though nothing close to the worst reports here. Riley was great to work with. So, at least these anecdotes suggest that Wiley >> SpringerNature. I don’t know whether Wiley uses Straive, but it does seem possible that automation can be introduced without creating such a ghastly mess.
I had a similar experience with one of the Springer journals. Instead of proofreading, the AI system messed up the citation system, changed the numbering of paragraphs, and destroyed the tables.
I had a bad experience with Straive at OUP (UK) already in 2021 as co-editor of a volume. OUP had an in-house person helping us with formatting the final manuscript. That person was excellent. The same goes for all other OUP-people we were in contact with.
As soon as the document was submitted to Straive for the final typesetting, however, things went downhill. (In fact, Straive only emerged throughout the process, at the beginning it was the predecessor company SPi-Global.) The proofs contained so many errors that it took us a good month of full time work to correct them. Anything containing math was particularly bad.
Some examples:
on one page, the symbol for conjunction was typeset as four different symbols (an ampersand and three versions of \wedge); negations were sometimes set as minuses, sometimes as ~, and sometimes correctly (albeit at different distances to the negated formula). The original manuscript did not contain these variations, so we were extremely puzzled why this happened. AI might be a good explanation.
It was not only math. Another example (one of many): For an inexplicable reason, a chapter that was supposed to be “Chapter 4” was numbered “Chapter 3.4”. Upon our request to change it to “Chapter 4”, we received a document with the heading “Chapter 3.4: Chapter 4”. Again, this seemed extremely puzzling and borderline passive aggressive at the time (mind you, one year before the release of ChapGPT). But in hindsight it seems like a typical error that not-so-well-trained AI-algorithms make.
This might also explain why the communication with Straive was painful and intransparent. The back-and-forth went on over more than six months, and we often had the feeling that the person at Straive did not understand us at all.
What is perhaps most egregious about this is that this can only happen in the world of academic publishing. Companies de facto outsource much of the typesetting work to authors. Given the prestige that comes with a publication at OUP (or Synthese etc), they can count on authors’ doing the work instead of simply choosing another publisher/outlet.
This isn’t about journals, but I published a book with Springer recently and production was handled by Straive (no choice of course). “Nightmare experience” sounds about right. The introduction of errors that were not in the original was intensely frustrating and correcting these errors was like pulling teeth. I don’t know for sure if they were using AI and I had chalked it up to human error, but this could be what was happening. A few errors made it into the final product, with the worst probably being a body paragraph that was repeated as a footnote of itself. While embarrassing, the damage could have been much worse if I had not personally proofed it so thoroughly. Their proofreading system is unfortunately also extremely glitchy and it deleted entire chapters of proofing work multiple times after crashing. I came away with the feeling I was the only real quality control or human copyeditor. My experience working with the people at Springer itself was great—I just wish they didn’t use Straive. I did let Springer know this and I hope others have done the same.
My book on William Blake, forthcoming from Palgrave Macmillan, was copyedited and proofed by Straive’s “agentic” AI system. Sheer nightmare from beginning to end. Almost all of the above-mentioned proofing errors occurred: random typographicals, systematic typographicals, chapter subtitles dropped into the main text, faulty line spacing of indented quotations, faulty numbering of illustrations, an illustration repeated in two different places and the correct illustration somehow made to disappear, dropped capital letters in the illustration captions and chapter titles, dozens of distracting paragraph indentations added (usually after an indented passage of quotation); a text paragraph made into a footnote of itself; reformatting of some bibliographical entries—and much, much more. In my case, too, dozens of new errors were added with each “corrected” set of proofs. Overall, four or five different proofs were required (and let’s see how many of these errors still make into in the printed book). The human editorial “agent” frequently disregarded my corrections; the same errors kept reappearing from one proof to the next. As Jeffrey White says above, Straive appears to be testing out a new bot. I wonder if that puts Straive in breach of contract with one of their publishing partners (Springer, Taylor&Francis, Palgrave Macmillan). Or don’t they mind?
I got a paper accepted with a Springer Nature journal three months ago, and am still dealing with the copy-editors. The production editor doesn’t seem to realize how bad the work is that he’s forwarding to me from his team. After the first round of proofs (which involved dealing with a sluggish web interface), I got a PDF. There then were hundreds of errors to be corrected, typically trivial ones, but not always. The copy-editor refused to correct most of them. He also introduced some bizarre new errors, like question marks in place of quotation marks. I don’t think that AI would make mistakes of this kind, but a highly careless and incompetent human just might.
I forgot to mention that the production company in question is Straive.
to straive (v.t.)
We just had the same nightmare with a Springer philosophy journal with my co-author. I don’t even remember how many proofs I checked. In the first version, 99% of the citations were wrong: the number in the text leading to a completely random reference in the end of the paper. Many of the mistakes were totally bizarre. Random extra spaces between words. References lumped together. The font changing in the middle of the paper. Words lined over suddently. I don’t think they were made by a human person. The publication process was prolonged at least four months because of all the edits to the proofs that we always sent back the next day. I hope this does not continue in future papers. Something’s wrong with proofs in Springer.
Recent experience with Springer journal AI & Society: I answered inquiries etc. in the proofing system. Double-checked everything (because a lil neurotic) before submitting the final proof. Paper published just a few days later and, to my great surprise, several errors had been introduced! Among the more serious ones were missing references in the reference list, and some references that had just been inserted at the end of the list (i.e., not in alphabetical order). I immediately emailed the journal with a list of the things in need of correction, and wasn’t keen on sharing my published article. Now, more than five weeks have passed and despite my reminders, the published article remains unchanged. I find it embarrassing that others would think the errors were mine, and worrisome that the journal doesn’t seem to take this seriously at all.
Post doc
It may be more expedient to ask to publish a Correction – in it, list all the things that were done wrong, and state clearly that these errors were introduced in the production stage. I worked on a Springer/Nature journal … incidentally the correction will be published and linked to your article (and it will be published OA)
I am publishing a historical book with Cambridge University Press and was just assigned a project manager at Straive. The book is text and image only (no formulas or tables). Does anyone have related experience? How exactly did they use AI as part of the process?
I’m the editor-in-chief of a Springer-published humanities/social science peer-reviewed journal that uses Straive for production. I can confirm that our proofs come back with bizarre and inconsistent errors that were introduced after copyediting and weren’t in the copyedited manuscript. The problem has also been getting worse, leading to significant delays in production as our authors and copyeditors then have to go back and correct all of the errors introduced at the proof stage. It’s exhausting. I wasn’t aware before reading this article that Straive was using AI; it explains a lot.