(Full disclosure: I have used all available tools for researching the current state of play on the subject, but the structure, the main arguments, and the wording of this article is entirely mine).
At least one writer who I am aware of has in the last couple of years produced a string of books with such rapidity and on such a range of subjects that I wonder if those books aren’t wholly or largely AI-generated.
On the other hand, everyone who uses a desktop, laptop, tablet, smartphone or any other such device anyway uses, perhaps even without realising it, operating systems, browsers, and a whole portfolio of tools that are now AI-based or at least AI-assisted. Such tools include those for email, spreadsheets, antivirus/security, music/media streaming, location/proximity/maps/navigation, product ranking/valuation, basic Q&A/search/research, photo/image editing, and even the most basic word processing and writing (such as Apple’s Pages, Google Docs, LibreOffice, Microsoft Word).
AI is now embedded in every aspect of modern life from money itself, to the purchase of groceries, to the use of an aeroplane bus, car, or train. Whether you are hiding from bombs in a city bunker or walking peacefully in some remote and rural area, your safety and security depend on the effectiveness of national security systems which increasingly use AI.
So the question for writers, editors and publishers is of course that of how best to use them.
For what sorts of purposes are AI editorial tools genuinely useful? Catching typos, checking sentence length variation, flagging repeated words, identifying inconsistencies – and, more widely, grammatical infelicities, pacing problems, and structural imbalances. You will notice that the list progresses from merely surface-level, specific, and bounded matters, to more important ones.
But what sorts of things can’t AI tools do at present? Assessing whether your book will matter to readers, whether your themes cohere at a deep level – and, in the case of fiction, whether your characters are genuinely alive or whether your work resonates emotionally with readers.
In other words, AI tools don’t have, and can’t have, aesthetic judgement. They are statistical pattern recognition machines that can and do pretend to aid aesthetic evaluation. As suggested above, of course AI can help to produce technically clean manuscripts, and all manuscripts do need to be technically clean. So, AI tools can help us achieve the qualifying round of becoming basically acceptable. But the litmus test for any work that is worth publishing nowadays is that of whether the work at least aspires to be good and, preferably, great.
For anything like greatness, or even goodness, the things that really matter (cultural authenticity, emotional truth thematic coherence, tone, voice) the tools that are generally-available aren’t yet of essential help – and, to the extent that they are, they need to be used with caution. However authoritatively presented by the tools (and aren’t they put totally authoritatively!) their suggestions are merely suggestions, and need to be treated as merely one input among many while making up your mind about what strategy to pursue or what matter of detail to accept – just as a spell-checker is a basic tool that is now so ubiquitous that authors more or less can’t help being attacked by them, but certainly no self-respecting writer can let a spell-checker rewrite whole sentences, let alone a whole manuscript.
And are there deeper dangers in all the “assistance” that AI offers us? How can we use AI as a slave rather than allowing it to enslave us?
Here is what I see, though I don’t detail them here in any order of importance or priority mainly because I have so far found all the following issues difficult to disentangle.
First, “ghosting”. However much of a beginner you may be in the use of AI tools for writing, editing and publishing, I’m sure you’ve experienced the problem, even in relation to facts, of your AI tool indicating a problem, or even wanting to “correct” you, while being entirely wrong! In other words, AI tools generate commentary which sounds accurate or at least plausible, when the machine is producing only what meets its own standard of statistical approximation (which may or may not be high). Every writer and editor, especially one who doesn’t already know a subject, shouldn’t act unthinkingly on the mere basis of the machine’s suggestions – as that might even introduce errors into a manuscript that was already accurate (and now may be no longer so!). The only reliable way of using AI for anything, in relation to every suggestion that might have any chance of being dubious, is to question further, and even argue against the machine: “How do you know that?”, “On what basis are you making that assertion?”, “Hey! What about….?” The remarkable thing you may already have found in your experience is that, once the machine has said something, it fights very hard to defend that, to justify it, to find all kinds of “evidence” to bolster its position. And it is only by repeatedly questioning everything that the machine presents that you have any chance of getting at anything like the truth according to even the body of material in that machine. And when you have got at that, it is still best practice to counter-check that with at least two different tools/platforms.
Second, bias. It is hardly surprising that all AI tools are trained predominantly on anglophone, Western material. Therefore, whether in the case of fiction or non-fiction, the machine inevitably operates on Western notions of realism, narrative structure, and quality. If your work draws on material that flies in the face of “facts”, conventions, structures, or registers shaped by other linguistic inheritances, AI tools are extremely likely to misread these as weaknesses. The fact is that they may indeed be weaknesses, but they may also be strengths.
Third, homogenisation. AI tools also have powerful preferences for certain sentence structures, rhythms, and vocabulary choices. Edits suggested by AI tools nudge all prose toward a statistical average of the published writing on which it has been trained. Applied across a manuscript, is that not likely to dilute the distinctiveness of your voice? The irony is that the very qualities a good human editor would fight to preserve — your idiosyncratic punctuation habits, your unconventional sentence rhythms, your characteristic vocabulary — are often exactly what AI tools flag as errors or weaknesses. Some tools proclaim that they work to “preserve authors’ voices”, but is that structurally at all easy to achieve given the AI’s actual architecture works against it?
Fourth, detail correction but aggregate falsification. You may have found, as I sometimes do, that while individual AI suggestions may indeed be accurate or reasonable, accepting many of them cumulatively transforms a manuscript— though one each change because one feels that it is improving manuscript, at the end one senses that it is now quite far from the author’s authentic voice. You, as the human writer/editor/publisher have to hold the character, spirit, intention, objective, strategy, and everything else about the whole book in mind. The machine’s strength as well as its weakness is that it focuses on a few specific aspects, and cannot consider all significant aspects of a whole work simultaneously.
Fifth, dependency risk. It is possible to have too much of a good thing. Just as children brought up with access to electronic calculators, often can’t even add up on their own, there is a real danger of becoming dependent on AI, so that you lose or at least weaken your own editorial judgement.
Sixth, detection and reputation risk. Even if an author uses AI only for editorial feedback and has personally written every word, there is a risk that AI detection software wil flag your manuscript if the tool’s suggestions have been a little too enthusiastically accepted. No, AI detection is not 100% reliable, but the problem for an author is that even if he or she is wrongly accused, that accusation is at present practically impossible to disprove, and the burden of proof falls entirely, and perhaps in current circumstances unfairly on the author. Perhaps in future there will be some software for an author to track use of AI tools in relation to a manuscript from research stage onwards which time stamps and records also the rights granted to every AI tool. But at present this is all a bit of a juggernaut moving at great speed with all sorts of implications that are impossible to assess. All that can be said is that authors need to be fully aware and do their best to ensure that they use AI minimally for their creative work.
Seventh, disclosure obligations and publisher relationships. As of 2026, all authors should take as standard best practice that they are required to declare if they have used an AI tool for writing purposes. You can take, as the fact to keep blazoned in your mind, that Hachette cancelled its contract for the horror novel Shy Girl on March the 19th 2026, *after* having published it in the UK and even after copies had probably been printed for publication in the USA on April the 7th – in spite of the fact that the cancellation cost the publisher at a very minimum, for the following:
- Editorial work
- Proofing, typesetting, and cover designs
- Printing and supplying ARCs/galleys 4–6+ months ahead of publication date for reviewers, sales teams, and blurbs.
- Catalogue reprinting
- Sales Conference materials, travel, hotels, campaign development, and related items
- Printing and then pulping the UK edition
- Legal costs
- Almost certainly the printing and pulping cost of the US edition since the publication date was less then 3 weeks from publication date (though there is no actual evidence, nor any official statement regarding this matter).
Hachette cancelled because AI use was initially suspected and then apparently established. What all that means for authors is that they need to disclose the extent of AI tool use. Most publishers include the requirement under their policy documents, though some already have on their actual Agreement with authors that they must declare that they have granted the AI tools that were used *no* other right than the right to use the author’s materials to provide the relevant service “exclusively to the author” (in other words, no rights to the AI tool to use the author’s materials for training of the LLM. If an author uses an editorial AI tool without disclosing the fact that he or she has used it, and what rights were granted to the AI tool, the author will almost certainly find her/his potential or actual publisher accusing her/him of breaching the agreement and withdrawing from the book. Now, it should be said that no publisher has done the following yet, but expect it to become more often the case that publishers will start going after authors to recover the publisher’s costs.
But let’s not focus exclusively on the negatives, let us look at one advantage of the use of AI tools – at the stage of research. The extent to which AI tools can help with brainstorming (suggesting avenues, aspects, and points of view that might never have been explored otherwise) is simply wonderful.
Let me conclude by sharing my experience of listening to a highly respected author who’s produced a lifetime of work on a particular subject, showing off a Large Language Model trained exclusively on all his own work, which could now respond to any question posed to it, by voice or text, and respond to that question with answers either in the exact words from the author’s vast output or in answers that he would have given had he written on that topic – in the style of that author.
I was terrifically impressed, both with the author’s pioneering use of AI and with the creative use of the generative capabilities of AI. But the immediate impact on me was moderated a moment later by three reservations:
- The author’s system, let’s call it X, was great for producing generations of “disciples”, which is what he was going to use the purpose for, but was it going to encourage critical thought, or balance, or any broad and comprehensive understanding, let alone any wisdom, any fresh insight?
- X was also excellent for enshrining all the author’s work to that date, but did he really never again want to write anything else? What if his new writing contained some alteration or modification of what he had thought earlier, or even some boundary-breaking material? Would his LLM refuse to accept it on the grounds that it violated the “fundamental structure” of the material it already contained?
- And how about style? Is it the case that a writer produces work in just one style throughout her or his life? Certainly not in the case of imaginative work, and I hope that is not the case even in the case of non-fiction. Would LLM refuse to accept work in any new style as being inauthentic? (BTW, that is why I regard with some suspicion the new machine-based methods for establishing the authenticity of ancient manuscripts, paintings, etc. – useful tools, but bad masters.
Anyway, I rest my case, and would love to hear your own thoughts on the subject.