Automation that spares people super-mundane jobs would pass without comment if AI were not involved. What has changed is not the displacement but who is being displaced, and that this time the affected staff are white-collar workers who talk to journalists.
Wired’s “investigation” into HarperCollins, Simon & Schuster and Hachette is being read as a scandal about AI. Our industry loves a scandal! And a scandal where AI can be pointed at as the bad guy is having our cake and eating it!
But the reality is this is a story about double standards, ownership and labour, and about disclosure rules publishers demand of authors but not of themselves.
In brief: On 9 October Wired published Adam Morgan’s investigation, based on interviews with more than two dozen anonymous staff. It reports that at least three of the Big Five US publishers (HarperCollins, Simon & Schuster and Hachette) are using ChatGPT and Claude to write publicity and back-cover copy, draft emails to literary agents, produce marketing material and, in places, cover art. Executives are driving it, in some cases through “AI Champions” drawn from junior staff.
The outline is well known by now, so this piece skips the recap. Two caveats: everything is anonymous testimony, and the companies describe narrower use. HarperCollins did not respond to Wired. Hachette says it supports AI for operational purposes but not creative uses, including communicating with authors. Simon & Schuster says staff have a limited set of vetted tools and use is not mandatory.
But as TNPS has previously explored, Big Pub is no stranger to riding the wave of moral indignation while milking AI for its benefits.
The real story is the asymmetry
The same houses are party to copyright litigation against AI companies, and they have cancelled books by authors suspected of LLM use. Hachette’s Shy Girl and Macmillan’s Call Me, I’ll Hide the Body are the public cases. One editor told Wired that every imprint has an unpublicised version of the same story.
So authors are held to a purity standard, enforced by cancellation, that the publishers do not apply to themselves. If AI assistance is fine for the publicist writing the jacket copy, the objection to an author using the same tool cannot be that AI is intrinsically illegitimate. It has to be about quality, contract or rights.
That is the test TNPS has argued for from the start: judge the work, not the toolchain. By that test the cancellations look less like editorial judgement and more like reputation management.
Two kinds of disclosure
TNPS has argued that reader-facing disclosure of AI use with argument the author’s name must be on the cover is nonsense while ghost-writing and packaged fiction are routine. Nothing in Wired changes that. But the story surfaces a second kind of disclosure, and it is a different matter: disclosure to the people publishers contract with.
Agents told Wired they have added clauses barring their authors’ material from being run through LLMs. One pointed out that clauses meant to keep tech companies out also bind the publisher. The questions here are contractual. Was a manuscript put into a tool at all, and on what terms?
Enterprise licences typically exclude customer data from model training, but that is a matter of the contract, and a publisher ought to be able to say which terms apply. One HarperCollins source says editors who asked about pasting full manuscripts into cloud tools were told no by the legal team. We don’t know whether that rule is followed.
Here’s the thing: Reader-facing disclosure is theatre, pure and simple. Counterparty disclosure is basic professional hygiene. Conflating the two lets both sides argue past each other.
The tells, and the detector
How did agents conclude that editors were using AI to write rejection letters? By comparing letters from the same editors and spotting familiar patterns: the “not this, but that” construction and the rule of three and all that nonsense.
Staff at all three houses told Wired the agents were right. Wired also reports, don’t laugh, that Pangram flagged parts of the Simon & Schuster CEO’s memo on workplace monitoring as AI-written, while acknowledging the tool’s contested standing.
TNPS has criticised exactly this kind of folk taxonomy, and nothing here rescues it.
The agents were vindicated by insider confirmation, not by the reliability of the tells. Default, unedited model output is recognisable – I think we all accept that much – but that is a quality problem about lazy use, not about AI as such. Neither style-spotting nor a detector score can establish provenance or clear a named person, and both will flag human writers. This case is the exception that illustrates the rule.
A business case we prefer not to talk about
Publishers run businesses, and cutting costs is what businesses do. But Wired’s account suggests adoption is running ahead of any articulated case. A HarperCollins employee says senior leaders bought Claude licences and quickly realised they did not know how to use them, so staff were “voluntold” into finding uses. Employees describe teams cut by layoffs and now handling double or triple the titles. The sequence matters: layoffs first, AI as the shortcut afterwards, not AI as the stated reason.
At Simon & Schuster, staff linked a trial of Skan AI’s workflow-analysis software to owner KKR and circulated an open letter. The CEO told staff no decision had been made. A Big Five publicist told Wired that AI-written pitches break good faith with critics and colleagues in the media.
That is a fair complaint only up to a point. Publishing has always run on templates: the form rejection slip, the mail-merged press release, the standard welcome letter, often with not a word changed.
Against that baseline, an AI draft that a human has read and tailored can be more personal, not less. The honest test is not whether a template or a model was involved but whether anyone paid attention. A form slip makes no claim to attention. A note that appears to praise a specific chapter, written by a model and never checked by anyone who read the manuscript, simulates attention that did not happen, and a first-time author who has never seen the machine from the inside has no way to tell.
That is the line about honesty and accuracy, not about AI. A business that trades editorial and publicity quality for headcount is spending the asset that justifies its margin. One that uses AI to clear the routine and keeps humans on the reading is not.
The scholar Dan Sinykin’s observation, quoted in the Wired piece, is the context: Big Five C-suites have more in common with other corporate C-suites than with their own editors. This is a private-equity and labour story with AI as the instrument. It fits the point TNPS keeps making: AI changes the scale of problems publishing already had.
The moral circle, widened
TNPS argued in June, in “The Hand That Feeds,” that the industry’s AI position is not anti-AI but about who controls the economic benefits.
Publishers drew a moral circle around creative labour because that is where the copyright sits, leaving warehouse staff, narrators, engineers and junior marketers outside it. Wired’s story is that circle uncomfortably widening.
The contrast is stark. HarperCollins is building a $160 million automated distribution centre in Indiana, due to open in 2028, with algorithmic labour-management systems that track worker productivity. At Simon & Schuster, the prospect of similar monitoring software for office staff produced an open letter, while its Riverside distribution centre has been automated for years, shipping more books without higher staffing.
Nobody wrote an open letter about that.
Automation that spares people super-mundane jobs would pass without comment if AI were not involved. What has changed is not the displacement but who is being displaced, and that this time the affected staff are white-collar workers who talk to journalists.
None of this makes the staff’s grievances wrong. Tripled workloads and workplace surveillance are real harms. It does make the industry’s selective outrage visible. The same house that sues Meta over training data is the one telling junior staff to become AI Champions, and the one partnering with ElevenLabs on cloned audiobook voices. Plus ça change.
The carbon line
At a HarperCollins brainstorming session this spring, staff who raised legal, ethical and environmental concerns were, according to one attendee, told it was “the cost of doing business.” Afterwards the staff portal gained a line saying AI’s environmental impact is a small part of an individual’s digital carbon footprint.
The dismissal was a management failure: three different kinds of concern waved away in one breath, to the people raising them. But the footprint argument deserves a harder look than either side gives it.
Consider what the industry already is. It fells trees for paper, pulps them in industrial plants, prints on instruction far more copies than will sell, wraps them in more paper on wooden pallets, and ships them by road and sea to warehouses and shops that are air-conditioned and, in the shops’ case, lit around the clock. Unsold stock is routinely pulped. Customers drive to the shop, while online orders are processed in a data centre and delivered by van. Publishing people fly to book fairs and conferences for a day or two. Little of this is audited for environmental impact, and almost none of it is questioned.
Against that, a publisher using an LLM to draft copy or an email that would have been written anyway adds an increment that is real but tiny. Proportion is the point, and the footprint that matters is what the copy is for. More pitches, sharper blurbs and cheaper translation all exist to sell more books, which means more paper, printing, freight and vans, plus the data-centre traffic of online sales. The industry celebrates market expansion at home and abroad as though it carries no environmental cost, and nobody wants t talk about that. AI data centres are a much easier target.
The wider concern about AI’s power and water demand is serious, and it is a matter for policy, but it concerns the technology sector’s build-out, not whether a publicist drafted a blurb with a chatbot.
The realty is – sorry, guys – publishing is a very small part of what AI is used for. We are actually not the centre of the AI universe. Protein-structure prediction, medical imaging and weather forecasting mostly run on systems trained on scientific data rather than prose, though generative language models now sit alongside them, reading the literature, writing code and explaining results to non-specialists.
Critics are right that those benefits do not license any particular practice, including how training data is acquired. But the logic cuts both ways: an aggregate energy case against “AI” cannot be pinned on publishers without applying the same test to print runs, returns and remaindering, freight and conference air miles.
So the test is consistency. Anyone raising carbon against AI in publishing should be asking the same of the rest of the supply chain, and a house confident in its position would publish its figures for both.
Not gonna happen.
The angle Wired buried
For much of the world, the most consequential detail is a throwaway line. Editors say they have seen LLM-translated manuscripts, with AI-use disclosures, at the Frankfurt and London book fairs, and HarperCollins’s CEO has said AI could expand translated books and audiobooks.
The rights trade has long run through a few languages and a few gatekeepers. Cheap machine translation could let publishers in Africa, South Asia and Southeast Asia sell rights into English at a fraction of the old cost. It could equally flood buyers’ desks with cheap machine-translated product and squeeze out human translators. Whether disclosure labels on translations become standard matters far more than back-cover copy.
One more detail for the Gulf watchers: Hachette staff expect AI policy to “evolve” under a new chief information officer who previously ran technology at SRMG, a Saudi state-backed company that has publicly embraced agentic AI.
What publishers should do
First, publish the policy: which tools, for what, under which contractual terms, and a flat rule on unpublished manuscripts.
Second, apply one standard. If AI use matters enough to cancel an author’s book, it matters enough to disclose inside the house. If it does not, stop cancelling books over tools.
Third, judge output on quality and desirability, which is what readers have always done and what no disclosure label can replace.
Readers will decide, as they always have. But we may not like answer.
This post first appeared in the TNPS LinkedIn Analysis newsletter
