HBO’s new Harry Potter TV series premieres this Christmas, expected to introduce the books to a new generation of readers – ker-ching! – at the same moment the company is banking real, recurring profit from licensing its catalogue to AI companies.
On July 15, China’s Cyberspace Administration switched off a piece of the AI economy that most of the publishing world never thinks about: companion chatbots.
ByteDance’s Doubao, Alibaba’s Qwen, and Tencent’s Yuanbao all pulled the plug on custom AI personas that day, complying with new rules – the Interim Measures for the Administration of AI Anthropomorphic Interactive Services – that ban emotionally sticky AI relationships for anyone under 18, and require anti-addiction safeguards for everyone else.
Millions of user-built companions went dark. Doubao users have until October 15 to export what they can before the data is deleted; Qwen isn’t offering even that.
It’s a story about consumer AI, not books. But it happened in the same week as a very different kind of AI story – one that publishing should be paying much closer attention to.
The steering wheel argument
Geoffrey Hinton, in a June interview on the Big Technology Podcast, made two claims that got most of the attention: that today’s AI systems are already conscious (in his view), and that corporations building AI are structurally incapable of prioritising safety over shareholder return – not through bad faith, but because fiduciary duty is a legal obligation that overrides voluntary commitments the moment they become expensive.
China’s July 15 shutdown became, in the retelling, his proof that regulation can actually steer AI behaviour rather than just react to it – something no Western jurisdiction I’m aware of has yet demonstrated at that level of specificity.
The consciousness debate is fascinating in its own right. Hinton is a Nobel Prize in Physics 2024 laureate and British-Canadian computer scientist and cognitive psychologist so don’t rush too quickly to dismiss that idea. Hinton’s functionalist case against Ted Chiang’s rebuttal (Chiang argues that attributing consciousness to AI conveniently lets developers off the hook on liability and moral-status questions) is worth a TNPS piece of its own.
But buried in the same argument is a claim with immediate, measurable relevance to publishing: Hinton’s “information collapse” thesis.
AI answer engines increasingly synthesise and surface publisher and news-organisation content without sending traffic back to the source. The Reuters Institute’s 2026 research has news executives projecting a 40% decline in search referrals over the next three years, driven by AI answer engines. If the organisations producing the source material can’t survive economically, the argument goes, the quality of what AI trains on eventually degrades too.
That’s the framing. In a world where the sky is always falling, it’s no surprise some publishers are playing the victim card. It’s in the job description.
Publishing’s actual experience of 2026 tells a more complicated – and far more interesting – story. It even involves Harry Potter, almost.
Bloomsbury: having it both ways
Bloomsbury Publishing is, basically, the cleanest possible counter-example to the “publishers as AI’s passive victims” narrative. And the timing could not be better for a case study: HBO’s new Harry Potter TV series premieres this Christmas, expected to introduce the books to a new generation of readers – ker-ching! – at the same moment the company is banking real, recurring profit from licensing its catalogue to AI companies.
The numbers: Bloomsbury’s FY2026 (year to February 2026) revenue came in at £325.9 million, with adjusted profit up 7% to £44.9 million. In its July 15 trading update – the exact same day as the China shutdown – Bloomsbury reaffirmed guidance for a record year: £354.2 million revenue, £49.9 million adjusted profit, citing continued AI licensing income as a named driver of its Academic & Professional division, alongside its 2024 acquisition of Rowman & Littlefield and a Google AI partnership struck last September.
Several reports note the Academic & Professional division is growing even excluding the AI licensing revenue – this isn’t a company propping up a declining core with AI money; it’s additive.
So a publisher whose consumer brand is inseparable from Harry Potter is less publicly building a parallel revenue stream by licensing its academic and professional content to the same technology many love to frame as an existential threat to publishing.
Why academic, and why not (just) trade
It’s tempting to read Bloomsbury’s pattern – AI money flowing through the academic arm, not the consumer/trade side – as evidence that academic content is simply what AI companies need most. That’s not quite right, and the distinction matters.
Here’s the thing: Academic and professional publishers typically hold consolidated, aggregatable rights: exclusive licences or outright copyright secured as a condition of publication, across large, structured corpora (journals, reference works, textbooks). That makes bulk licensing tractable in a way it rarely is for trade fiction and nonfiction, where rights sit fragmented across individual author-agent contracts, territorial splits, and reversion clauses. One deal with Wiley or Taylor & Francis can licence a huge, well-organised corpus in a single negotiation. There is no equivalent single counterparty for “all of trade fiction.”
The scale bears this out. Wiley reported $40 million in AI licensing revenue for fiscal 2025 (up from $23 million the year before), and $49 million for fiscal 2026, with net income jumping 163% to $221.6 million on essentially flat revenue – AI licensing was, as one analysis bluntly put it, not a side story but the story.
Informa’s Taylor & Francis struck a reported $10 million deal with Microsoft in 2024, without notifying authors beforehand – a move researchers called a “worrying precedent” for author consent in academic publishing. Wiley has since also partnered with Anthropic directly on scholarly research tools.
Trade can be licensed – HarperCollins, the first Big Five publisher to do so, struck a deal with Microsoft (reported, not officially named) in late 2024: $5,000 per nonfiction title for a three-year training licence, split 50/50 between author and publisher, with guardrails limiting AI output to 200 consecutive words or 5% of a book’s text, and an explicit pledge not to scrape pirated content.
But it required individual author opt-in, book by book – a fundamentally different, slower, more contested mechanism than an institutional bulk deal, and the Authors Guild has been openly critical of the 50/50 split, arguing the underlying rights belong to authors, not publishers. That interpretation of course will need a court of law to settle for sure.
Trade content, evolving and culturally current in a way academic content structurally can’t be, may well be more valuable per word to a frontier lab. It’s simply much harder to buy at scale – which is precisely why so much of it ended up being taken instead of licensed.
The other ledger: what piracy pays out
That’s the second half of the story. The Bartz v. Anthropic settlement – $1.5 billion, covering 482,460 works pirated from shadow libraries like Library Genesis, with roughly 92.77% of eligible works claimed – reached final approval on May 14, 2026, with pay-outs tentatively expected from around August. After plaintiffs’ lawyers cut their fee request from $300 million to $187.5 million, the estimated pay-out sits near $3,000 per claimed work, typically split 50/50 between author and publisher.
Those shadow libraries skew heavily toward general trade fiction and nonfiction – exactly the material that’s hardest to licence in bulk and easiest to find pre-scanned online. Which means the retrospective piracy pay-out is landing disproportionately with trade authors and publishers, running in parallel to (and largely separate from) the forward-looking bulk licensing deals flowing mostly through academic publishers.
Two very different relationships with AI money, sorted almost entirely by how tractable each sector’s rights infrastructure happens to be.
The pattern is escalating, not settling. In January 2026, music publishers led by Universal Music Group and Concord filed a separate $3 billion suit against Anthropic – alleging more than 20,000 pirated compositions, torrented and, they claim, re-seeded back to the public – naming Dario Amodei and Benjamin Mann personally as defendants. The court denied Anthropic’s motion to stay the case in April; a motion to dismiss is due in August.
Legal analysts see the size of the claim as leverage toward another Bartz-style mega-settlement. Note the word settlement. The likelihood of any of these cases getting legally confirmed one way or another is remote. And both sides gain, which means expect much more of this merry-go-round in the future.
The steering wheel, in publishing’s terms
But let’s bring it back to Hinton’s framing: regulation as a steering wheel, not a brake, because this is hugely important for publishing and for wider society.
China has just proved a government can direct AI behaviour with real specificity when it chooses to.
Publishing’s own experience in 2026 shows something adjacent – that where an industry already has consolidated rights infrastructure (academic publishing), it can steer AI money toward itself on its own terms, forward-looking and negotiated.
Where that infrastructure doesn’t exist (trade), publishers and authors are instead relying on courts to claw money back after the fact, for content already taken.
Bloomsbury is at the exact seam between those two worlds: a byword for one of trade publishing’s biggest franchises, quietly profiting from AI on the academic side while its consumer side waits, like the rest of trade, for licensing infrastructure – or litigation – to catch up.
This post first appeared in the TNPS LinkedIn Analysis Newsletter
