The industry does not need to resolve every open legal question, or abandon every reservation about how its content might be used, to start acting. It needs to stop waiting for the fight to be won outright before it starts negotiating the terms of what comes after.
For most of the past three years, the oh-so comforting story in book publishing has gone something like this: generative AI is entirely built on theft, the courts will eventually sort it out, and if the industry simply holds its nerve, the “threat” will recede and AI will disappear.
After all, AI can only produce slop that no-one wants and will never pay for, and AI can never match human creativity at any level (us “professionals” can spot AI-created content at a glance), so really there is no actual threat. We just have a duty to protect the ignorant general public who don’t share our finely tuned powers of discernment.
Let me bring in the Monet saga here. Per an earlier TNPS op-ed, someone posted an AI-created copy of a Monet masterpiece online and invited comment. The painting was, by near-universal verdict, a perfect illustration of AI’s aesthetic bankruptcy: derivative, soulless, technically impressive in a hollow way, proof if proof were needed that machines cannot create.
It was not made clear how many of the respondents were in publishing, but the response suggests many may have been.
Only problem being, the painting was a Monet. A genuine one.
The critics had not responded to the work. They had responded to the label. The label ‘AI-generated’ activated a set of preloaded judgments so powerful that they overrode the evidence in front of almost every pair of eyes.
The event highlighted a psychological phenomenon known as post-hoc rationalisation, or confirmation bias. When people are told a piece of media is AI-generated, their brains instinctively look for flaws to support their pre-existing biases against artificial intelligence. The experiment proved that audiences frequently judge the origin and provenance of an artwork rather than its actual visual quality.
Repeat for AI generated text, books or images. We convince ourselves, by constant repetition and peer-group reassurance ,that something is true or not true, depending on what case best suits our needs.
So we tell ourselves AI slop is everywhere, no-one wants it, and by the way, AI is stealing all our jobs because everyone is buying AI content.
Some publishers still believe it. But more and more publishers are coming to terms with AI, and realising none of that existential threat nonsense is actually true. But it sounded good a the time. The sky is always falling in publishing land.
Getty Images believed a version of that story too. It had more reason than any book publisher to believe it, and more resources than most to test it in court.
It tested it.
It lost.
And what it did next is the part of this story that book publishing needs to take a close look at.
Not the usual sky-is-falling moment
Book publishing’s anxiety about AI has, until now, been mostly anticipatory – a sense that something disruptive is coming, argued out in op-eds and conference panels around the globe, even as the actual business of acquiring, editing and selling books carries on much as before.
It is possible to overstate the threat to books, because long-form, edited, fact-checked narrative text is hard to fake convincingly at scale, and yes, readers still generally know the difference. Whether they care or not is another question. The 50-million minutes of AI-created audio content listened to over at Audible may, as Chris Kling noted, be a tiny fraction, but what matters for this argument is that it is happening at all. There has been no mass protest against AI from consumers.
But back to Getty Images, faced with the worrying problem that image-generation AI has improved from crude and obviously synthetic to commercially usable in the space of about two years, and it is still improving weekly.
Yes, we can all repeat the stories from 2023-4 when six-fingered, long-necked characters were the hall-mark of an AI-produced image. Not any more.
A publisher can reasonably tell itself that nobody’s ChatGPT-written novel is about to outsell their frontlist. Getty could not tell itself that nobody’s Stable Diffusion output would substitute for a stock photograph – because it already had. This was not a hypothetical erosion of value. It was underway, priced into the stock, and visible in the numbers before any deal was signed.
Publishers, keep that in mind as we read what follows. Becuase Getty’s response wasn’t a luxury option chosen from a position of comfort. It was survival strategy, executed under real pressure, and that is precisely what makes it instructive rather than merely admirable.
Getty fought first – just like publishing
Let me be clear on the Getty history here, because the industry narrative has quietly rewritten this as foresight when it was, for a long time, resistance.
In January 2023, Getty sued Stability AI in the UK, alleging that its images had been scraped without permission to train Stable Diffusion, and pursuing claims of primary and secondary copyright infringement, database right infringement, trade mark infringement and passing off. It was, in other words, exactly the posture book publishers and authors’ organisations have taken against OpenAI, Meta and others: assert the rights, go to court, wait for vindication.
Only, the vindication didn’t fully arrive. By the time the case reached trial in mid-2025, Getty had conceded there was no evidence that Stable Diffusion’s training had actually taken place in the UK, and abandoned its primary copyright and database right claims outright.
When the High Court delivered judgment on 4 November 2025, it rejected Getty’s remaining secondary copyright infringement claim, ruling that Stable Diffusion’s model weights did not themselves store or reproduce copies of Getty’s images in the way UK copyright law requires.
Getty salvaged only a narrow trade mark finding, related to watermarks appearing in some early outputs – a fraction of what it had originally sought.
This inconsiderate detail is what gets lost in the tidier version of the Getty story, but it’s the detail that makes the comparison to publishing seriously useful.
Getty didn’t out-negotiate the courts and then magnanimously choose partnership. It took its best legal shot, at real expense, and came away with a partial and largely symbolic win. The pivot to licensing that followed wasn’t the reward for winning. It was what remained once litigation had shown its limits.
The swing: from courtroom to contract
Less than a year before the Stability AI judgment, Getty had already started building the alternative. In October 2025, it struck a global multi-year licensing agreement with Perplexity, giving the AI search platform access to Getty’s creative and editorial imagery through its API, with an explicit condition: Perplexity would improve how it credited and linked back to the images it displayed.
Then, on 21 June 2026 – after losing the bulk of its case against Stability AI – Getty announced a comparable display partnership with OpenAI, bringing its licensed libraries into ChatGPT’s search and discovery experience.
Both deals share a structure, that is the single most important nuance for publishers to absorb: they are display agreements, not training agreements. OpenAI is not being licensed to use Getty’s photographs to build or refine its models. It is paying to show Getty’s images, with attribution, when they answer a user’s query. Perplexity’s deal is built the same way. Getty protected the one asset that actually frightened it – the prospect of its archive being absorbed wholesale into someone else’s model – while opening a new, recurring, low-friction revenue line built on exactly the trust and provenance that anonymous scraped imagery cannot offer.
Neither company disclosed financial terms. The market reaction to the OpenAI deal, while positive, is a messy story in its own right – reported share-price surges ranged from roughly 90 to over 200 per cent across different outlets, an inconsistency that says as much about the state of financial reporting on fast-moving AI news as it does about the deal itself.
But the direction of the reaction was unambiguous. A stock trading near penny-stock territory earlier in the year, with a Q1 2026 net loss of $4.4 million (a marked improvement, in fairness, on the $102.6 million loss a year earlier), jumped sharply on a deal that generated no disclosed revenue guidance whatsoever.
Investors weren’t pricing in the details. They were pricing in the model.
Why the training-versus-display line matters more than it sounds
It would be easy for book publishers to read “display agreement” and assume it’s a lesser, more cautious cousin of the training deals that News Corp, the Associated Press and Axel Springer have already signed with OpenAI and others – some of them worth hundreds of millions of dollars.
But that would be to misread what’s actually being demonstrated.
Getty’s insight is that there are at least two entirely separate products an AI company might want to buy from a rights holder, and they carry different risks and different prices.
Training rights are the more familiar territory: a one-off or recurring fee in exchange for a model absorbing your content into its parameters, after which your control over how it’s used effectively ends.
Display and retrieval rights are different in kind. The content stays external to the model, is surfaced only at the point a user asks a relevant question, and can be attributed, tracked and – in principle – renegotiated. For a company as protective of its archive as Getty, that distinction was the entire point.
Book publishers weighing their own AI strategy might consder not treating “should we license to AI companies?” as a single yes-or-no question. It isn’t. A publisher might reasonably refuse to license a backlist for model training while actively pursuing retrieval and display deals that put verified, attributed excerpts of their books in front of readers asking AI tools for recommendations, summaries or reference information.
Those are different products, sold on different terms, and conflating them is how publishers end up either giving away more than they intended or refusing opportunities that carry little of the risk they’re worried about.
The wider pattern Getty confirms rather than creates
Getty is a useful test case precisely because it isn’t an outlier. News Corp signed a five-year, $250 million-plus deal bringing The Wall Street Journal and The Times into OpenAI’s ecosystem.
Axel Springer, Le Monde and Dotdash Meredith have struck comparable arrangements.
Reddit is being paid by Google for access to its user discussions.
In publishing specifically, HarperCollins opened an opt-in AI training arrangement for nonfiction backlist titles, and both Wiley and Taylor & Francis have signed multimillion-dollar agreements with AI developers.
I might add there that the latter two were not always with full author consultation, which is its own cautionary tale about how not to do this. Or how to make millions without rewarding the authors. But to be clear, that’s an issue of publishing contracts, not the fault of AI.
The throughline across all of it is the same one Getty’s pivot makes concrete: AI companies building consumer-facing products have discovered that scraped, unverified content is a legal and reputational liability, at least where using pirated content is concerned, and that paying for authenticated, well-described, rights-cleared material is now cheaper than the alternative.
That is a- at risk of using an “AI word” – genuinely new commercial fact, not a talking point. It didn’t exist in this form two years ago.
What this actually asks of publishers
None of this requires book publishing to declare itself suddenly enthusiastic about AI (in time honoured tradition we shall continue to decry the new tech until the day we fully embrace it, and then we’ll pretend we always loved it);
And there’s no need to pretend the industry’s caution has been groundless – plenty of publishers have legitimate, unresolved concerns about author consent, compensation and the terms on which their backlists might be used.
But there is a difference between principled caution. Luddite resistance and simple inertia, and Getty’s experience draws that line clearly.
A few practical steps follow directly from it:
Audit before negotiating. Getty could only strike these deals because it controlled its archive cleanly. Publishers who haven’t confirmed exactly what AI-related rights their author contracts do and don’t grant are not in a position to negotiate anything, on any terms, when the offer arrives.
Separate the products. Treat training, retrieval and display as three distinct commercial questions, not one referendum on “AI.” A publisher can decline one and pursue another without contradiction.
Insist on attribution as a term, not a courtesy. Perplexity’s commitment to image credits and source links wasn’t incidental to the Getty deal; Getty made it central. Publishers negotiating retrieval or display arrangements should treat visible, verifiable attribution back to the source book or publisher as non-negotiable, both for author trust and for driving readers back to the point of sale.
Read the legal record honestly. The Getty v Stability AI judgment was not a win for rights holders – itwas a warning that the current copyright framework, built for a pre-AI world, does not straightforwardly protect content once it has passed into a model’s training data.
The lesson, stated plainly
Getty Images did not thrive because it found a clever way to make AI companies pay tribute. It thrived because it stopped treating the courtroom as its only strategy the moment the courtroom stopped delivering, and because it had already done the unglamorous work – clean rights, structured metadata, a defensible archive – needed to negotiate from strength once it did.
Book publishing possesses exactly the kind of asset AI search platforms are short of: long-form, edited, fact-checked, professionally curated knowledge and entertaining narratives. The industry does not need to resolve every open legal question, or abandon every reservation about how its content might be used, to start acting on that fact. It needs to stop waiting for the fight to be won outright before it starts negotiating the terms of what comes after.
This post first appeared in the TNPS LinkedIn Analysis Newsletter.
