Context Window 55
This edition covers a methodological dispute over Karen Hao’s Empire of AI and the wider problem of poor sustainability reporting from AI companies, an audacious AI music-generation claim from Suno, OpenAI’s forecast of 220 million paid ChatGPT subscribers by 2030, fresh disclosure pressure on OpenAI in the authors’ copyright suit, Creative PEC analysis of UK job ads showing complementary demand for AI and creative skills, Tim Sweeney’s argument against tagging AI-made games, Mistral’s new range of small, on-device models, and new research on how sentence structure can bypass AI safety rails.
This was a week in which AI refused to behave according to anyone’s preferred narrative: strong environmental claims unravelled, legal cases tightened, policies were challenged, and AI models themselves showed new points of fragility. I was particularly interested in new research that challenges a common false dichotomy between AI use and creativity. Talking to publishers, the environmental impact of AI is regularly cited as one of the major reasons not to use it—and it is one of the central criticisms in Karen Hao’s bestselling book Empire of AI. But researcher Andy Masley recently highlighted significant methodological errors in Hao’s calculations, which overstate water usage by several orders of magnitude. In response to Masley, Hao has acknowledged the error and will correct the book. The fine detail of this is quite abstruse, but for me the takeaways are simpler. The response to the controversy is a Rorschach blot for how people feel about AI. Many of the people commenting in support of Masley condemned Hao for being ideologically anti-AI. In turn, some of Hao’s supporters dismissed Masley out of hand for his links to the Effective Altruism movement. Reality is more nuanced than either characterisation: Hao may be directionally right even if some claims are discredited, while Masley’s analysis highlights an important epistemic risk, namely, how easy it is to take weak analysis at face value when it aligns with a preconceived worldview. Second, the uncertainty over environmental impact exists in large part because the standard of sustainability reporting from AI companies is so poor—it was interesting to consider this in the context of Satya Nadella’s comments this week that tech companies need to earn social permission for their resource use, a standard the AI sector is currently failing to meet. A fundraising document for the AI music platform Suno claims it is generating the equivalent of Spotify’s entire catalogue every two weeks. I’d love to know the equivalent number for generation of self-published books; Amazon/KDP is probably the only organisation with data robust enough to attempt that calculation, and they don’t publish it. It’s also worth noting that music’s subscription economics make this sort of volume easier to monetise than in books, which remain largely sold a la carte. So if this isn’t happening at this scale in publishing yet, it’s not because it would be technically difficult—it’s because the incentives haven’t lined up. The Information (paywalled) reported that OpenAI is forecasting 220 million paid subscribers to ChatGPT by 2030, which it compares favourably to Spotify’s 280 million paid users. I’d suggest a better comparison is with Microsoft Office, which has just shy of 500 million paying users: to get to just under half of that in less than a decade would imply a remarkable level of ubiquity. Whether the forecast is achievable, or enough to justify OpenAI’s stratospheric valuation, is another matter. However, not everything is going OpenAI’s way: a court ordered that internal communications related to the company’s use of pirated books for training, which OpenAI argued were legally privileged, should be revealed to authors suing the company. OpenAI’s in-house legal team will now be deposed. As I noted after earlier disclosure losses, there is a long road to resolution of this litigation. But Bartz v. Anthropic set an important precedent: a court suggested that training on lawfully acquired books may qualify as fair use, but obtaining books from pirated libraries may be copyright infringement. If OpenAI’s disclosures reveal extensive use of illicit material, pressure for a settlement would likely increase. A major new analysis of 168 million UK job ads finds that demand for AI skills and creativity skills is rising together, especially since the release of ChatGPT, with the strongest overlap appearing in highly-skilled roles and the UK’s established creative clusters. For publishers, this reinforces a point many of us are already seeing in practice: AI skills are becoming a complement to good creative judgement, not a replacement. It also suggests that publishers who invest in both sets of skills, rather than treating them as competing priorities, will be better positioned as GenAI becomes embedded across publishing workflows. Epic Games CEO Tim Sweeney started a debate in the games business by arguing that games marketplaces like Steam shouldn’t label games made using AI, because it will soon be part of how all games are made. Consumer signalling is an equally hot issue in books, though no-one is asking for less transparency, with at least four organisations offering an AI-free checkmark for book covers. Mistral launched a new range of models. On a like-for-like basis, they are not really competitive with larger models from OpenAI, Google or Anthropic, but the interesting aspect is some of the smaller models, designed to be run locally on smartphones or laptops without an active Internet connection, which reportedly benchmark in line with competitor models four times their size. Researchers from MIT, Northeastern and Meta have shown that some LLMs can over-rely on sentence structure rather than meaning, answering questions based on familiar grammatical “shapes” even when the words are nonsense. For book publishers, this helps explain why models sometimes hallucinate confidently in unfamiliar contexts, misinterpret bibliographic or rights information, or respond erratically to out-of-domain material. It also highlights a potential safety and quality risk when AI is used in workflows such as metadata creation, research, fact-checking or customer-facing discovery tools, where subtle shifts in phrasing can produce unreliable results.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.