Context Window 92
The themes of this week’s newsletter are copyright and money: the Department of Justice weighed in behind OpenAI in its copyright litigation with the New York Times, and a similar interpretation of Fair Use sits behind a new academic initiative. Meanwhile, OpenAI announced a billion dollar revenue line, and Google introduced a new commercial channel for publishers.
The US Department of Justice filed a statement in support of OpenAI in the copyright infringement litigation brought by The New York Times against the AI platform, arguing that training on written works is Fair Use, “exceedingly transformative”, and that restricting it would “give a competitive advantage to foreign adversaries who are not so encumbered”.
It’s unclear what effect, if any, this will have on pre-trial motions, but it’s the clearest statement yet from an interventionist administration and will worry publishers involved in other litigation (on which subject, Anthropic faces yet another lawsuit this week, this time from music publishers).
Also on the copyright and usage front, HathiTrust announced a new five-year partnership with Northeastern and the Authors Alliance funded by the Mellon Foundation to use AI to improve discoverability and metadata in the HathiTrust corpus and to offer a legitimate source of training data for language models.
If I am reading this correctly, it treats the Fair Use of lawfully acquired books for training as settled, and there is potential for this to cut across publishers’ licensing activities—or as the press release characterises it, “corporate mediation” of access to data.
OpenAI announced that its advertising platform reached $1 billion in annualised revenue in less than 200 days: that is, nearly $3 million per day in sales. One of the prominent criticisms of AI companies has been that they cannot earn sufficient revenue to support their unprecedented valuations and capital investment, but after several years of apocalyptic predictions, it’s clear that there is a path to revenue.
That path is not always straightforward though. I’ve written before about OpenAI’s Prism, a research and editing environment for scientific authors. Now, less than nine months after it was introduced, the executive behind it is leaving the company and Prism is being folded into OpenAI’s general product strategy. Product churn has always been an issue in tech, with Google a particular byword for withdrawing popular services (I still miss Reader). But the combination of that culture and hyperscaling AI platforms means users need to think carefully about committing to products that might not see another winter.
A new revenue source for publishers this week: Google launched Expert Intelligence, a new feature allowing users to integrate books purchased from Google Play into Gemini Notebook. It launches with support from PRH, Bloomsbury, Macmillan, O’Reilly Media and other publishers.
Anthropic released a new learning programme, Claude Academy, with short courses, explainers and webinars for their products. I’ve been using AI tools for four years, training people in them for three, and I picked up some new tricks. Well worth a look if you use Claude.
We’ve all got used to reading about Pangram AI detection results over a long season of doubt about the authenticity and provenance of leading books. Wired profiles the company and its founder Max Spero, highlighting some interesting ethical and practical questions.
Thanks to Alex Jeffries for sharing MIT’s new report on AI in teaching, learning and research, which will be of interest to academic and educational readers in particular. Pace Pangram, it argues against reliance on AI detectors for practical and cultural reasons, highlighting the impact of false positives on, for example, neurodivergent students, and the possibility of creating an adversarial culture (lessons that apply to the publisher/author conversation). Longer term, if this report is right, we’re going to see a significant level of churn as curricula are adapted for AI learning and assessment, and that will create a content opportunity for publishers.
An interesting piece of data from Benedict Evans to contextualise the sustainability arguments around AI use. Annual US data centre water use: 17.4 billion gallons. That seems astronomically high until put in context of total US water usage: 102 trillion gallons. Data centres are less than a fiftieth of one percent of the total—and that covers the servers behind your Netflix binge or social media, not just the AI use you might be worried about.
Finally, not really a publishing thing but thanks to Tom Abba for sharing a very cool use of AI: Claude Fable solved a cipher in 44 minutes that had previously been unsolved for nearly 400 years. The conclusion is very relevant to anyone using AI, even if you’re not interested in cryptography: