Context Window 70
I’m sending this week’s newsletter a day early, as Friday and Monday are public holidays here in the UK—judging by the out-of-office messages I’ve seen this morning, it looks as though a lot of publishing has already checked out for the long weekend. But if you’re still at your desk, some reading matter to take you into the Easter holiday: there’s a lot going on, and a surprising number of strategy signals in this week’s news.* The latest copyright lawsuit to be filed is a fascinating one: Penguin Random House is suing OpenAI in a German court for copyright infringement. The biggest trade publisher in the world taking that step is notable in its own right, but the interesting aspect is that PRH is, via parent Bertelsmann’s strategic partnership, an OpenAI user as well.
This points to one of two things: either another OpenAI media partnership unwinding, or a publisher taking a pragmatic enough view to use the product set while simultaneously litigating against its developer. Maybe Coopfrontation should be the title of my airport business book.
OpenAI is probably good for any potential damages, having just raised $122 billion of funding on an $852 billion valuation—the largest raise in venture history. The company also revealed monthly revenue is now $2 billion, 40% of which comes from enterprise customers like Bertelsmann.
Digiday has details of the Guardian’s first AI product: Storylines uses an LLM to curate the most relevant related stories on a topic (something that’s highly relevant to any publisher looking to drive readership and cross-sales). It’s also interesting that they made a deliberate decision to work off headlines alone, not full story content, to reduce the risk of hallucinations: good design is as much about context and constraint as capability.
MCP is becoming one of the foundational elements of a smart publisher AI strategy, but that exposes it to the risk of becoming a management fad: as one good, recent piece puts it, “We’re Building an MCP Server” Is the New “We Hired a Data Scientist”.
For authors and publishers looking to use AI as a research tool, Microsoft shipped an update to its Copilot Deep Research agent with two new multi-model features, Critique and Council. Critique uses one AI model to generate research on a topic, and others to critically review it. Council brings perspectives from different models together. I haven’t always thought of Copilot as a leader in this space, but this is a potentially very useful approach to improving the quality of research.
I took part in a webinar hosted by the Crius Group earlier in the week, on a panel with my friends Cameron Drew and Simon Mellins. One of the topics that came up was common failure patterns in AI: I wrote a quick blog post on the topic with the assistance of my webinar notes and Claude (see below), though as I went through it, it became clear that these issues are by no means restricted to publishing.
Friend and subscriber Alex Boden hosted an inaugural webinar for his data publisher Asymmetrix, and the notes in his newsletter this week are very relevant to other areas of publishing: the emerging strategic discussion centres on defensibility through proprietary data and rights, the difficulty of AI replicating high-value or high-stakes content, and the growing importance of human expertise and workflow integration in sustaining premium positioning.
With an early newsletter publication this week, I haven’t had time to form a considered response to Jonathan Woahn’s latest post on AI and content marketplaces—I’ll come back to it next week, but you should read the piece in the meantime.
A follow-up to last week’s research data from HEPI on the use of AI in UK education: the Lumina Foundation and Gallup released similar research on use of AI by college students in the US. The studies are not directly comparable, but on a directional basis, suggest that AI is embedded in both systems but that usage may be slightly higher in the UK.
I queried the accuracy of AI detectors last week in the aftermath of the Mia Ballard imbroglio. Max Spero of Pangram, which had an outsized role in making story prominent, replied to my questions on AI detectors and neurodiversity. I accept his point on absence of evidence, but essentially we seem to be working off anecdotal evidence in both directions.
As a small, limited experiment, I curated and wrote this week’s newsletter by hand as usual, adding one section which ChatGPT rewrote (see if you can guess which), and put it into Pangram: 100% human. Similarly, my blog post above was outlined by Claude from my original notes and then written by me: the result, 100% AI. Your mileage may vary, but these are very definitive figures for what seems at best a directional class of tool.
Finally, to end on a lighter note as we head into a holiday, Tom Fishburne nails the practical problem with human-in-the-loop review of AI outputs…
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.