Context Window 38
This edition covers fresh data on student AI use and ChatGPT’s new Socratic Study Mode, NotebookLM’s new multilingual audio and video overviews, Adobe’s Firefly-powered Photoshop upgrades, content licensing deals from Gannett/Perplexity and Johns Hopkins University Press with benchmark pricing for academic books, Neil Perkin on where AI still falls short, fresh research on the scale of “Shadow AI” use in workplaces, and Ethan Mollick on the “Bitter Lesson” and outcome-trained AI in messy organisations.
Salvete lectores. For the last ten months I’ve been figuring out what subjects resonate most strongly with subscribers. The missing element was clearly the classics: I had more email from you about a joking reference to the Cambridge Latin Course last week than any of the more controversial AI topics discussed in the last year. I enjoyed every one of them. Please keep your questions and feedback coming. AI is becoming woven into the fabric of education: quantitatively, Kortext/HEPI data shows 92% of students using AI, and this recent piece looking at three students’ use of ChatGPT over eighteen months gives a more qualitative view. So it is interesting to see ChatGPT responding with a new feature, Study Mode. It’s designed to work Socratically with the learner, guiding them through a topic. Of course, it depends on the user wanting to learn that way, as opposed to just getting a quick answer… For publishers, Study Mode’s potential impacts include new formats for interactive textbook companions or assessment tools, but uncertainty remains about whether ChatGPT will directly license external educational content into its outputs. But in addition to how it reshapes educational content, it’s also a really useful tool for lifelong learning, and something we could all use for professional development. Google’s AI research assistant NotebookLM has added new content creation features, including multilingual audio overviews and the ability to create video overviews of material. This is quickly becoming one of my most-used AI services, and for anyone looking to create ancillary content around their books or research (e.g. downloadable resources), it’s a great tool. Adobe announced a range of new AI features for Photoshop, powered by its Firefly image model. This includes enhanced Harmonisation and Remove tools, more AI model options when using Generative Fill and Generative Expand, and a beta of their Generative Upscale tool. This is particularly valuable for illustrated publishers needing to upscale archival photos or original artworks to modern print standards, solving a common and costly design problem. Two new licensing deals this week. US news publisher Gannett has announced a strategic partnership with Perplexity. And in academic publishing, Johns Hopkins University Press has approached authors asking them to opt into a licensing deal, though the exact tech partners and usage aren’t clear from reporting. The key takeaways here are that like other university presses and trade houses, JHUP felt the need to consult authors, as opposed to the more executive decision making at Wiley. And there is a useful benchmark in the reporting: JHUP is suggesting authors might expect $100 per book per license, significantly less than the $2,500 HarperCollins offered non-fiction authors, though trade publishing is not a direct comparator. Neil Perkin has a great piece this week asking a key question: what is AI not good at? Let me know what you think, but I found this very reassuring in that many of the areas where AI models struggle are those where I’ve seen great publishers excel: context, cultural nuance and judgement. A new piece of terminology for an established problem: Shadow AI, or unauthorised AI tools being used in the workplace. New research from US and Canadian businesses paints a concerning picture: 93% of employees have uploaded information into AI tools without approval; 91% of employees believe such unsanctioned use is either without risk, or that the benefits outweigh any risk; 61% are using unapproved tools more than they were a year ago; and 60% of employees want fair and practical policies, with 66% wanting better education on understanding the risks. Publishers must urgently define clear, pragmatic AI usage policies and provide practical education to employees to mitigate growing risks. A fascinating new blog post by Ethan Mollick explores why enterprise AI adoption is so hard—and what might make it radically easier. Drawing on Ruthanne Huising’s research into organizational messiness and Richard Sutton’s “Bitter Lesson” in AI, Mollick argues that many companies struggle to scale AI because their internal processes are undocumented, bespoke, or just plain irrational. But here’s the twist: if generative AI can be trained on outcomes rather than processes, organizations might no longer need to untangle the chaos—just define what good looks like. For publishers, this could shift the focus from automating workflows to training AI on high-quality editorial outputs. The real challenge may not be operational complexity, but clarity of vision.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.