Context Window 74
Happy Friday. I’m catching up after a week on the road and there is a lot to process. This week my reading and thinking clustered around two themes: what happens when the cost of producing content falls, and whether that cost will be stable over time.
I am just back from the ECPA Leadership Summit in Chicago, where I gave a keynote on the past, present and future of publishing. I wrote up my impressions of the event here, including a particularly useful framing of AI writing as a proxy failure raised by one of the other speakers, which I then develop from an economic perspective.
On that point about writing and authenticity, MIT Press published a thoughtful book extract on the long history of machine-writing anxiety, from Roald Dahl and Calvino to ELIZA and ChatGPT. Its most useful point for publishers is that today’s fears are not misplaced, but that they are less unprecedented than they feel.
Books are clearly the focus of concern for many subscribers to this newsletter, but for another frame on questions of authenticity and authority, we can look at the web. New research suggests over a third of new websites appear to use AI-generated or assisted content, that this is contributing to semantic homogenisation, but that there is no statistically significant evidence for accuracy dropping off. I’d caveat that this relies on software detection, but the results are directionally very interesting. If the finding holds, the risk isn’t that AI makes content wrong—it’s that it makes it indistinguishable. For publishers, that shifts the competitive question from accuracy towards differentiation.
A new paper published by the AI Task Force of the academic journal Organisation Science analyses five years of submission data, showing that the rate of submission and the rate of AI usage in submissions and peer review have grown sharply. Interestingly, the researchers associated high rates of AI submissions with institutions that are highly focused on publication activity and journal rankings as metrics.
That “publish or perish” dynamic is well known in academia, but it’s interesting to contrast it with, say, self-published authors and KDP. In other words, AI adoption may be less about the technology itself and more about the incentive structures around it. Where output is tightly measured and rewarded, AI use accelerates—regardless of concerns about quality or authorship.
Last week, I wrote about the cost/benefit ratio of Claude Code and Anthropic experimenting with price tiers. This piece in The Verge goes further, arguing that the current pricing of AI models versus the underlying cost of providing those tools is unsustainable. If that’s right, it has a particular impact on smaller publishers who are getting the (subsidised) deal of a lifetime at the moment, but might struggle to justify higher subscription tiers. The practical implication is that today’s AI workflows should be built with cost volatility in mind—model portability, usage tracking and fallback options are not nice-to-haves.
Relatedly, Adam Hyde published some useful token cost benchmarks for using different AI models for scientific publishing workflows, arguing that publishers could manage costs by building model-neutral systems which could be switched to different LLMs. Hyde’s own platform is running on the Chinese model DeepSeek, which costs a fraction of the latest OpenAI and Anthropic models.
Meta’s proposed acquisition of agentic AI platform Manus has been blocked by the Chinese government. This is a complex picture: Manus was founded in China, but headquartered in Singapore, and its website already states that it is “part of Meta,” which may look a touch hubristic. For publishers, this is a reminder that control of AI models is increasingly political.
Ethan Mollick highlighted an interesting aspect of GPT-5.5, OpenAI’s latest reasoning model: the developer documentation recommends shorter, goal-oriented prompts rather than detailed instructions. The second of my two blog posts this week looks at what that means for users, and suggests based on a historical parallel that for that to work, context and trust are going to be key. It’s a small change in interface but a big change in responsibility.
Adobe’s creative Firefly AI Assistant is now available in public beta: like GPT-5.5, this envisions users focusing on their desired output and providing iterative feedback, and the agent working behind the scenes to deliver that vision using the full range of Adobe software. With a busy period of travel, I haven’t had time to experiment with this yet, but I would love to hear from anyone who has.
In other tech developments, Google added native file creation capabilities to Gemini, which can now generate Google Workspace files, Microsoft Office documents, PDFs, TXT, CSV, Markdown and LaTeX from within chats. No EPUB support yet—though Workspace and Markdown both give simple pathways to EPUB creation—but as other Google AI products support the format, one can live in hope of it coming with a future update.
Thanks to Will Crook from PLS for drawing my attention to IFRRO’s newly published AI Governance Guide: for anyone on the legal and rights side of publishing, this has useful context on training, licensing and labelling of content.
Warmest congratulations to Burleigh Dodds Science Publishing for winning the AI category at the Independent Publishing Awards for the second year running. As I work with both BDS and the IPG I should declare that I had no role in the judging, but I’m very happy for my colleagues, and congratulations to everyone who was shortlisted or won other categories.
Finally, we’re probably all used to LLMs being people pleasers and agreeing with everything we say, so this response from Claude to one of my requests this week was a bit of a change In all humility, I have to admit that having rethought my prompt, it wasn’t entirely wrong. But I’m not standing on the other side of the pod bay doors from it anytime soon…
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.