Context Window 67
I’m catching up after three intense days at London Book Fair. That intensity has been mirrored in the AI world, with a slew of copyright and licensing developments, some new practical tools for publishers, and catastrophically misjudged AI developments at Grammarly, Amazon and McKinsey.* I wrote up my reflections from LBF here, including some thoughts on AI at the Fair: more businesslike, but unevenly distributed. What struck me most was the gap between the publishers talking about it abstractly and the people using it fluently.
It was a big week for copyright and AI, building towards the UK government’s next update on the topic, due by 18 March—though apparently well-informed reporting suggests that the government intends to kick the problem down the road. A delay buys time politically, but it prolongs uncertainty for rights holders and AI companies alike.
The most visible aspect of this issue on stands around the Fair was the Don’t Steal This Book publicity stunt, but there were other consequential developments. More substantively, Oliver & Ohlbaum released a new economic assessment commissioned by the News Media Association, debunking recent lobbying claims on copyright by the tech industry. And Publishers Licensing Services launched its opt-in, collective licensing scheme for Generative AI training.
Outside the Fair, the big AI story was writing platform Grammarly offering ‘expert AI reviews’ in the persona of various writers, journalists and academics, alive or dead (in some cases, very recently dead) who were not consulted. The company went through several rounds of poorly worded communication before pulling the feature. They are now facing a class action lawsuit. I try to take a balanced and pragmatic view of this wider area, but the legal risk here was obvious; the reputational risk even more so. This was a grossly misjudged product and Grammarly deserves every dollar it will lose in the ‘finding out’ part of FAFO.
Related: Ingrid Burrington’s genius description of Grammarly’s editorial AI agents as ‘sloppelgangers.’ No notes.
Microsoft announced Copilot Cowork, an agentic AI tool that works across Copilot and other Microsoft applications. As I commented on my LBF piece, many publishers I’ve spoken to end up with Copilot as their default AI tool because it’s integrated with email and documents. We’ll have to see whether Copilot Cowork performs as well as its Anthropic namesake when it rolls out to customers later this month.
Claude Code added support for running tasks on a predetermined schedule, useful for any publishers looking to use it for regular workflows. Small feature, big implication: once tasks can be reliably scheduled and delegated, AI starts to look less like an assistant and more like background infrastructure.
Hosting provider Cloudflare introduced a new endpoint allowing users to crawl an entire website with a single API call. This is something more associated with AI companies than publishers. But for publishers thinking about metadata, crawl controls, and AI discoverability, the ability to review a site’s structure and content programmatically is way more relevant than it might first appear.
Even the biggest companies need guardrails for their use of AI. Amazon, which previously leaned in hard to AI agents, is reportedly introducing a 90-day safety reset after an AI coding tool led to 6.3 million orders being lost in a day-long outage across North America earlier this month. That’s potentially a lot of books. And McKinsey’s internal chatbot, including confidential messages and client files, was pwned by an AI agent. These stories provide a salutary reminder that speed without governance scales mistakes just as efficiently as it can scale productivity.
I’m off to Liverpool next week to give the Wednesday morning keynote at the ALPSP University Press Redux Conference—I hope to see some of you there.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.