Context Window 43
This week’s practical insight focuses on two linked topics: onboarding and oversight of AI models. But there are also more philosophical questions from news publishing and Seth Godin, and details of the most open LLM to date. Have a great weekend. Sometimes it’s the simplest insights that stick. In a recent webinar with Rich Ziade and Paul Ford of AI developer Aboard, Rich said something that stopped me in my tracks: “AI is a new hire.” When you onboard a new team member, you think carefully about their role—who they’ll work with, what systems they’ll need access to, how they’ll get up to speed. In a hybrid workplace, we’ve learned that structure and documentation matter more than ever. But with AI, I’ve seen people open up a general LLM with no context or preparation, and then feel disappointed by the results. It’s the equivalent of throwing a new hire in at the deep end and hoping for the best. Here’s my start on a checklist to avoid that. Have you done each of these things for your AI projects? And let me know if there’s anything you’d add to the list.
- Clarify the role: what tasks is this AI responsible for, what outcomes define success, and who will work with it?
- Grant the right access: what data, documents or systems does it need to access, what information would not be available through its general training data, and what should be restricted?
- Share onboarding materials: brand voice, style guides, catalogues, examples of good work, plus glossaries, common terms, key people and projects. Like any new team member, AI needs oversight. Human-in-the-loop is one of the key principles in AI strategy: anyone who’s attended one of my training sessions has heard me say the phrase a couple of dozen times. But not every AI use has the same level of risk. It’s not always clear what appropriate levels of control look like. This superb, practical framework breaks it down by end goal and risk level, identifying sixteen levels of control and autonomy, and explaining exactly how humans should be involved in each. It’s an invaluable reference for any publisher developing AI use cases. Semafor’s Gina Chua, the incoming director of the Tow-Knight Center at CUNY, has an interesting piece exploring the impact of AI on news. Her central question is what happens to news publishing if value shifts from a fixed output to something that AI personalises to an end user (Nieman Lab has a good round-up of experiments with conversational news here). It’s a fundamental point, and one with considerable relevance to publishers of all sorts—as a thought experiment, I’d be really interested in how non-fiction book publishers think this through. As Gina Chua puts it: “Will our value be increasingly measured in the questions we ask, the facts we gather, the insights we have, and the relationships we have with our audience, rather than in our words?” The Swiss sovereign LLM that I wrote about in July now has a name: Apertus, as its name suggests, is fully open: models, weights, training data and documentation are all accessible, and it promises compliance both with Swiss national law and the EU AI Act. This is looking like a potentially compelling alternative for European publishers. Seth Godin’s writing is always highly engaging and provocative: he marked the Labor Day holiday with a personal perspective on the dilemma that writers face with AI—walk away, or dance. Reports this week described a troubling new dimension in ransomware attacks: hackers who took over an art website threatened to provide all of the artists’ content they had stolen to LLM training sets unless a ransom was paid (though how they would do so is not entirely clear). Given the relative ease of working with text files, a varying degree of security awareness across the industry (I know of agents and smaller publishers running everything through personal Gmail accounts) and previous cases of pre-publication manuscripts being obtained through phishing attacks, it’s a good reminder to check that your information security policies and procedures are sound. There’s been a lot of discussion about the impact of AI on web traffic, but in this detailed piece, Dan Petrovic argues that SEO has never been more important as ChatGPT 5 is engineered to reason and refer to external data sources, not to store information. For context and data, it relies on live web searches, and the visibility of publishers’ content depends on how good their SEO is. The levers for publishers are well defined, but in many cases there’s still scope for publishers to get the basics of SEO right. On this topic, I’m going to be taking part in a panel on SEO, GEO and the web at the Independent Publishers Guild autumn conference on 16 September—if you’re in London, I hope to see you there. For academic readers, this is a sobering first person essay exploring the impact of AI on peer review processes through the experience of reviewing a suspect submission. It’s interesting not just in terms of the technology, but of the structural incentives that lead to people using it—and beyond academia, I suspect the author’s experience will be replicated by a lot of commissioning editors in the coming year. Finally, if you read that opinion piece in the trade press yesterday on the books business redefining its purpose in an AI age, I took issue with it here.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.