Context Window 78
This week’s newsletter focuses on the economics of AI becoming more visible. Sometimes that’s positive for publishers: a new tool for estimating the value of licensing your content. Other perspectives are more challenging, particularly on the real cost of AI usage. For publishers, the lesson is that AI strategy must now include cost, value and platform-risk assumptions.
Setting aside announcements from big publishers about major (and by definition, exceptional) deals, there’s a general lack of data on the size of opportunity for licensing of publisher content to AI platforms. It’s great to see data platform Cashmere releasing a calculator to help publishers estimate potential licensing revenue. Even if the estimate is directional, playing with the variables is a very helpful way to explore what drives revenue. Ultimately, only the market can determine value. But it’s a welcome starting point. YouTube announced improved detection and labelling of AI-generated content, though it isn’t clear whether the automatic AI detection they refer to will just inform labelling or could in future also block problematic content. From a publishing perspective, the latter would be helpful in addressing the glut of audiobook copyright infringement on the platform. This announcement reinforces the impression that solving that infringement question is a question of will and priorities rather than technical feasibility. The publisher question is not really whether YouTube can detect more synthetic content. It is what the platform chooses to do once it can. Anthropic released Opus 4.8, its latest AI model, which is state-of-the-art on the benchmarks that matter most to publishers, including accuracy in knowledge work and data analysis. It also allows users to control the level of effort (and therefore cost) that goes into answering prompts. That cost point really matters. Simon Willison has a great piece on AI features and pricing that argues that April was the month OpenAI and Anthropic found product-market fit and started charging enterprise customers for what they actually used, as opposed to subsidised prices. Simon quantifies the level of subsidy through his own use: over $2,000 of token costs for $200 in subscriptions. I’ve previously characterised investor-subsidised AI pricing as the Uber model, so it was ironic to read Uber’s CTO complaining about token costs (the company used its entire 2026 budget for Claude Code by April). Meanwhile, DeepSeek bucked the trend by making a permanent 75% price cut on access to its leading reasoning model. Pricing and access to functionality are now critical issues for anyone thinking about building AI workflows for the long term, so I reiterate my advice that the smart play is building model-agnostic systems. I gave a presentation at the IPG Summer Summit last week with a framework for how publishers can do due diligence on vendors of AI tools: my colleagues at the IPG have now turned that into a short, on-demand course available through the IPG Skills Hub. Norway’s National Library has been tasked by the country’s Ministry of Culture with developing a sovereign LLM, leveraging the library’s legal deposit mandate, collections and publisher relationships. The published analysis on this has largely focused on technical aspects, but it will be fascinating to see how this develops and whether other countries follow suit. The Economist is the first major publisher to release a dedicated ChatGPT app which allows users to interact with the magazine’s charts and data through the LLM. This may be the most strategically interesting link this week: not because every publisher needs a ChatGPT app, but because it shows one way to experiment with AI-native discovery without handing over the whole archive. This is a bet on audience acquisition as much as technology. Last week I mentioned that a highly anticipated book on how AI affects truth contained hallucinated quotes. New York magazine ran a follow up arguing that trade nonfiction publishing is unprepared for the wave that is about to hit it. For me, the most interesting conclusion was that the underlying problems are structural and predate AI: “Nonfiction publishing is uniquely vulnerable to AI because the industry has long neglected to do anything to ensure the books it publishes are factually accurate.” Journalist Dorian Lynskey made a related point on Bluesky that inaccuracies and errors were not exactly unknown prior to LLMs. The sheer volume of AI-generated content is overwhelming structures that were rickety to start with.