Context Window 10
This edition features stories on the UK’s national AI strategy, a study of LLMs in education, Bloomberg editor-in-chief John Micklethwait on AI in journalism, and new features from ChatGPT and Copilot.
The UK government rolled out its national AI strategy, accepting all fifty recommendations made in a report by venture capitalist Matt Clifford. There’s a lot that is positive in this, but for many publishers the key question is recommendation 24: to reform the UK text and data mining regime (that is, use of copyrighted material for training) so that it is at least as competitive as the EU. Pegging the UK’s approach to the EU is potentially problematic: the EU’s AI act allows use of copyrighted material for training unless the rights holder has opted out, and the propriety and efficacy of opt out mechanisms, as opposed to requiring specific permission and licensing, is contested by virtually all rights holder groups. It also throws into question the government’s own consultation on the subject, due to run through February: it seems to rule out options that the government presented in that consultation, such as strengthening the copyright regime. While regulatory debates dominate in the UK, AI’s impact on education provides a striking example of its potential in practice. The World Bank published a piece on the impact of using LLMs to support school students in Nigeria: their headline claim is that learners using AI support made nearly two years of academic progress in six weeks, outperforming 80% of traditional educational interventions studied through randomised controlled trials. From what I can find, the AI model in question was generic Microsoft Copilot: if I was an educational publisher, I would be very interested in researching the impact of AI tutors linked to specific, trusted content, which would also help to address potential ethical issues of accuracy and representation in general training data. Bloomberg editor-in-chief John Micklethwait published a piece on how journalism will adapt to AI, based on his James Cameron Memorial Lecture at my alma mater, City. There’s an interesting benchmark that around a third of Bloomberg’s stories have some element of AI automation. He then makes eight predictions, many of which seem equally relevant to book and journal publishing: his view that AI changes jobs more than it replaces them, and that the primary impact is on editors rather than writers (or, to paraphrase, on back office processes rather than creativity) is consistent with what I’ve seen in most book publishers. My only significant disagreement is his suggestion that hallucination will be easier to solve in text than in audio/video: I would argue it will be more apparent in video, but even with improvements in model accuracy, LLM outputs will still be capable of subtle, pernicious errors that require human scrutiny. On a much more practical level, OpenAI has started rolling out the ability to schedule tasks in ChatGPT—if you are on a Plus, Pro or Team subscription, you’ll see it first as a beta feature, though it will roll out to all users. There are some limitations in the beta: only 10 concurrent tasks per user, and it can’t work with uploaded files or custom GPTs yet. However, this promises to be a really useful feature for publishing workflows. There was a parallel announcement from Microsoft this week, also pushing further into agentic AI, or models which can perform more complex tasks rather than just responding to a prompt. Microsoft announced a new, free pricing tier, Copilot Chat, sitting underneath the regular $30 user/month Copilot. This allows users access to the ChatGPT 4o model without paying an OpenAI subscription, and to create agents on a pay-as-you-go basis.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.