Context Window 68
It’s been a busy week, with significant developments in copyright policy and a steady drumbeat of new research and product launches. The key question posed by new research linked to below has stayed with me: for the right task, AI brings real benefits, but how do we determine what the right task is?* I gave the morning keynote at the ALPSP University Press Redux conference in Liverpool on Wednesday, talking about the impact of AI on scholarly publishing. It was one of the best conferences I’ve attended in a long time, with particularly good follow-up discussions in the breaks afterwards. My slides are available here.
For UK readers, the big news this week was the publication of the government’s update and economic impact assessment on AI and copyright. Last year, only 3% of respondents to the consultation on the issue backed the government’s preferred position of a broad copyright exception for AI training with rights holder opt-outs, following a significant lobbying campaign by authors and the wider creative industries. As a result, the government no longer has a preferred option and will consult further.
On the subject of the impact of AI on authors, I was asked a question at Redux by Anthony Cond from Liverpool University Press: how did I feel about AI as an author? I’m quite relaxed about the idea of AI training on my writing: what matters to me is whether it creates new layers of engagement rather than simply extracting value.
What I would really like is for it to provide an interactive layer around my work, similar to what Stephen Witt outlined last year. I acknowledge many other authors won’t share those sentiments, but I was interested to read this interview with author and Wired co-founder Kevin Kelly, where he sets out a more AI-positive view.
A date for the diary: on 31 March, I’m delighted to be taking part in a webinar hosted by Crius Group on AI in publishing, along with colleagues from around Europe, including Simon Mellins. All the details are on LinkedIn.
European AI unicorn Mistral launched a new product called Forge, which allows enterprise customers to create custom AI models trained on and grounded in institutional data. For the largest publishers with significant archives and proprietary data, that opens up strategic possibilities beyond generic API or MCP access.
Katie Parrott at Every published an excellent, step-by-step guide to creating a manual of style for AI. The promise is that this will teach an LLM to write like you, but even if you don’t want to go that far, I found this a really useful exercise for identifying unconscious patterns and characteristics in my writing.
Anthropic released a huge study on attitudes to AI, which is incredibly valuable context. But as interesting as the results are, the key thing I noticed is that this is the second major piece of research where AI was used to interview users at scale: 81k user interviews in 159 countries and 70 languages, in a single week. The ability to run structured, multilingual qualitative research at that scale changes the economics of user insight—something which publishers could leverage, even if on a smaller scale.
On the wider impact of AI, I really liked this observation in a Bloomberg piece about how AI is used as narrative cover to justify job cuts that are really down to other factors: “We’re restructuring around AI” is a growth signal. “We over-hired during the pandemic and revenue softened” is an accountability signal.
A new research paper on productivity benefits showed that consultants using AI tools completed 12.1% more tasks than colleagues without AI, and to a higher standard. However, the kicker: where they tried to complete tasks that were outside the capabilities of the AI model, results were 19% less likely to be accurate. The key question, therefore, is a great, transferable one for every AI user: understanding what models are capable of and selecting tasks appropriately. And for that, experimentation > documentation.
On that question of capabilities, Google’s most recent spreadsheet agent, Gemini in Google Sheets, achieved a remarkable benchmark of being within a single percentage point of expert human performance and accuracy. It’s a reminder that specialised agents are narrowing the gap in structured tasks, even if general reasoning remains uneven.
Finally, for a wider take on the nature of AI productivity, Matt Jones’s essay on the disconnect between the time taken by AI agents and humans, and what it means for us, is a really excellent long read to settle down to with a coffee. There’s also a fascinating conclusion section on how AI helped to write the piece.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.