Context Window 66
It has been a week in which AI moved decisively out of product demos and into politics, newsrooms, courts, and war. Across very different domains, the same underlying tensions keep surfacing: control, accountability, and whether speed is crowding out judgement.* The outbreak of war in the Middle East is uncomfortable territory but impossible to ignore. The relevance to this newsletter is that arguments about the role of AI in those military operations led to a very public break-up between Anthropic and the Pentagon, OpenAI stepping into the gap, a consumer backlash with 1.5 million users leaving ChatGPT (importantly, it isn’t clear what proportion of that number is paying users) and Claude hitting number one on Apple’s App Store chart.
Sitting behind this is a quiet but consequential move by Anthropic: building an import function to bring memory, preferences and context into Claude from other AI models. All of the major models have decent export options, but I think this is the first explicit import option I’ve seen. Data portability underpins consumer choice and experimentation.
One of my favourite AI tools, NotebookLM, released updates including the ability to create cinematic-style video overviews of research materials, and finer-grained controls of visual exports like infographics and slides. It’s increasingly the first place I go for both data analysis and content creation.
Semafor has good analysis of internal conflict over the use of AI tools at the Associated Press. Slack threads shared with Semafor showed a senior AI leader suggesting “resistance [to AI] is futile”, which for my fellow late eighties kids is quite a culturally loaded statement. There was a far more critical view from reporters (“insulting and abhorrent” according to one).
AP said that the leaked threads don’t reflect its official position on the use of AI, but they probably reflect ground truth in that publisher, and many others, more than PR bromides. (What would your internal discussions look like if surfaced through a leak, litigation, or a data subject access request?)
In a different professional domain, Alexander Kustov posted a ten-point provocation on the role and value of AI in social sciences research and publishing. It’s worth reading even if you disagree with his conclusions: many of the points map directly across to publishing, and the piece represents a useful “yes, but what if” scenario for publishing strategists.
On the value of human writing, the Authors Guild has expanded its Human Authored certification programme from members to any author or publisher with a US publication willing to pay a small fee. It will be interesting to see how much this and similar commercial schemes cut through with consumers, or if it’s just a tax on publication that could be avoided with a simple logo on the book cover like the ‘also available as ebook/audio’ notes that appeared on everyone’s books all at once in the 2010s.
Congratulations to the UK Publishers Association on the publication of its Content Superpower report, released ahead of the UK government’s expected response on AI and copyright on 18 March. The PA report shows, among other things, that every major UK academic publisher is expected to be actively licensing content to AI this year. It comes as campaigners raise alarm bells about the possible direction of that government response.
On the other side of the Atlantic, an update in the long-running legal arguments between AI developer Stephen Thaler and the US Copyright Office over the copyright status of AI-generated artwork: the Supreme Court declined to hear an appeal from Thaler to lower-court judgements against him. Mr Thaler is now 0 for 2 at Supreme Court litigation, having earlier lost in the UK.
Andrew Savikas has a great piece which highlights why it’s worth reading past the headlines in AI coverage: a new study showed that AI users scored lower than non-AI users on a quiz measuring their learning on a task—but the results were more nuanced. Patterns like delegating fully to AI or iterative development led to poorer results. But the group that used AI for conceptual enquiry—asking questions, not going straight to answers—scored highest.
Prompted by Paul Ford’s recent argument that there are “billions of software products that don’t exist but should,” I wrote this piece to wrestle with my own ambivalence about how easy AI has made building. It explores the tension between capability and judgement, and picks up Rich Ziade’s warning about using new technology to do old, low-value work faster. The conclusion is a simple but increasingly important one: as the cost of making collapses, the real challenge is not whether we can accelerate, but whether we are accelerating the right things.
It’s London Book Fair next week: I hope to see many of you around Olympia.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.