Context Window 103
I feel sorry for anyone who tried to get out any news about AI this week, as Anthropic’s IPO dominated the media to the extent that there might have been a full two-hour period when I saw nothing new said about Jensen Huang and Nvidia. While the IPO is not directly relevant to publishing, it coincides with important milestones in Anthropic’s copyright infringement settlement, and helps to fund the next stage of it. Besides some immediate checks that authors and publishers should make, I’ve tried to highlight some of the AI developments and views that fell below the fold this week.
Anthropic’s near-$2-trillion IPO and the latest back-and-forth on the company’s copyright infringement settlement with authors and publishers produced an uncomfortable juxtaposition of tech triumphalism and authorial anger. Thanks to the IPO, Anthropic is good for the final tranche of settlement funds, which was contingent on an IPO or other funding event. The entire settlement is less than 2% of what Anthropic raised in its float.
In the meantime, the first payments to rights holders are due to be made next week. For many authors and publishers there are ongoing questions, particularly regarding the dispute resolution process for claims. Today (20 November) is the last day to flag a claim you disagree with. For anyone in the midst of it, the Authors Guild has a good explainer.
The Authors Guild also highlights perceived deficiencies in publishers’ rights reversion processes: where disputes exist, publishers may want to check their position before those disagreements are passed to the settlement Special Master for resolution.
Retraction Watch has a fascinating piece on academics including prompt injection—such as a hidden instruction to favourably review a paper—in submissions for peer review. This isn’t a new phenomenon, and it’s a challenging one for authors and publishers of all sorts. It reflects a foundational assumption by the author that AI is going to be used to assess their work, which in turn says a lot about their trust in review processes that are generally stated to be human-based. And it opens up an arms race: if you assume AI is going to be used, and that your competitors for rivalrous publishing slots are doing this, the rational (if not ethical) response is to do the same.
This example is about academic publishing, but I’d be curious to know whether any literary agents or trade publishers have explicitly checked for prompt injection in submissions made to them—based on a couple of quick chats this week, I suspect the incidence of prompt injection may be higher than the proportion of agents and publishers actively checking for it…
The tech publisher Tim O’Reilly linked to an interesting piece on the role of a software developer when AI writes code. It’s immediately interesting to anyone working in development and sets out a framework for humans to define specifications and steer the process. But I’d suggest it has a wider relevance to publishing.
Back in February, I wrote about the similarities between coding and publishing, and why publishers’ experience of language and systems gives them an unfair advantage working with AI. I’d suggest that a thoughtful editorial or operations person could take the basic approach set out in the O’Reilly piece and apply three ideas to large parts of the publishing workflow: every rule needs a proof, every error becomes a check, and there is a named person responsible for each decision. In publishing terms, that is a style guide, proofing, and editorial sign-off.
Writing code (or developing agentic workflows) is different to prose. There’s a basic, binary test of whether code runs. Writing good narrative text is much more subjective. In this piece for AI platform Every, Mike Taylor explains the six-stage process he followed to create an AI skill that writes like him. I’ve experimented with this a bit in the past and have never been happy with the results, but Taylor’s process is another level of preparation and optimisation.
For many people in publishing, it will be deeply uncomfortable reading because it describes putting AI into the part of the process that many (most?) want explicitly reserved to humans. However, I’d suggest that makes it essential to read to understand just what is possible with some effort: better replication of prose style, but at the cost of the writer’s grasp of their argument.
Google’s Gemini Notebook has a new creative output, allowing users to build interactive webpages based on the content they’ve curated in a notebook. For educational and non-fiction publishers, this is a fast way to prototype explainers, activities and other dynamic content, though not yet a production-grade tool.
Gemini also got an upgrade this week, with the Tasks functionality made available free in six countries, and the ability to make calls and payments on behalf of users in selected territories. The UK is not among them, so I haven’t been able to trial it. I have to admit that, even with the various safeguards that Google describes, this makes me very nervous. I’d still recommend reading the announcement as it provides a clear set of examples of where tech companies see agents fitting into their customers’ lives: even if you don’t use agents yourself, you may very well end up interacting with other people’s agents. Bookshops may well be the first part of the trade to experience this.
For academic and scientific publishers, OpenAI published a new paper on the results of its ChatGPT for Academic Researchers summer pilot, which reached 1,800 researchers. It found that more than two thirds of researchers used ChatGPT to develop research ideas, and that nearly a quarter of their idea-development messages asked it to suggest a research question outright. Researchers also stayed highly engaged, checking and iterating on outputs. It’s important to remember that the research base was those who opted into a pilot rather than a representative population sample, but even if we treat the results as directional only, it’s an interesting bellwether for researchers.
OpenAI also introduced a ChatGPT for Sales product, including integrations with leading third-party CRM and sales tools, and a set of prepared workflows. The whole thing looks very geared to software and services sales—I’m unconvinced it’s ready to automate the process of working with bookshops. For publishers selling B2B subscriptions, for example, it could be worth investigating.
Amazon and Audible seem to be introducing a new AI feature almost every week: the latest is a discovery tool that lets customers experience a traditional audiobook sample combined with an AI-voiced overview of the book generated from public information, metadata and reviews. There’s no word in the announcement about publisher opt-outs or controls, but it says it will roll out on selected titles in the US first, with plans to expand support in time.