Context Window 95
I’ve been thinking about AI risks in preparation for a BISG webinar last night, and this week’s newsletter includes my thoughts on risk mitigation from that event, Jane Friedman’s excellent, calming advice for authors, and a blanket no-AI policy in an adjacent industry that may be hard to enforce for the reasons Jane and I set out. There’s jeopardy too for the big AI companies, and it comes in the form of their own employees’ thoughts on copyright and training…
It was a great pleasure to open BISG’s autumn events series last night, giving a webinar on AI, trust and risk. I presented four categories of risk for the publishing supply chain—technology, governance, content and relationships—and proposed six practical mitigations for each of them. The recording of the webinar will be available to BISG members, and you can download my slides from the presentation here. Thanks to BISG and everyone who attended.
Many of the issues discussed last night involved authors: for example, questions about the provenance of manuscripts, or publishers using AI without authors’ knowledge. Uncertainty on these issues has led many authors to worry. If you—or the authors you publish or work with—are in that category, Jane Friedman published a clear, comprehensive post on where authors should and shouldn’t worry, and what they can do. Truly excellent advice.
Meta’s Connect 2026 event this week saw a raft of AI and hardware announcements. Its personal AI agent Muse will work across phone, desktop, smart glasses (now with an audio-only option, bowing to criticism of video recording in public spaces) and a new personal hardware device, the Muse Charm (quickly and sharply dubbed a “scamagotchi” by John Willshire).
It’s been a busy but unharmonious week for the tech giants. Amazon aren’t Muse fans: the retailer blocked Meta’s agent from accessing its site to browse or shop on behalf of users. As a reminder, Amazon has fought a running legal battle this year with Perplexity over agentic access to its store, and blocked agents from Google and OpenAI. Of course, it has skin in the game with its own shopping tool, Buy for Me. For now, you’ll just need to buy your own books.
Amazon hosted Accelerate, its annual seller conference in Seattle, and announced upgrades to its Seller Assistant AI tool, and a plugin integrating it with Amazon Quick and Claude, with other integrations to follow. A line from the announcement, “meeting sellers where they already work, and bringing our guardrails and security standards” is as good a one-sentence explanation of the value proposition of MCPs and connectors as I’ve seen. It’s available to US store sellers only for now, but for any publisher selling direct on Amazon.com, this is worth investigating. It will be interesting to see if the underlying approach rolls out to Vendor Central and other areas of the Amazon ecosystem.
OpenAI’s ads business continues to be one of the more interesting AI topics for brands and marketers: the latest development is Sponsored Agents, which will allow users who click on an advert to have a conversation with an agent developed by the advertiser. However, a cautionary note: a security researcher set out evidence that OpenAI is setting cookies that let it track user behaviour on third-party advertiser websites.
Setting aside advertising and shopping, the other big development this week was in copyright litigation. Unsealed briefs quoting internal documents in the actions led by The New York Times and the Authors Guild respectively were staggeringly revealing about how insiders at OpenAI and Microsoft viewed their own companies’ training efforts: take your pick from, “the largest theft of labor in human history”, “a complete mockery of the idea of ‘fair use’”, that “LLMs are a product that destroys its supply chain”, and that LLM outputs are “largely substitutive, period”.
Meanwhile, Axios reported a coda to the Department of Justice’s recent intervention in the same case: its statement of interest in support of the tech companies was signed by senior officials rather than career attorneys, and took interested agencies such as the Copyright Office and Patent and Trademark Office by surprise.
A review by the Forecasting Research Institute found that expert predictions on AI progress have been less than accurate. Last year, its experts suggested a 10% chance of one of the Millennium Prize math problems being solved by AI by the end of 2027. OpenAI claimed a solution to the Navier-Stokes problem weeks ago. Similarly, earlier this year, FRI experts suggested maximum year-end revenue for an AI company at $16-25 billion, compared to around $100 billion already. Figuring out practical ways to use AI in the short to medium term may be a better use of time than prognosticating on where it goes.
Finally, an interesting update from an adjacent industry to publishing. UK Games Expo, the country’s largest tabletop games convention, announced a ban on all merchandise, books and artwork “generated completely or in significant part” using AI tools. The show has a significant crossover with publishing, with publishers like Hachette Boardgames and Bloomsbury’s Osprey Games exhibiting. More broadly, it speaks to a significant consumer backlash which trade publishing has arguably not experienced to the same degree. Outside a few behemoths, the tabletop games sector is overwhelmingly made up of small companies, hobbyists and individual creators, as much a craft as an industry, so the sentiment is perhaps understandable. Quite how they are going to enforce this fairly, across thousands of exhibitors, or manage erroneous or malicious reporting of AI use from tens of thousands of attendees, is beyond me.