Hello from Dublin, where I’ve been speaking at the Publishing Ireland conference. It’s been a really excellent day and good to see many subscribers and friends here. The big theme this week is AI developments from Big Media: new features and guidelines from Spotify, Netflix and Time all have implications and lessons for publishers of all sizes. ​ Spotify has introduced a new feature called Recaps, which uses AI to generate a short narrative reminder for listeners when they come back to an audiobook after a while. The carefully worded announcement states clearly that book content is not used for LLM training—though it seems likely that it is processed by an LLM for generation, even if the data isn’t retained or used to train the model. Functionally, this is similar to Amazon’s recent Story So Far feature. There are two immediate implications for publishers. With two of the leading retail platforms offering this feature, similar functionality is likely to come to other reading systems. Second, the tight integration of AI into core platform functionality shows how difficult it is becoming to accommodate requests from authors who want to avoid any AI use associated with their work. ​ Generation of summaries is also the focus of this opinion piece from the Authors Alliance, which highlights an implication of the recent court ruling in the OpenAI class action: that certain short summaries of fictional works might, in some circumstances, be interpreted as infringing derivative works. This isn’t an issue for Spotify or Amazon, who work directly with rights holders, but the piece speculates about potential risks for platforms like Wikipedia. ​ AI audio platform ElevenLabs has launched a new marketplace for ‘legendary voices’, synthetic versions of leading figures such as Maya Angelou, Michael Caine, Laurence Olivier and Judy Garland for text-to-speech applications. Most of the featured voices are no longer dead, with relatively few contemporary figures. It will be interesting to see whether more living rights holders choose to participate, given the varying legal frameworks around voice likeness and posthumous rights. ​ Netflix has published a set of guidelines for the use of generative AI in creative applications, which is a useful thought starter for publishers. What’s particularly relevant is the thinking on ethics and how AI works with creative workflows and third party intellectual property. Ideally publishers should have policy documents that are internally facing like this one, and for key external partners like authors (as in the Wiley examples last week). ​ Google has introduced a new retrieval augmented generation (RAG) functionality to the Google Gemini API. Put simply, the new File Search tool allows Gemini to interact with your own repository of files rather than relying solely on its general training data and web search. This is priced aggressively at $0.15 per million tokens (roughly 5-6 non-fiction books). ​ Google has also introduced a series of improvements to one of my favourite AI tools, NotebookLM, including a Deep Research model and better integration with PDFs and other files in Google Drive. This makes it significantly more useful as a research tool. ​ As an example of a really expansive publisher AI strategy, Time magazine has introduced an AI Agent offering interactivity, summaries, translation and audio versions alongside its reporting. ​ Helen King’s PubTech Radar has an interview with Jonathan Woahn, the founder of AI content management platform Cashmere, which sits between publishers and AI companies. It’s all useful context, but the key takeaway for me is that publishers can either engage with licensing and shape how it works, or watch it happen to them.

I sent the first of these newsletters a year ago this week, to twenty-seven friends and colleagues. Since then, it’s grown to over 1,500 readers and has become one of the highlights of my work. The cadence of researching and publishing weekly keeps me up to date before I hit send, and I really value all of the comments and connections that have come from it. Thank you for your time and attention, and I hope you’ll stay with me for another year. ​ Thad McIlroy has an excellent, thought-provoking piece on the publishing industry and technology adoption, using the slow take-up of ONIX 3.0 as a frame to ask, what if publishing is equally dilatory in adopting AI. Thad evidences his argument in part with the absence of AI requirements in current job listings—to back up that argument, I looked at a number of open roles at UK publishers that talk a good game on AI, and even in technology-forward areas such as digital marketing and analytics, there’s not a mention of AI skills… ​ Authors and publishers suing OpenAI have gained access to internal company discussions on the use and deletion of books from LibGen, which could be seen as destruction of evidence. The plaintiffs are a long way from their day in court, let alone damages or an Anthropic-type settlement, but it’s not a stretch to imagine that as an outcome. ​ Getty Images largely lost its UK lawsuit against Stability AI. The court found only limited, historic trademark infringement relating to Getty watermarks and rejected secondary copyright claims. For rights holders, the UK still lacks a definitive legal position on whether training AI models on copyrighted works is infringement—meaning policy developments and future cases will likely shape the landscape more than this ruling. ​ One of the fundamental challenges for publishers in the last year has been balancing practical opportunities from AI with litigation over use of content in training. Litigation requires confrontational positioning, but companies have to be careful not to box themselves into a corner. That said, I’m not sure I’ve seen anything quite as quick-footed as Universal Music announcing the settlement of litigation and a strategic partnership with the opposing party in the same press release… ​ Amazon filed a new lawsuit against AI platform Perplexity for allowing its AI agents to shop on behalf of users (Amazon is, of course, developing similar functionality itself). Perplexity responded with an understated argument that this is a threat to all internet users. For publishers who are grappling with issues of AI agents accessing information, it’s a small foretaste of the bigger arguments when agents are transacting with websites. ​ Separately, on a more practical note, Perplexity published a new set of guidelines on AI in the workplace, built around its own tools but with some useful general advice on prompts and capabilities for any AI user. ​ Slate uses recent discussions between local history publisher Arcadia and its authors as the starting point for a piece on publisher licensing to AI companies. Besides providing general context, it’s also helpful in highlighting the opacity of a lot of licensing deals and other friction points for authors. ​ A week after a series of largely incremental product updates from Adobe, Canva announced a complete overhaul of its product line, now described as a Creative Operating System. AI tools are present throughout, including a proprietary design model, integration with LLMs and AI-powered analytics for marketing campaigns. And Affinity, the InDesign competitor acquired by Canva last year, is being made free in perpetuity (and has been downloaded over a million times in less than a week). Overall, this looks like a really compelling alternative to Creative Suite for publishing design and marketing teams. ​ For academic and educational readers, Sage published a new white paper by Tom Chatfield on AI and the future of pedagogy. It strikes a good balance of highlighting issues and opportunities, particularly the potential for AI to be a guided learning partner, not merely a source of content. Anyone developing courseware and textbooks could get a lot from it. ​ On that theme of AI as partner not replacement, belated thanks to Shimmr founder Nadim Sadek for sending me a copy of his excellent new book, Quiver, don’t Quake, a combination of philosophical musing and practical guidelines on AI and creativity. It made for a thought-provoking return journey from New York last week. ​ John Wiley published a new set of guidelines on the use of AI for journal authors, editors and peer reviewers. This complements their earlier guidelines for book authors, which I’ve often held up as the gold standard for publisher policy and communications. Both sets of guidelines are essential reading and Wiley has done the entire ecosystem a service by making them publicly available. It would be good to see other publishers being similarly proactive (if your publisher has a set of guidelines, let me know and I’ll share it). ​ The Wiley guidelines underpin clarity and transparency on where AI is being used, which seem essential principles. But there’s a tension in this: new research suggests that readers’ perceptions of content shift when they are made aware of AI involvement, with lower levels of trust and engagement—though this is offset when readers have greater levels of AI literacy. To be clear, I do not think anyone should take this as justification to obscure AI use, but it’s a reminder that AI remains a nuanced and contested topic. ​ Overdrive recently added an AI book curation and recommendation feature to its library app Libby: in this interview, CEO Steve Potash discusses the feature, dealing with user scepticism, and whether it can help to arrest the decline in reading.

Greetings from New York City, where I’ve been on a working vacation for the last week. It’s my first time back since Covid, and besides the inevitable changes to the city (I could date the last fifteen years of photos by the maturity of the foliage on the High Line) the thing that’s struck me is the ubiquity of technology. Every billboard in Midtown is for an enterprise AI tool, and while this was primarily family time, I had some good conversations with publishers about their use of AI.

I’m in New York with the family. We were up early for Brooklyn Bridge Parkrun—a very enjoyable, fast, flat course along the waterfront—followed by a cooldown walk around Cobble Hill and Brooklyn Heights.

For the first time in a while, I didn’t go to Frankfurt Book Fair and experienced it vicariously through Slack, WhatsApp and press updates. My impression was a lot of conversation about AI, and a lot of publishers talking more than doing. Many years ago I was introduced to Seth Godin. After our mutual friend had explained my role at a major publisher, Godin smiled and replied with one of his catchphrases: “Great, but what have you shipped?” I thought of that moment while reading some of the CEO platitudes from the Fair. So what did I miss? If you shipped something new this week, hit reply and let me know.

It’s been a heavy week with travel, an IPG Lunch and Learn update and interesting discussions at the BISG AI Working Group. But before publishing breaks for the weekend and packs for Frankfurt, some interesting AI developments across audio, author apps and retail… ​ AI audio platform ElevenLabs has introduced a new managed production service combining AI voice reading and human editing and post-production. Audiobook production is one of the launch use cases, and pricing starts at $2 per minute. It’s an interesting middle ground: cheaper than human voice audio, arguably better than out-of-the-box AI voice but still not quite convincing, to my ear at least. I also wonder if offering a managed service around it is also a tacit acknowledgement that tweaking outputs for quality isn’t easy. ​ A couple of weeks ago, I commented on a piece about AI use in education and speculated about which self-help author would be the first to create their own AI coaching app. Thanks to subscriber Nikki Howard from Gill Books for pointing out that Tony Robbins was way ahead of my speculation: for $99 a month you could have access to 24/7 advice from his chatbot. Similarly, this week, hedge fund manager and bestselling author Ray Dalio launched a beta version of an AI advice app using a voice interface. For publishers working with brand authors, there’s an interesting question of whether this kind of AI product becomes a core part of the format mix, or if it’s the equivalent of the iPad apps that were launched with great fanfare in the early 2010s, few of which survive. ​ With the holiday season in sight, Adobe Analytics published new research on how shopper behaviour is being influenced by AI: 53% of respondents had used generative AI for research, 40% for product recommendations and 36% for finding deals. This growing usage explains why ChatGPT is integrating buy links into LLM outputs, and Google recently introduced a payment protocol for AI agents. ​ For publishers, this introduces a trade-off between restricting AI crawlers from their websites to prevent training, and allowing them to encourage visibility of products in outputs. OpenAI and Anthropic use different crawlers for training and servicing user queries, so it’s possible to take a granular approach. But Google and others do not, and most publisher websites that I’ve reviewed take an all-or-nothing approach. ​ Web content being scraped doesn’t just affect publishers and retailers, but also universities. For academic readers, the Authors Alliance has a good piece here looking at AI scraping of institutional repositories (which often contain versions of publisher content) and tensions between restricting AI training and Open Access publishing. ​ Ultimately, across both domains, this will be a judgement call for publishers, but I’m reminded of Tim O’Reilly’s aphorism that obscurity is a greater threat than piracy. ​ At OpenAI’s Dev Day, Sam Altman claimed that ChatGPT now has 800 million weekly active users—a staggering figure that, if accurate, would make it one of the largest digital platforms in the world. For context, Microsoft 365 has around 400 million paid seats. In less than two years, generative AI has gone from experiment to near-universal tool, a shift no other technology has achieved at such speed or scale. ​ For a more independent take on usage statistics, there’s a great new report from the Reuters Institute at Oxford University: 90% brand awareness of leading AI tools, all-time AI use increasing from 40% to 61%, and weekly use of AI tools nearly doubling. For publishers, it’s striking that information seeking is the number one use case, ahead of content creation. ​ There are some interesting points in a recent Econsultancy study on marketing and AI: in particular, 47% of marketers described their knowledge of generative AI as advanced or expert: only 4% then passed a test on their knowledge. ​ Finally, Dan Shipper has an interesting essay here commenting on sharp progress in AI performance but putting it in the context of human intelligence. It’s a timely reminder that the good results that are possible with AI depend on enormous amounts of human framing—not least designing tasks, prompts, and evaluation criteria.

Another week, another flood of AI product announcements from Microsoft, Amazon, OpenAI and Anthropic. But some of this week’s news feels particularly consequential: if you publish books to Kindle or have a direct-to-consumer sales strategy, read on…

The approval of the settlement in the Bartz v. Anthropic litigation overshadows most other developments for publishers this week. But elsewhere the week has seen new research, strong opinions on LLMs and web traffic, and a new creative AI tool that could get traction with creative teams in publishing. Have a good weekend. ​ Big news overnight: Judge Alsup has given preliminary approval to the copyright infringement settlement in Bartz v. Anthropic. There will be more information on the operation of the settlement in coming days, and I’ll be covering this in detail in my next policy update for IPG members. As a reminder, it’s not enough for works to be within LibGen: eligibility for the settlement depends on copyright having been registered within five years of publication and before August 2022 (the date of Anthropic’s infringement). It’s also worth reiterating that the settlement is about copyright infringement, not the subsequent training, which Judge Alsup found to be Fair Use. In considering the sufficiency of the roughly $3,000 per author, the Judge asserted that Anthropic could have paid as little as $1 for a (presumably used) copy of a book and trained from that. ​ I was talking to a client this week about how publishing breaks down into individual and group tasks. Most of the successful AI use cases I’ve seen have been about making individuals more productive—fewer have addressed collaborative working. So I was interested to see the latest experiment from Google Labs: Mixboard, an AI-powered mood board tool. It’s currently only available in the US, so American friends, have at it, while the rest of us wait or VPN. But this looks like a really interesting way for design, marketing and other creative teams to collaborate on visual tasks, and it’s the first AI tool I can imagine seeing on meeting room screens. ​ Useful snapshot data in Google Cloud’s 2025 DORA report on the state of software development: 90% of software developers are using AI, 80% reported increased efficiency, and 59% claimed increases in code quality from using AI tools. ​ Research from the Tony Blair Institute tends to divide opinion. But its latest report on AI usage and attitudes, based on polling by Ipsos, also gives a useful contextual snapshot. Some highlights: more than half of UK adults have used Generative AI tools in the last year, and just under a quarter use Gen AI weekly; two thirds of people who feel confident in their AI skills expect it to help them at work while leaving core responsibilities intact, but only 45% of people with lower confidence are optimistic; and looking at secondary education, 37% of respondents were comfortable with the idea of AI tutors compared to 32% against, with 33% in favour of AI taking on some routine teaching tasks versus 38% against. Lots of undecideds on both, but even small moves in this direction would have far-reaching implications for the sector and for educational publishing. ​ On the education theme, HBR has an interesting report on an experiment using Generative AI for professional training which found significantly better outcomes from personalised AI coaching. It’s interesting but all a bit vague: I’d love to know more about the model they used, as opposed to generic references to a Gen AI Tutor. But it’s an interesting direction. I’m curious to see who is the first self-help author to release an AI coaching tool alongside their books and courses. ​ Continuing the theme of discovery and web traffic from LLMs from last week, my friend Ani Attamian has a great piece in The Bookseller on AI optimisation, including recommendations for tools to monitor performance and visibility. I’ve known Ani since she was working with publishers at Google, and if you’re looking for help in this area, she’d be a great person to talk to. ​ Cloudflare’s annual letter outlines a future where AI agents, not humans, are the main readers of the open web, and traffic-based business models continue to erode. Their proposed fix is to facilitate AI companies paying creators. But as Paul Ford points out in his newsletter, that vision turns writers into prompt-chasing content suppliers. The real lesson for publishers? Paywalls work. In a world of aggressive scraping and vanishing referral traffic, subscription models—not scale—are becoming the most reliable path to sustainability, leverage, and autonomy. ​ Nieman Lab has a piece comparing the AI strategies of the New York Times and Washington Post, which provides some helpful advice on AI deployment for publishers: I was particularly taken by the Post’s policy that while generic information can be processed through enterprise AI tools like ChatGPT, sensitive data can only be used with internally hosted language models. Few publishers will need that level of security, but I suspect if you’re writing about All the President’s Men, sensitive data really is quite sensitive (RIP Robert Redford). ​ Finally, since we’ve established that a non-trivial proportion of you are classics nerds, you might enjoy this tongue-in-cheek suggestion from Ethan Mollick for properly acknowledging the role of AI in work: tagging it with the Latin phrase, Fieri iussit.