I’m writing this from a very sunny Edinburgh, where I presented to the lovely Canongate team at their away day. It was great to get off camera and out of the office and have conversations with a group of brilliantly creative people on their event theme, ingenuity, and how AI can support it. If you’re looking for a speaker for your conference or event, do let me know. ​ In less happy news, it has been a tumultuous week for AI and copyright. Just after I sent the last newsletter, the US Copyright Office issued a “pre-publication” version of its third report on copyright and AI, tackling the issue of Fair Use and training data in a pretty balanced way, pointing to case-by-case determinations and licensing. The Authors Alliance has a good summary. Then, over the weekend, there were reports that the Trump administration had fired registrar of copyrights Shira Perlmutter, though the Copyright Office still lists her as holding the post and her successor is unclear. This follows the earlier firing of the Librarian of Congress. Many commentators have highlighted the administration’s close links with AI companies and the rapid publication of the report. For publishers it’s another sign that we’re way past business-as-usual. ​ Audible announced a range of production options for AI-narrated audio. Their stated aim of closing the gap between the number of published books and the number of available audiobooks is important both commercially and in terms of accessibility. It’s unclear at the moment of publishing this, but very likely that audio produced via this route will be exclusive to Audible. And there is no sign of them changing their longstanding policy of prohibiting audio produced with alternative technology, such as the ElevenLabs/Spotify partnership. From the perspective of author, publisher and consumer choice this is a horrid example of platform lock-in. I wonder whether how many publishers are ready to go all-in on Audible at a time that Spotify is growing fast (or think that Spotify and its wider network is ready to replace Audible). ​ The BBC’s Director General Tim Davie gave a major speech on the corporation’s role and forthcoming Royal Charter renewal. Particularly relevant to this audience was the suggestion that BBC Bitesize could create a personalised AI learning companion for students aged 7-16. The corporation is able to consider such a move because of the scale of its content and marketing platforms, but it raises important questions. From an AI perspective, such a project needs to solve for accuracy, authenticity and how it uses personal data. But even more fundamentally, this development could have a significant effect on the market for other educational products and services. Anyone publishing educational content in this area in or for the UK will want to watch this closely. ​ As marketing and publicity departments experiment with how to promote authors and books using AI, Semafor suggests an old school approach: “the best way to get your… message into the output of ChatGPT, Claude, Gemini, and the rest is by talking to journalists”. It’s a great reminder that traditional earned media can still shape discoverability, even in an AI-driven environment. ​ On the subject of the news media, CJR has a fascinating set of short interviews with leading journalists discussing their approach to AI. There are obvious parallels for other writers and editors, and I’d love to see a similar exercise for book and journal publishing. Would you be open to being interviewed about your AI use and profiled in a future newsletter? Drop me an email if so: hello@outsidecontext.co.uk. ​ Human-in-the-loop reminder of the week: a new Coca Cola ad campaign generated with AI hallucinated details of J.G. Ballard’s writing, and conflicting statements from the ad agency and the Ballard estate suggest that the latter was unaware of the use of AI. Probably just as well Coke is no longer using the “it’s the real thing” ad line. ​ Finally, I was happy to be asked to join Simon Mellins and co-host Simon Holt on the Digital Publishing Podcast last week, and the show is now available. Over the course of an hour we discussed how AI affects strategy, competitive dynamics in large and small businesses, accessibility and employment in publishing. The Simons are great interviewers and I recommend checking out their other episodes.

This week’s newsletter starts with a question that’s been quietly nagging at many of us working with generative AI: are we saving time, or just skipping the thinking? A brilliant post introduces the idea of “cognitive debt”—the mental version of technical debt—where shortcuts today can cost us clarity tomorrow. It’s a useful lens for publishers figuring out how to scale AI responsibly, especially as new tools promise more speed, but not always more understanding. ​ Creative strategist John Willshire published a great piece this week building on the concept of Technical Debt in IT development: with generative AI, the risk is Cognitive Debt, “where you forgo the thinking in order just to get the answers, but have no real idea of why the answers are what they are.” It’s a really important point, which highlights the difference between text generation with LLMs and writing as an embodiment of a thought process. As a lot of my strategy and writing work is solo, I really appreciate AI as a resource and sparring partner, but I’ve definitely experienced revisiting old LLM outputs and being uncertain about the underlying logic. To mitigate this, here are a few approaches I’ve found helpful: retaining older prompts and outputs in an archive so they can be revisited (ChatGPT’s Projects feature is a great way to structure that—does your organisation have a prompt retention policy or guidelines?); including a Chain-of-Thought request in prompts to see the process of arriving at an output, e.g. “Explain your reasoning step by step and highlight any assumptions”; similarly, if you’re outputting code such as Python or VBA, asking the model to include inline comments section-by-section; and, for non-trivial use cases, keeping metacognitive notes in the same way a scientist keeps a lab notebook: what I asked, why, my assumptions, how I evaluated the answer, and whether and why I made any decisions based on the output. (I do this by hand, using a Remarkable tablet that also stores my client notes, and I find that just transferring something into a handwritten medium seems to help my recall of it.) I’d be really interested to know if you have strategies for working with LLM outputs and will share any ideas in future. ​ Amazon has released a new seller tool, Enhance My Listing, to generate A+ content and optimise product pages on the store. It claims a 40% increase in listing quality from the use of AI tools, and that 90% of AI suggestions are accepted by sellers. It’s being rolled out to US sellers initially. ​ Kevin Anderson has been posting from WAN-IFRA’s World News Media Congress 2025: this post includes some of the highlights from WAN-IFRA’s forthcoming AI survey, including a gap between the perceived importance of AI and its usefulness. Still, the results show nearly half of publishers seeing increased productivity, and only 8% not using AI. ​ The same conference saw the launch of a new initiative to protect news integrity: the key principles of consent, value, attribution, plurality, and partnership between publishers and tech companies are extremely relevant to other parts of the publishing industry. ​ Speaking of partnerships, John Wiley continues to power ahead, with a new deal with AI platform Perplexity to integrate AI search and purchased Wiley content for institutional users. There’s a lot to unpack here: it positions publishers not just as providers of training data but as valued content partners. It also reinforces the importance of attribution and ensures alignment with how students are increasingly accessing content via AI-assisted search tools. ​ The Wikimedia Foundation published its AI strategy, which is based around creating time for human editors to concentrate on “deliberation, judgement and consensus-building”. Unsurprisingly, it also includes a strong commitment to open source models. There’s a lot here that would serve as a model for any publisher. ​ I know many of you are interested in the copyright ramifications of AI from the number of clicks on related links: if you’re in that group, the Authors Alliance published an interesting take on Studio Ghibli’s visual style being replicated by AI, explaining why it’s important that copyright should not protect style, for the sake of creators, and so the copyright system promotes creativity for the public good. ​ Thanks to Gavin Marcus from Storywise for sharing the facepalm moment of the week: an Australian radio station used an AI-voice presenter for six months. That they were able to do so in the first place speaks to improvements in the technology. But the instant consumer backlash when the truth came out is a reminder: transparency matters, and trust is hard to earn—and easy to lose.

There’s a lot to unpack this week, including a couple of quite contrarian views on the environmental impact of AI and on the use of copyrighted material as training data. But I’m starting with something that’s both exciting and very practical… ​ OpenAI is rolling out an update to ChatGPT’s web search capabilities to include online shopping results and personalised recommendations. With around a billion users, this has the potential to disrupt existing retail channels: perhaps not like TikTok Shop in its ability to drive bestsellers, but more like Google Shopping’s impact across a long tail of titles. (OpenAI expressed interest this week in buying the Chrome browser with its 60%+ market share if Google is forced to divest it, which would accelerate that impact). For publishers wanting their books to be visible, there are two immediate action points. First, products are discovered by a search crawler, OAI-SearchBot, and you need to ensure that robots.txt on your website allows access to this. To my mind, OpenAI have implemented this quite thoughtfully: OAI-SearchBot only crawls for product information, not model training, so it would be possible to opt your website out of its other training bots, but still make titles visible. Secondly, referral URLs from OpenAI include the UTM parameter utm_source=chatgpt.com so this can be tracked in Google Analytics. Forward this paragraph to whoever looks after your website and analytics. The second action point: at the same link, OpenAI has opened up a waitlist for suppliers to provide product feeds directly to ChatGPT—if your website has D2C sales, you should sign up today to be notified when this is available. ​ Earlier this year, researcher Andy Masley published a deeply researched 9,000 word blog post on the environmental impact of individual ChatGPT usage (or similar models). Last week, he published a concise version that presents updated findings and puts the environmental impact in perspective. This doesn’t address data centres/infrastructure, but it gives useful context on individual choices. While he doesn’t include publishing in his comparisons, the uncomfortable truth is that any reasonable individual level of LLM use probably doesn’t signify against the impact of one traditional publishing cycle. ​ Anthony Finkelstein, president of my alma mater City, has a provocative take on AI training and Fair Use, highlighting that copyright is a balance between creators, consumers and society, and arguing that many human creators overestimate their originality and underestimate the proximity of their work to others. His view is fundamentally critical of Meta’s business ethics but supportive of the underlying AI development they’ve undertaken (with 150 papers in LibGen, he cannot be accused of being a disinterested observer, though as a senior academic he can potentially be more relaxed about the impact than an early career researcher or more general writer). ​ YouGov has new data on what people think AI is good at: for publishers, it’s instructive that 39% are positive about AI writing on academic topics, 35% positive on AI creative writing, and 50% positive on ability to summarise complex topics. There are also insights on AI in other areas of life that map to non-fiction publishing. It’s interesting to compare that research with recent figures from HBR on the top AI use cases in 2025. ​ After the recent story about LLM outputs in a medical textbook, researchers have found over 700 journal articles that contained telltale signs of AI, phrases such as “as of my last knowledge update” or “as an AI language model”. In some cases the response to this has been to remove the offending text but not to revisit the article itself, which seems ethically dubious. As an industry, we need to do better than this. ​ In terms of lessons from outside the industry, I was really interested to read a LinkedIn post from Walmart executive Jose Chapa, whom many of you might know from his time on the Amazon EU Kindle team, about Walmart’s new AI tool, Trend-to-Product, which collects fashion trend data, creates ideas and mood boards for the apparel design team to refine and sign off, and then generates a spec for the manufacturer. The key metric is that it reduces the production timeline by up to 18 weeks, while retaining human control. It’s an impressive looking solution for trend-driven products: it’s interesting to think about what an equivalent for books/content commissioning would look like. ​ Finally, huge congratulations to my colleagues at Burleigh Dodds Science Publishing, winners of the inaugural PLS AI Award at this week’s Independent Publishing Awards. (Disclosure: I am on the board of Burleigh Dodds, work with the IPG on AI training, and played no role in either submission or judging of that category).

This edition covers a correction to last week’s link on LLM task length, the new native AI text-analysis formulas in Google Sheets, the latest Adobe Firefly upgrades, eye-catching data on AI-generated music on Deezer and rising submissions to publishers, Wikipedia’s training-optimised Kaggle dataset, an Anthropic paper on how Claude expresses values, Ethan Mollick on “Jagged AGI”, Tom Goodwin’s AI-flavoured update to the Eisenhower Matrix, Trump’s executive order on AI in schools, the new CLA/PLS/ALCS Generative AI Licence, Ahrefs research on AI overviews suppressing click-through, and a contrasting Washington Post deal and Ziff Davis lawsuit involving OpenAI.

This edition covers the headline findings of Stanford HAI’s 2025 AI Index Report, fresh data on the rapidly growing length of tasks LLMs can perform, OpenAI’s o3 model release, the importance of human leadership in AI initiatives, a Bloomberg investigation into Inkitt’s AI-powered romance factory, and an embarrassing case of an AI-generated passage slipping through Springer Nature’s editorial process.

This edition covers Shopify CEO Tobias Lütke’s AI-first hiring memo, a behavioural study suggesting reader preferences for human authorship are weaker than stated, an argument that AI may improve content quality by writing for the model as well as the reader, Anthropic’s launch of Claude for Education with Socratic prompting, a salutary LSE blog on “efficient inefficiency”, OpenAI’s persistent-memory upgrade to ChatGPT and the Temporary Chat workaround, Meta’s Llama 4 multimodal models, and a court ruling allowing the New York Times’s copyright case against OpenAI to proceed.

This edition covers Amazon’s new Nova Act browser-agent model, Google’s Gemini 2.5 reasoning model, Rachel Coldicutt’s responsible-AI dos and don’ts, fresh CLA research showing 82% of UK professionals upload third-party content into AI prompts, FT analysis on the limited employment impact of AI so far, Tyler Cowen on AI in his writing workflow, a Nature piece on AI and academic peer review, and new data points on AI scraping bots overwhelming publisher infrastructure.

This edition covers OpenAI’s inline image generation in 4o and the Studio Ghibli controversy that followed, a US court’s refusal to halt Anthropic’s training on song lyrics and what that means for publisher litigation strategy, a new world map of AI copyright lawsuits, MIT Tech Review on the dangers of total autonomy for AI agents, OUP’s AI Discovery Assistant with Silverchair, the worsening impact of AI crawlers on open access infrastructure and Cloudflare’s response, and a BookBrunch op-ed on AI in bookselling.