Context Window 101
I’ve been travelling most of this week, and watching the extraordinary political developments in the US from trains and hotel rooms. “Close to an extinction-level event” is how one Republican politician characterised it, but it introduces some serious issues for AI companies as well—and as a third-order effect, that poses questions for those of us who use AI. This week the newsletter links to a framework for understanding that which you could use with colleagues or your favourite AI research tool.
This newsletter is about AI in media and publishing but it’s impossible to ignore the political consequences of the midterm elections in the United States. Public opposition to data centres was one of the key issues, but antipathy to AI is broader. I must admit that I had thought it might have been a lower-order consideration, but over half of voters said that AI was a crucial or major factor in their choice—so not just preference but real salience. Verity Harding has an excellent summary of what went wrong for the tech industry, what might happen next, and what should happen next.
For larger organisations, HBR unpicks the policy implications across six key areas—but if you’re in a smaller business and don’t have the CTO, General Counsel and strategy and policy leads that the piece suggests convening, the list would make a useful starter for a Deep Research prompt.
With significant political attention on big tech and its links with the administration—one Democratic lawmaker stated memorably that “we’re going to need to plant new forests for all of the subpoenas we’ll be printing”—one unsurprising development was that Meta read the room and chose election day to announce that its forthcoming terms of service change, which publishers argued would have allowed it to scrape customers’ websites for AI training, is now paused indefinitely.
Longer term, I wonder if increased congressional scrutiny means that the willingness of AI companies to settle litigation and licence content may increase.
Meta is not the only tech company to trim its sails to the changing political winds. On a more practical note, Anthropic, OpenAI, Google and Microsoft are supporting an update to the robots.txt protocol for website owners to express preferences on web crawling for AI training.
Anthropic is supporting the update already, with the other tech companies promising to do so in future. Meta made a carefully worded statement saying it supported “publishers’ ability to express preferences about how their content is used for foundational generative AI training“—though, like the update, the latter bit is very specific and doesn’t include crawling for other purposes, such as agentic AI or answering user queries. In other words, while worthwhile, the update largely covers the activity that’s already been litigated, not where most of the value now sits. It’s still worth reviewing your data and updating your controls, particularly if you publish the sort of timely content that AI companies want to train on, or if you’re like the client of mine who saw fully one third of their web traffic last month come from Anthropic alone. But be aware of the limitations.
As disruptive as all of this is, it’s worth remembering that generative AI isn’t going away, and continues to be broadly used. The Economist ran a piece on the “surprising popularity” of AI around the world: two thirds of respondents in Latin America and Asia-Pacific and nearly 80% in Africa and the Middle East felt that AI would have a positive impact on their life. It’s Europe and North America where sentiment has soured most. This is not just a point about vibes, but about product strategy: a British or American publisher selling a platform or content with AI features internationally may well have a stronger market than their home customers alone would suggest.
In product news, OpenAI announced the latest version of its image generation model, which is now able to create and export animated GIFs, work with transparent backgrounds and create 4K images, potentially useful for anyone creating social media content.
Just in time for that product release, Pangram, the AI detection platform at the heart of many of this year’s controversies over provenance, took its images and video detector out of preview.
Finally, two useful pieces of vendor content for publishers.
Marketing platform HubSpot published a detailed case study on using AI in its own writing process. Of course, writing corporate communications has a different cadence and tone to books or journals. But there are transferable lessons around research, writing and quality checking, and for a piece of vendor content, I thought this was pretty reflective and balanced. I’d refer back to it if I were rethinking my editorial workflow.
Secondly, Zapier published a survey on AI and automation in small and medium-sized businesses: exactly what Zapier sells, so they are talking their own book. The takeaway: more than three quarters are automating some processes, and almost a third are using AI across multiple systems with limited human review. Most publishers are in the SMB tier, and there’s a lot of useful practical advice. I particularly liked the suggestion that the price of new AI tools should be benchmarked against the cost of doing nothing, which is often in my experience less visible, or not discussed at all.