Context Window 84
It’s been a long, hot week on the road, seeing publishers in Dublin, Cambridge and London. Lots of the conversations were about choosing platforms: a timely question as models get better, but also become more complex to configure, integrate and use well. This week’s stories pick up that theme: agents are improving, switching costs are rising, and companies need to measure the impact…
After several months in which it felt like competitors were making the running, OpenAI shipped its latest ChatGPT 5.6 models, along with ChatGPT Work, an agentic AI that works with your existing files and apps. In the same way that the advent of coding tools led to a boom in application development, now that every major AI company offers a desktop app, I would expect to see a surge in new workflows.
One immediate practical impact: when AI experimentation was about simple prompts, switching between models was relatively quick and easy. To make the most of desktop agents, users will need to integrate data and set up workflows, which makes chopping and changing between agents a less trivial matter.
As tools like ChatGPT Work become part of professional life, there’s a new challenge for companies: how to measure employee performance. This piece sets out some useful ideas for managers and HR teams. At the moment, evaluation of AI-assisted work still seems to be largely about whether a task is completed faster or cheaper.
Plaintiffs in a long-running copyright infringement lawsuit against OpenAI, including the New York Times, went to court to argue that the AI company has withheld evidence on its training data and outputs. Of course, where the media companies characterise this as discovery misconduct, OpenAI frames it as protecting user privacy.
Another interesting update in a long-running case: image generator Midjourney, which is being sued by Disney, Universal and Warner Bros, is trying to force the studios to disclose how they use AI, suggesting that they are guilty of the same behaviour they accuse Midjourney of. It’s an interesting tactic, particularly because many media companies (including the New York Times and PRH) use AI tools from companies they are independently suing.
There’s an interesting piece here on TIME’s decision to move away from WordPress for its web CMS, driven in large part by a lack of support for AI integrations. I’ve seen or led several major replatforms in my career and it’s never a decision taken lightly. The fact that AI functionality is now an important enough factor to outweigh the cost of change says a lot about publisher priorities. There’s also a lesson for software vendors that prior scale and market share is not a substitute for updating core products.
Digiday looks at the phenomenon of AI poisoning—where brands go beyond optimising their own products and start trying to harm competitors’ reputations through bad reviews and online posts that are ingested into LLMs and thus repeated. That’s a depressing enough thought, though there was also a terrifying statistic cited: apparently just 8% of users fact-check AI outputs.
Library app Libby will add filters allowing users to select whether they see AI-generated content. Interestingly it will be based on publishers disclosing AI use in metadata, rather than detection software. That does admit the possibility of a false negative if a publisher gives inaccurate information, but I would be less worried about that than false positives if content was uploaded into potentially inaccurate detectors.
A coda to the story last week about Microsoft planning to embed 6,000 employees inside customer organisations to help them develop AI solutions. Semafor reports on business unease about the role of forward-deployed engineers and whether their real purpose is helping customers to build, or developing AI companies’ knowledge of customers that could help them build competing products.
It’s not a fanciful notion: AI labs continue to hire experts to advise on how to make their models more applicable in business settings—here’s a live example. Practically, this is likely to hit bigger industries than publishing first, but the pattern matters. The more AI vendors embed themselves in customer workflows, the more value shifts from selling software to learning the customer’s operating model.
Finally this week, this new paper investigates the AI policy frameworks in use across the largest academic publishers, including Elsevier, Springer, Wiley, Taylor & Francis, and SAGE: some basic consensus, but also “performative assertions”, which sounds very like publishing.