Context Window 94
The biggest story this week is about the gap between what organisations are doing with AI and what their staff are paying for. It demands urgent attention from leaders, but it will be solved. In the longer term, I think the story of a student in Nanjing with a $2 budget is the one that should most concern international publishers…
Deloitte released new research on AI in the workplace showing nearly two-thirds of 25,000 UK workers surveyed used generative AI at work, nearly a third without the knowledge of their employer. Strikingly, one in six AI users said that they used an AI tool in their work that they paid for personally. Extrapolated to the working population, that represents £958 million of annual spending on AI subscriptions by individuals—or billions of dollars if the same dynamics play out internationally. How big might that number be in your business?
Taking the research at face value, there’s a lot going on here. One in five of the workers using unapproved services said they paid for an AI tool because it was superior to what their employer provided. This is the Copilot problem: the model the business got by default isn’t the one people want to use. One has to assume that rational employees wouldn’t pay personally for something that didn’t deliver a real productivity benefit—and if employers haven’t seen a drop-off in quality, they either have a confirmed AI use case or a performance management problem. And of course, individuals may be getting a benefit, but this kind of shadow AI use sits outside key company policies including data protection. A separate report from Perplexity cites data suggesting senior leaders are twice as likely as junior staff to do this.
What to do about this? I know of examples where companies have used approaches ranging from automated detection to anonymous surveys and AI amnesties to get on top of this. Whatever the approach, treat how your people are actually using AI as a data point on your policies, tools and use cases, not just a risk to be managed.
Speaking of policies, Microsoft published a draft Code of Conduct for its in-house MAI models and is seeking feedback over the next six weeks. There are shades of Asimov’s laws of robotics in the phrasing. Of course, other tech businesses, including Google, found a tension between high-minded statements and real-world implementation. But this is a good start. My feedback: pay copyright owners for their work.
An example of the challenges of real-world use: OpenAI recently trumpeted the use of its models to solve one of the Millennium Prize math problems, but is now facing questions as to whether use of OpenAI’s tools by human researchers helped to inform its approach (the company denies using the researchers’ prompts or proof). Helen King published a smart take on this, pointing out the limited protections available in terms of service and the broader relevance to any researcher sharing work-in-progress with an AI model.
Underlining that privacy question, a new report highlighted the extent to which users’ ChatGPT prompts, including personal information, are reviewed by OpenAI and contractors working on improving the service. This isn’t a fundamentally new problem: the same issues came up when documents moved from local storage to the cloud.
Ofcom released new research on how consumers experience news media, which showed that one in five adult consumers are using AI to find news stories, with obvious implications for visibility and accuracy of what they find. The research found AI apps scored 44% for trustworthiness, versus 65% for broadcast television news. High/growing reach and low trust is a challenging combination. AI use was concentrated among younger users and higher socioeconomic groups—not irrelevant to the wider publishing audience.
With that growth in AI discovery, Google is piloting a new commercialisation option for online publishers, offering payments to publishers where their content “significantly contributes” to responses in Gemini and AI summaries in search. The money on the table is small, but the precedent matters: Google has conceded that content in an AI answer has an economic value that should flow back to the source.
In parallel, an earlier pilot programme to add AI search and discovery reporting to Google Search Console was extended to every Search Console property at the end of August. This is a really useful data point for GEO, and something anyone running a publisher website should be across.
AI’s impact on rights and translation was highlighted by an unauthorised translation of Haruki Murakami’s new novel, rendered into Chinese in four hours using an AI script called Wenyi created by a literature student with a $2 token budget. Murakami’s previous Chinese translator conceded that it was fluent, but argued it missed the author’s voice. Readers, judging by the reaction, have been less fussy.
Indexing is another specialist area where the same arguments are playing out. Thanks to subscriber Ted Kosan for drawing my attention to the ASI’s recent critical review of Indexia’s AI tool, and the company’s open response to the ASI. I haven’t used the tool in question, but it seems to me both sides make interesting points. The real disagreement is over what to measure: raw output, or edited result. It shows that tensions over expert work, the quality of AI outputs and the pace of change aren’t going away.