4 November 2026

This week I summarize three new studies on how people use AI, with links. The key point is what current use tells you if you are a publisher or journalist.

But first, a small but notable change is coming for robots.txt. This file sits on every website and tells search bots, spiders and AI agents what they may do. The news: publishers can now say no to AI training with one line in robots.txt. Until now a site needed a separate rule for each AI crawler. On 30 October, OpenAI, Google, Microsoft and Anthropic said they will honour the same line. The line is __usage_ai-training, from an IETF draft. Anthropic has started. Microsoft said it will follow on 1 February 2027, Google said it will follow in early 2027, and OpenAI says "coming months".

That the big AI companies agree to use robots.txt is a step in the right direction, but only a small one. In October I looked up the number of content transactions via RSL (Really Simple Licensing), which also works by extending robots.txt. Finding: no reported deal since the standard was introduced (issue 23). This concession is a step, and negotiations between publishers and AI companies need to continue. Now to the news of this week.

64% of people who use AI to find information say they do it because it is faster and easier. Only 21% say it gives them more reliable information (Reuters Institute, YouGov, about 2,000 people in each of nine countries, several answers allowed).

Why care? Speed drives AI use, and journalists have to accept that. Many editors say accuracy is key, and that is understandable. Newsrooms often argue that "AI is not trustworthy" and point to its mistakes. The concern is real, but it is not what drives use, and it does not keep people from using AI for all kinds of information requests.

29 October 2026 · Reuters Institute

Weekly use rose from 34% to 51% in a year, in the six countries with data for both years. Only 12% are comfortable with news made entirely by AI, 19% if humans oversee it. And 17% of those who use AI for news say they used it in the week before the survey to get around a paywall. It is an online panel with self-reported answers, so read it as direction, not as a count. Getting around a paywall is interesting. A publisher or journalist could find a new path to provide access and create a package that people accept, even with a subscription. Simplification and faster delivery may be the elements to start with. Full report.

October 2026 · Rutgers University

State of the Prompt is yet another new study. It is a huge collection of real usage and complements the Reuters findings. Reuters asks people why they use AI. Rutgers researchers looked at what people actually did, in 3.4 million ChatGPT conversations from an opt-in panel in the US, July 2025 to June 2026. Health is the largest of 28 single subjects, 10.2% of conversations, and 49% of people had at least one health conversation. The chart below groups the subjects into eight categories. In political conversations, 0.22% ask for evidence. Topics were sorted by a language model, and the panel covers one tool in one country. The study lets you explore how people use AI interactively: State of the Prompt.

22 October 2026 · Tow-Knight Center, CUNY

I like the name of this study, "The Three-Layer Cake". It is memorable, and that helps, because so much new information is published in the AI boom. Gina Chua and Adiel Kaplan bring together what they learned in a year of conversations with journalists, technologists and funders. Their starting point: news will be intermediated by AI, whether we like it or not. They map the information ecosystem in three layers: consumption, synthesis and creation. And they show where newsrooms, technologists, funders and allies can act while the new infrastructure of information is still being built.

Two key findings:

It is a synthesis of a year of conversations, not new data. Report.

Based on the studies above, here is a short summary of what we can learn from current use. A core concept is the desire path: people will find a way to get the information they look for. That path may not be what publishers want, with their paywalls, advertising and so on. But the desire path is a powerful concept if it is understood and applied. Give people what they want in the age of AI, as simply as possible.

A landscape gardener plans one paved path. A year later, a trodden line across the grass shows where people really walk. This is what a desire path looks like in real life. It is not about UI/UX, nor aesthetics or design. It is a testament to what people really do. The three studies above show the desire paths of AI use.

People walk to AI because it is easy. 64% say faster and easier, 21% say more reliable. 17% of AI news users used it to get around a paywall. And 18% have paid access to AI, the same share as have paid access to news (Reuters Institute). They bring their own lives: work, money, tech, games, home and health. Almost nobody checks the answers.

Newspapers paved their old path for years when the web arrived, and many are under severe pressure now. My view: the same mistake is open again. On many news sites I visit, a reader passes a cookie banner, a login and a wall of ads before the first paragraph. Next to a chat window that answers in seconds, most news offerings are an obstacle. The task is to reduce friction and simplify information offerings, so they are lean and digestible for overstimulated readers.

The authors of The Three-Layer Cake point in the same direction. Three more of their key findings:

For journalists, content creators and large publishers my recommendation is: think ahead and allow for simple, new solutions to present information. Do not let the existing problems and mistakes of AI blur the picture. Similarly, do not let the status quo of production block the path to something new. People crave information, and that need refreshes every morning. Be where the answer appears. Work out what will get your offering there, how easy it is to access, and what value it has beyond being easy to use. Start with one test: a reader or an AI should get the answer from your page in seconds.

Limit: all figures come from online surveys and one US opt-in panel. They show direction, but not the solutions yet.

Meta's terms meet a court and a regulator. The Oslo District Court stopped them for Norwegian publishers on 28 October, and the UK CMA opened an investigation on 30 October. Meta says the 20 November start may slip.

86% followed the agent. In an Anthropic study of 27 October, 50 analysts followed an AI agent's hidden lean on four analysis choices 86% of the time, and only 15% credited it. Vendor research on its own product, one task. I use Claude.

Rest of World will stop as a standalone newsroom. It plans to close at the end of January 2027 unless a merger partner takes it, Semafor reported on 25 October. Reported causes are falling search and social traffic.

Languages & The Media · 4 to 6 November 2026 · London · UCL page · I speak on Friday, 6 November at 14:00, on a panel on Workflows & Technology. If you are there, find me and say hello. I am happy to meet for a coffee. Watch whether anyone shows numbers on quality.

Hacks/Hackers AI Real Talk with Gina Chua and Adiel Kaplan · 12 November 2026, noon ET · online · link · Watch for concrete proposals on standards that carry sourcing with facts.

Language Intelligence 2026 · 17 to 19 November 2026 · Vienna · link · Watch for independent quality data, not product launches.

A small, notable data point from the studies this week: in political conversations, only 0.22% ask ChatGPT for evidence. Source: State of the Prompt, Rutgers University, 3.4 million ChatGPT conversations from a US opt-in panel.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany.

Three projects you will hear about in this newsletter:

AI use: Full disclosure, I do use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. My goal is to find out where AI is reliable and where the hallucinations come in. Before publishing I check all facts and links.

Error log: To track the issues of using AI I maintain an open Google Doc where I collect the small and big problems that a tool like Claude introduces to editorial work. The idea is to get a better understanding of where AI is good and where it is not. Read it here.

babylon-newsletter.com · 7,000 languages. AI works for 20.