Babylon Newsletter, Issue #27: Simplify, simplify

Please note: this newsletter got shorter. The goal is to make it easier to read. There is so much information about AI, content and language. Easily too much. a quick overview that gives you a (partial) sense of what is going on in the field. No more, no less.

Last week I was in London for Languages & The Media. Translators, subtitlers and researchers met for three days, and AI was everywhere. But the tone was more cautious than at tech events. Germany has a big tradition of dubbing films, and dubbing professionals said even very good voice clones lose a lot of nuance.

In my talk I reported on what the Research and Cooperation team (ReCo) at Deutsche Welle learned from co-developing plain X with Priberam. Three years ago DW made it the official tool for content adaptation. It runs on several engines, with benchmarks per language pair, so DW can move along as engines get better. The goal is to offer more, by adapting existing content faster. Last week DW reported a record of 406 million users a week, up 20%. Almost 80% of them live in countries where press freedom is restricted. If you would like my slides, send me your email. (Languages & The Media, DW)

Anthropic signed its first paid licensing deal with a publisher, the Associated Press. On election night, Claude users who asked about the results got live AP data. The terms are not known.

My take: Anthropic is paying for fresh data, not for the archive, and this could point to the future. Large parts of what we know today will be outdated in a year. That is the big opportunity for news gathering and research: deals with the AI engines based on how the world has changed. (Nieman Lab)

Less friction, fewer distractions: possibly the biggest opportunity for journalism

There is a big opportunity right before our eyes: simplifying news pages and information packages, by getting rid of all the friction, pop-ups, calls to action and the myriad of other things standing between the news and the user.

That there is a problem is not new: Developer Shubham Bose opened the New York Times website to look at four headlines. His browser made 422 requests and loaded 49 megabytes. (Bose)

John Gruber of Daring Fireball published a screenshot of a Guardian page where hardly any news could be seen. (Gruber)

When I open the NYT (which I do almost daily), it wants me to consider a family subscription. I have clicked past this offer maybe 50 times, but it keeps showing up. (The main reason is that my daughters would not want to read an English source. But I will never be able to tell that to the NYT.) And then the cookie banner, then the offer to download the app. It never stops and creates massive friction.

Both publishers and readers might see all this noise and distraction as inevitable. But only for a moment. The first alternative offering will make this setup look stupid. Agentic AI could do just that. Meta is already giving its assistant a memory, so it knows what you care about. Preferences could be stored once. Someone might want the news plus fashion plus cooking, but none of the distractions. AI could even switch the writing style, from smart brevity to elegant long form, depending on what the reader prefers.

Users could also do it themselves, and this is where it gets really bad for publishers. To practice what I preach, I spent about an hour creating a mockup of a decluttered, very personal news update. See below.

The message is simple: Start using AI now to build simpler, better ways to read your journalism. Try one clean page, one personal briefing, one format in a different style. Push it forward this year, not next year. Otherwise someone else’s agent will do it for your readers.

Fake Mark Cubans on Facebook. AI fakes of Mark Cuban sell a $294 "diabetes cure" to older people on Facebook, and he calls the platform "the worst of the worst". Internal Meta documents, reported by Reuters in 2025, estimated that about $16 billion, 10% of Meta's revenue, could come from ads for scams and banned products. Meta says the estimate is too broad. In October 2026, Rhode Island became the first US state to sue Meta over scam ads. (FinanceBuzz, Rhode Island AG)

A voice actor against Netflix. Matti Klemm, the German voice of Jason Momoa, used a data request to find out that Netflix trained AI on his voice. A lawsuit is planned for the end of the year. (Slator)

Who measures which model works best. The language industry is building its own AI leaderboards, says Konstantin Dranch (Custom.MT). More and more of them let AI judge AI, which is fast and keeps up with new models, while human evaluation is slower but builds more trust. (LinkedIn)

Futures of Journalism. Nordic AI Journalism and the Fojo Media Institute have published eleven essays on journalism when AI sits between news and readers. I liked Shuwei Fang on machines that turn information into decisions, and Kasper Lindskow on selling trustworthy knowledge instead of finished articles. (anthology, Fang, Lindskow)

Hacks/Hackers AI Real Talk with Gina Chua and Adiel Kaplan · 12 November 2026 · online · link

Language Intelligence 2026 · 17 to 19 November 2026 · Vienna · link

Newsrewired · 25 to 26 November 2026 · London · link

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.