Babylon Newsletter, Issue #15

AI is moving in two directions at once. Adoption climbs every month. So does the list of things going wrong. The question everywhere right now: how do you disclose the level of AI use honestly, and what happens when you do?

This week alone: a German court told Google it is responsible for what its AI writes. Pew says most Americans now read AI answers at the top of their search. And two studies found that telling readers you used AI can lower their trust, even when the work is accurate. It arrives at a time when trust in news, and interest in it, is at a record low.

2 August 2026

This is the day EU law starts requiring you to tell people when they are talking to a machine.

In May, Brussels delayed the hard parts of the AI Act to 2027 and 2028. This rule was not delayed. From 2 August 2026, any AI system that talks to a person must tell them so. AI-generated content must be marked too, though that piece has a short grace period. It applies to you even outside the EU, if your output reaches people inside it.

Practical guide to Article 50 · EU AI Act implementation timeline

June 2026 · Reuters Institute, Digital News Report 2026

What's new: More people are avoiding the news, worn down by negativity. And too many publishers cannot resist speculating, chasing the story that might happen tomorrow. The Reuters Digital News Report puts numbers to it: trust in news has fallen to 37%, the lowest since measuring began in 2015. For the first time, social and video platforms have overtaken news sites. 10% now use AI chatbots for news weekly, rising to 17% among under-35s.

Why care: The audience is moving to AI answers, whether journalists like it or not. Google and the big platforms are pushing it. The real question is whether those systems are reliable. Built on strict rules, they could be better than people at filtering out misinformation. Built carelessly, they could hide bias, spread it, and polarise public debate worse than social media did. It is in the hands of the builders to do the right thing.

22 June 2026 · Nordic AI Journalism summit (Rasmus Kleis Nielsen)

What's new: In his speech Nielsen sorted newsroom responses to AI into four postures, and proposes a "public test": could you explain this use of AI to your readers without flinching?

Why care: Refreshing to see the four side by side, because that is what I observe too, only in real life it comes mixed, not this clear. Many newsrooms are in confusion mode as a result. What would settle it is one genuinely convincing example of agentic AI doing news well. So far there is none.

Reported 9–12 June 2026 · Regional Court of Munich (case 26 O 869/26)

What's new: The court treated the AI Overview as Google's own words, not a list of links. It barred Google from repeating false claims that had tied two publishers to scams the sources never mentioned.

Why care: This puts the blame for a wrong AI answer on the company that built it, not the pages it drew from. The court said the reasoning could apply more widely.

Current status: A preliminary ruling. Google can appeal, and German law sets no binding precedent. The Next Web

Can I use AI, and how far? Where does it help, where does it fail, and when it fails, is that the tool or me expecting too much of it? That is the challenge I've been tackling for fifteen issues of this newsletter.

These troubles are not mine alone. They are the big current questions. Two studies in Digital Journalism, covered by Nieman Lab this week, look at one piece of it: disclosure. Readers say they want to be told when journalism used AI. Being told can lower their trust, even when the work is accurate.

Dietmar Schantin says the question is not whether AI can write journalism. It is who can be asked why. Why this angle, why this source, why this sentence. And who answers for it. AI gives you the text. It cannot stand behind the choice. With no human at the core, you can ask the question but find no one willing to answer for it. That role might be an "AI editor at large": someone checking the prompts that govern the system, tracking its errors, working it towards reliability.

Jesper Nielsen, a localisation expert, wants to drop "human in the loop." The phrase increasingly describes a powerless role: not a guard for quality, but the one who takes the blame. He proposes flipping it to "AI in the loop." In the same LinkedIn discussion, Arle Lommel of CSA Research refined the old idea into a sharper framing: "AI in the loop, human at the core."

Florent Daudens has been testing these "virtual newsroom" systems. He found this: Feed one a dataset and it will find the patterns inside it. What it cannot do is know the part that was never in the data: the context, the reason, the angle a reporter gets by going out and asking. The machine handles what is on the page. The judgment for what matters is still yours.

My view: Information and news people must move faster. We take too long to get to grips with AI. We need a vision for where the ball is heading, what do we need to start now to have it ready in three years? The fight over how to label AI use is blocking a bigger question. Do we know what people actually need to read? Not the next update on Trump or Putin. Information to stay healthy, handle money, choose a school, weigh a diagnosis. I do not care whether it comes from a human or a machine, as long as it has real value for some or many people. Value first, news second.

"Journalism" is an old word for people who wrote daily entries in journals. The work that matters next is not to feed printing presses, but to manage information collections of value: gathering and curating facts, and knowing what people already know, what they think they know, and what they should know. Maybe AI labels become normal. Maybe they vanish, because everything is touched by AI. Either way, the quest is for better information, not hiding the tool, not over-trusting it.

Spotify says more than half its listening is now non-English, with 16 languages in last year's global top 50 [Rest of World]. On AI music, its executive draws the same line this issue keeps hitting: covers and remixes only with consent, and a badge so no listener is fooled into thinking a machine is a person. The growth story is language; the AI story is disclosure.

Mistral released OCR 4, a document-reading model covering 170 languages that runs self-hosted, so files never leave your server. Mistral says the biggest gains are on rare and low-resource languages, where rival systems often fail. Vendor-stated, but for European newsrooms wary of US clouds, it might be the right move.

New York passed the FAIR News Act on 8 June, the first US state law requiring news outlets to label content substantially made by AI, enforced with fines of $1,000 rising to $5,000. It still needs the governor's signature, and sponsors expect other states to copy the text. Where the EU rule (see The Number) targets AI systems, this one targets newsrooms. The same disclosure fight, landing on both continents the same month.

At a European Commission workshop in Brussels on 24 June, Marcus Bösch, who researches propaganda in the age of AI and consults for media and political institutions, will argue one thing: stop calling it "disinformation." He treats it as a condition of the whole information environment, not a single piece of content you can label. If the problem is the water, labelling one drop at a time will not get you far.

DW Global Media Forum — "Journalism out loud" · 23–24 June 2026 · Bonn · At the conference plain X live-transcribes and translates speakers on stage.

GenAI in Localization 2026 — The Builder's Playbook · 23–26 June 2026 · online · The implementation-focused localisation conference, fourth year, on retraining teams, redesigning workflows, and deploying AI responsibly.

WAN-IFRA — Strategic Overview of the AI Licensing Landscape · 24 June 2026 · online • Where publisher content meets AI licensing, with Kevin Anderson and Elena Perotti.

WAN-IFRA / FT Strategies — Future Newsrooms Study 2026 · 1 July 2026 · online · Findings from 448 newsroom leaders across 86 countries.

AIMEDIA 2026 · 5–9 July 2026 · Nice, France · AI in media production, LLMs in creative industries, trust and transparency.

INMA Media Tech & AI Week · 21–25 September 2026 · San Francisco •The LLM ecosystem, content licensing, and the AI content supply chain for news.

Languages & The Media · 4–6 November 2026 · London · This year’s theme is “Moving Images That Move Audience: Localising with Intent”.

I am Mirko Lorenz. I work as an innovation manager on language technology projects at Deutsche Welle in Germany. I am a co-founder of Datawrapper, a data visualization tool.

Three projects you will hear about in this newsletter:

AI use: I use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. The goal is to get help, and to find where AI makes mistakes. It makes them; a diligent author can catch them. Increasingly, working with AI feels like having a tireless assistant, sometimes a whole team of them. Responsibility for stated facts, names, and links is entirely mine. This week I added something new: An open Google Doc where I track problems, mistakes by AI or me.

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