Babylon Newsletter, Issue #16

AI improves gradually, and our thinking about how to use it should too.

Instead of going back and forth on the problem of the week, the mistakes, the hallucinations, we need a more systematic view, so it moves our organisations forward.

The question is not whether to use AI for transcripts or translation. It is what the translation is for. Where does it create a real advantage? How is an audience served better by it?

This issue is about deciding that yourself, instead of hoping the AI will tell you.

The official EU languages a new open model has promised to cover. On 19 June the European Commission picked the EUROPA consortium, led by Italy's Domyn: more than 400 billion parameters, run on European public supercomputers, weights given away so governments can self-host. Domyn, founded in Milan in 2016 by Uljan Sharka and called iGenius until last year, builds sovereign AI for regulated industries and has already built its own supercomputer with Nvidia.

For scale, 400 billion is the class of the largest open models, like Meta's Llama 3.1 at 405 billion. Far bigger than GPT-3. Not provably better than ChatGPT or Claude, which do not publish their size. Size is not quality.

Then the small languages. The 24 are the official ones. Basque is not on the list. Nor Catalan, nor Welsh. The languages that need the help most are still outside it.

India: Something comparable is happening there. Sarvam raised 234 million dollars for 22 languages. The plan is to build natively, not translate from English.

Sources: European Commission on EUROPA · TechCrunch on Sarvam

June 2026 · NewsGuard

What's new: NewsGuard launched a chatbot, at newsguard.ai and as a widget publishers can embed on their own sites, that answers only from about 12,000 sources its journalists have rated reliable. It cites and links every source, runs answers against a catalogue of 64,000 debunked false claims, and shares revenue with the publishers it quotes. It costs six dollars a month, is free to start, and does not train on publisher content. French, German and Italian versions are due in September.

Why care? The founders, Steven Brill and Gordon Crovitz, built it from NewsGuard's existing source-rating business and pitch it as the opposite of ChatGPT: pay publishers, cite them, refuse the open web. Instead of labelling how much AI is in the output, it limits what the AI can draw on.

Current status: NewsGuard is a for-profit, and advertising agency Publicis was a lead launch investor. Whoever decides which 12,000 sources are reliable holds the real power, if this project grows. The company has faced US bias complaints over exactly that.

Update to Issue #14 · 9–25 June 2026 · Press Gazette / Techdirt

What's new: Local news sites blocking the Internet Archive have risen from 241 to about 340. The cease-and-desist against Common Crawl hit a wall: its archive cannot be edited after publication, so it filters URLs from its tools rather than deleting them.

Why care? Common Crawl is the data floor under most large models, and it is mostly English. When these archives close, the part that thins first is the local, regional and non-English record that was already thin.

Update to Issue #14 · June 2026 · Google / practitioner reaction

What's new: Google's Gemini 3.5 Live Translate is the fourth frontier-lab live translator in two months, after OpenAI, Zoom and DeepL. The launches no longer make headlines.

Why care? Elena Chernysheva, who co-founds the dubbing firm Dubformer, splits the work in two. Utilitarian: a call, a meeting, done once meaning crosses. Media: a film, a match, where timing and voice carry it. The first is going free. The second, she says, the live tools cannot hold past a few minutes.

Start with the real question. Last week a large media company asked us about the real productivity gains from a system like plain X. The question was focused almost entirely on one number: how fast can a translation be done? That got me thinking about how organisations could and should answer it.

AI is the tractor: Powerful, but only one element. Picture AI as a tractor arriving on a farm that had no machinery at all. It goes to work and plows the whole field, fast. That is what AI does for content. But the machine does not tell you how to be a good farmer. The tractor knows nothing about the quality of your soil, how much water the crop needs, which crop to plant, or whether there is a market for it. Translation is the same. The engine produces fluent output fast. It cannot see the mistranslated legal clause, the idiom that lands wrong, the term that is correct and contextually wrong. The machine does the visible labour. The judgment is still yours.

Measure the whole field, not the tractor. Most AI productivity claims time only the machine step: the transcription that took a day now takes an hour. The real sum counts everything around it. Time the full job as it runs today:

Add it up, and the tractor is often a small part of the total.

Then improve in small steps. That is not a reason to drop the tool. It is the reason to measure before you claim a saving, then improve the whole process in one-percent steps, not one big bang. You are not alone if there is no system in place, yet:

In order to make this a bit more practical I have put together a free Google Sheets template for mapping a whole workflow this way. It is a first step and still rough, but it gets the calculation going. Take a look, then enhance the whole thing to fit your needs. Share results, eager to see how the whole approach changes for you.

The other side of sovereignty. The AI Now Institute argues Europe's AI winners still send value to US firms, because they rent the model, the cloud and the distribution. Max von Thun of the Open Markets Institute makes the matching point against the "Europe 2031" scenario (Issue #14): right diagnosis, wrong cure, because more dependence is not leverage.

Institutions are hiring human translators again. The World Bank and the European Central Bank are both recruiting in-house language teams. Where accuracy cannot fail, the human stays.

New startups skip the translation tool. Slator's 2026 list finds newcomers launching as orchestration platforms, not engines. A notable trend: Specialists are going after the languages big labs ignore.

Slator's analysts write: “Much of the language AI market conversation is focused on broad, multilingual workflow ownership within a single environment. Yet specialization remains a viable strategy. Alongside broad workflow platforms, this year’s cohort includes a number of startups focused on narrow but strategically important capabilities.”

Who Gets to Crawl? · 30 June 2026 · online · Ezra Eeman and Kevin Anderson with Cloudflare's Sam Else, on publisher control over AI crawlers.

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

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

LocWorld56 · 19–21 October 2026 · Vancouver · The localisation industry's largest event, with the Multilingual AI track.

Languages & The Media · 4–6 November 2026 · London · "Moving Images That Move Audiences: Localising with Intent." Disclosure: I have a speaker slot.

Publications, old and new, need to think about Category Entry Points. Grzegorz Piechota of INMA says reach is no longer a count of people on your site. What matters instead is whether your brand is part of the situations: the morning catch-up, the commute, the clarification someone wants about mortgage rates, the five minutes with a puzzle or a small game. We are losing the click, which gave us big numbers but not much connection. 50,000 real readers may soon be worth more than 500,000 who only ever passed through.

I am Mirko Lorenz. I work as an innovation manager on language technology projects at Deutsche Welle in Germany. I co-founded Datawrapper, a data visualisation 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. Responsibility for stated facts, names, and links is mine. I also keep an open Google Doc tracking problems, mistakes by AI or by me. This is to learn more, week by week.

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