Babylon Newsletter, Issue #3

Hi, last week I received feedback from a journalist who had read the previous issue. His verdict: Relevant, but slightly too packed for a quick read. Point taken. We are all overhelmed with AI updates and speculation. Starting with this edition the newsletter is slimmer: Fewer items, written for easier reading.

More money, more problems:Every week someone announces AGI. Meanwhile, the numbers are hard to comprehend — until you see them visualised. Lisa Charlotte Muth at Datawrapper put AI investments in perspective this week: Google's AI budget could build a Burj Khalifa every month. Amazon's could construct a Wembley Stadium every week.

Share of enterprise teams now using multi-provider AI translation strategies, according to a Crowdin survey of 152 B2B professionals (March 2026). The same survey found that 1 in 5 enterprise teams have already experienced a quality incident since introducing AI translation. This means: Humans still needed.

Why care? A new type of middleware is emerging — multi-model language applications working between corporate IT and the big LLMs. There is good reason: Companies using AI services need to ensure they are not locked into one model or one platform without the option to switch anytime. Model quality shifts happen fast in this field.

March 2026 · Slator

What's new: The Language Council of Norway (Språkrådet) raised the alarm that AI-generated errors are becoming normalised in Bokmål and Nynorsk — the two official written standards of the Norwegian language, used by a population of around 5 million. AI-generated errors are being accepted as correct rather than flagged as mistakes, says the Council.

Why care? Norway is one of the most digitally advanced countries in the world, with a strong institutional tradition of language policy. If AI error normalisation is happening there, it is happening in every language community that lacks the volume of training data the major languages enjoy — just without anyone measuring it.

Reality check: There is a need for systematic measurement to track and understand how AI might introduce technology-driven language changes.

2026 · Sources: Ofcom, European Accessibility Act

What's new: In the UK, effective since February 2026, there are now rules for the percentage of subtitled content. The new regulations are just one initiative of many in other countries, pointing into a similar direction: Going towards 100% subtitling on most videos.

Some specific details for the UK, because the rules apply differently to different media.

Video-on-Demand (VoD) Services

Broadcasters (Linear TV)

Why care? More people will be able to actually use the content - which is good. The downside is that most media companies have reduced revenues and will have to bear the costs to comply.

Open question: When the first broadcaster is fined for non-compliance, will regulators follow up on quality on platforms like YouTube as well?

March 2026 · NL Mediacluster 2040

What's new: "De Nederlandse mediasector staat op een kantelpunt" — the Dutch media sector is at a tipping point. A Dutch position paper argues that the media sector should be classified as critical national infrastructure, on a par with energy or transport. Towards the future the NL Mediacluster 2040 paper, by Vivian Opsteegh and Jeroen van Mastrigt, calls for a national MediaTech ecosystem. The framing is economic, not cultural.

Why care? The core argument: if you wait until the infrastructure is gone to decide it mattered, it is too late. That logic extends directly to language tools — subtitling, translation, voice synthesis. If Dutch media is infrastructure, so are the tools that make it work across languages.

Reality check: The paper is in Dutch. An argument for media sovereignty published in a single national language has limited reach in EU policy circles — which is exactly the problem it is trying to solve. Other EU countries need to formulate their own direction.

In issue #2 Mistral CEO Arthur Mensch argued in the FT for a levy on AI companies to compensate content creators. The NL Mediacluster paper argues for protecting the infrastructure that produces that content.

Google is rewriting media headlines
The Verge, March 23, 2026 — Google confirmed it is running an experiment replacing publisher-written headlines in traditional search results with AI-generated alternatives — not just in Discover, where it already does this. One documented case: a rewritten headline endorsed a product the original article criticised. Google called it a "small experiment."

Mistral releases open-weight voice synthesis
TechCrunch, March 26, 2026 — Mistral released Voxtral TTS — an open-weight text-to-speech model, 4 billion parameters, runs on a laptop or smartphone. Nine languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. Voice cloning from under five seconds of audio. A European AI company releasing open-weight voice synthesis at this quality level lowers the barrier for any organisation that cannot afford proprietary pricing or wants to keep voice data on-premise. Nine languages is a narrow slice — but open weights mean the community can extend it.

Google Research: Strong AI model compression without accuracy loss
Google Research, March 24, 2026 — Google published research on TurboQuant, a compression algorithm that reduces AI model memory requirements by at least 6x with no loss in accuracy and no retraining required. Lower inference costs and simpler deployment change the economics of building and maintaining language models for smaller language communities — the ones where the business case for expensive infrastructure has never existed. In February Logical Intelligence made headlines with an energy-based model called Kona 1.0. According to information from the company the system solved 13,000 Sudoku puzzles at a cost of $4 instead of $11,000 for a similar workload on standard LLMs.

WAN-IFRA Frankfurt AI Forum · 13–14 April 2026 · Frankfurt · wan-ifra.org
AI in newsrooms, content creation, distribution. Bilingual DE/EN.

DeepL Spring Launch — #VoiceToVoice · 16 April 2026 · Virtual · deeplspringlaunch.com
DeepL announces three breakthrough advances.

EBU HORIZONS · 5–6 May 2026 · Geneva · EBU headquarters
Media distribution technology strategy. Content reach, broadcast innovation.

TAUS Massively Multilingual AI Conference · 3–5 June 2026 · Rome · taus.net
Focus on European language technology industry.

EAMT 2026 · 15–18 June 2026 · Tilburg, Netherlands · eamt2026.org
Main European academic conference on machine translation.

Languages & The Media · 4–6 November 2026 · Senate House, University of London · languages-media.com
Theme: Moving Images That Move Audiences — Localising with Intent.

A new term worth remembering: prompt injection attack. This is a new type of cybersecurity threat: a malicious input in a webpage, document, or email hijacks an LLM. The more access AI agents have to your data and systems, the larger the attack surface.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany. Three projects I am involved in as innovation manager — you will hear about all of them here:

I cover all three honestly — including when competitors do something better or when our approach has limits.

AI use: Claude (Anthropic) is used to research and edit this newsletter, based on refined and specific prompts. Responsibility for stated facts, names, and links is entirely mine.

Hey, if you made it down here you must be a friend I have not met yet - from journalism, language technology, or AI development. Consider connecting on LinkedIn and drop me a line: what would make this newsletter more relevant for you?

babylon-newsletter.com · Weekly updates

7,000 languages. AI works for 20.