Less than 1%

That is the share of pageviews AI platforms are returning to publishers — including ChatGPT, which grew more than 200 percent year-over-year, according to Chartbeat. AI models rarely credit sources unless explicitly asked.

Why should I care? AI reads and ingests everything, charges for subscriptions, but sends back nothing — this model is not sustainable. Revenues at large AI companies already run into the billions. The time to negotiate is now, before the defaults are locked in. If journalism, translation, and serious content creation dry up, future AI systems will have less to learn from. Two responses to this problem are covered below: a proposed European content levy from Mistral, and RSL, an open standard that lets publishers define their own licensing terms.

23 March 2026  ·  Slator

What’s new:  There is weekly speculation which professions might be negatively affected or replaced by AI. A Slator study places translators among the most AI-exposed professional groups — and less able to adapt than most.

Why it matters: The core work of a professional translator is not word-for-word transfer — it is capturing what a text meant to say. Oxford researchers estimated roughly 28,000 US translator jobs were suppressed by machine translation between 2010 and 2023, before generative AI entered the picture. An Acolad survey in 2025 found 84 percent of professional translators predict falling demand.

Reality check: Post-editing AI output is often as much work as translating from scratch. Errors are plausible, not obvious — catching them requires more domain expertise, not less. Past disruptions suggest new markets can end up larger and needing more experienced people. Cold comfort for a translator today, but relevant for organisations making decisions about language teams right now.

19 March 2026  ·  Meta AI Research

What’s new:  Meta published a research paper on Omnilingual Machine Translation (OMT). Headline claim: 1,600 languages. The paper itself says the system understands around 400 languages “sufficiently well” — meaning it can convey core meaning in most cases. Production-reliable translation exists for roughly 20–30 language pairs.

Why it matters: One massive model stretched across 1,600 languages is one answer. Breadth and quality pull in opposite directions. A model covering 1,600 languages can very likely not be deeply tuned for any of them.

The Huniki Federation represents another: community-rooted language technology companies in Africa, each building and maintaining AI tools for their own language communities, with native speaker involvement and local ownership. The focus is on one language at a time, not many.

Reality check: The OMT concept is a research paper, not a deployed product. At the same time the Huniki community model needs sustained funding — harder to secure than a single large research grant. European organisations managing smaller or regional languages should look at what Huniki has built.

14 March 2026  ·  ElevenLabs

The product I reported on from a webinar last week is now live. Klarna and Deutsche Telekom are cited as early enterprise deployments. The system claims support for 70-plus languages. Emotional nuance in a second language under real stress — a cancelled flight, a billing dispute — is a harder problem than a controlled English demo suggests. That test has not yet been reported publicly.

The compensation gap is a genuine problem. Two proposals are now in active discussion. They approach the same issue from opposite directions — one mandatory and top-down, one voluntary and bottom-up.

Mistral: a mandatory European levy.  On 20 March, Mistral co-founder and CEO Arthur Mensch wrote in the Financial Times proposing a revenue-based charge on all commercial AI providers placing models on the European market — between 1 and 5 percent. Proceeds would fund content creators and the cultural sector. In return, AI developers would gain legal certainty for training on publicly available material.

This is not a publishers’ lobby speaking. It is a European AI company arguing that legal fragmentation disadvantages European developers relative to US and Chinese competitors operating under permissive or non-existent copyright rules. The logic is economic: the same principle as ASCAP and BMI in music licensing, not a moral argument.

Open question: Big US AI companies will resist. Finding legal and financial form for a levy like this could take years. The main haul of training data has already been scraped. A mechanism focused on ongoing use, not just historical training, is still worth building — if agreement is possible.

RSL — Really Simple Licensing: the voluntary layer.  RSL is an open standard, launched in September 2025 and reaching version 1.0 in December. It lets any publisher attach machine-readable licensing terms to their content — free access, attribution required, pay-per-crawl, pay-per-inference, or no AI use at all. By December 2025, more than 1,500 organisations had signed up, including The Guardian, Associated Press, Vox Media, Cloudflare, and Creative Commons.

RSL is what robots.txt should have been: a licensing standard, not a blocking standard. The model is B2B infrastructure — AI companies paying publishers automatically at scale through a standardised protocol.

Open question: RSL is not a technical access control. A bot that ignores it can still ignore it. Whether the RSL Collective builds real negotiating leverage depends on adoption at a scale not yet tested.

Both proposals are early-stage responses to a structural problem. Watching how they develop means watching whether the content economy finds a new equilibrium — or repeats the familiar pattern of displacement first, reckoning later.

Webinar: WAN-IFRA is hosting a webinar on AI content licensing today — “The Power of the Collective: Navigating AI Partnerships & Content Licensing” — 17:00 CET. It will likely be recorded and available afterwards.

Schibsted open-sourced Videofy — GitHub. The Nordic media group built an internal tool that converts news articles into short videos automatically, used it across its brands, then released it. The honest verdict: it is a base layer, not a finished product. Worth watching for newsrooms producing multilingual content who need format conversion, not just language conversion.

DeepL Spring Launch — 16 April.  A virtual launch event teasing “three breakthrough advances” under #VoiceToVoice. Worth a calendar entry: deeplspringlaunch.com.

TAUS Massively Multilingual AI Conference — Rome  3–5 June 2026  ·  taus.net  —  The main gathering for translation technology and MT quality on the Babylon beat in the first half of the year.

EAMT 2026 — Tilburg, Netherlands  15–18 June 2026  ·  eamt2026.org  —  The main European academic conference on machine translation. Keynote: Rachel Bawden on LLMs and translation for low-resource languages.

WAN-IFRA World News Media Congress — Marseille  1–3 June 2026  ·  wan-ifra.org  —  AI and content rights are central themes. The RSL and Mistral levy debate will likely surface here.

Languages & The Media — London  4–6 November 2026  ·  Senate House, University of London  ·  languages-media.com  —  Theme: ‘Moving Images That Move Audiences: Localising with Intent.’ The main European conference on audiovisual translation and media accessibility.

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.

Let me know how you perceived this issue. Hit reply. Tell me what you are working on, what tools you are testing, what questions need answering. If you happen to have colleagues or friends who work in language technology, please forward this issue to them.

See you next week.

babylon-newsletter.com  ·  Every Tuesday at 10:00 CET

7,000 languages. AI works for 20.