Covering AI, language technology and the future of content.
7,000 languages. AI works for 20.

Le Monde (665,000 paying subscribers) has made deals with AI platforms — first ChatGPT, then Perplexity and Meta. The terms are not public. But CEO Louis Dreyfus says stories surfacing on ChatGPT convert to paid subscriptions 20 times more than on Facebook, and 50 times more than on Google Discover. He is urging other media companies to follow.

Media analyst Thomas Baekdal is skeptical that one-time deals like this work over time. His argument: usage-based licensing — where compensation scales with actual AI demand for your content — produces better long-term returns than fixed agreements negotiated before anyone knew what the traffic was worth.

Why care? Every publisher is trying to solve this equation. There is some genuine hope here — AI interfaces may convert readers to paying subscribers better than social media ever did. But there is a deeper reason the formula matters: if AI trains on copies of copies, original verified journalism becomes the scarce input everything else depends on. Its value is not going down. It is going up. When we will know which model wins: not yet.

Sources: Baekdal · Press Gazette · Wayfinder #163, Ezra Eeman

March 20, 2026 · Google Developer Documentation / Vantacron / Thomas Baekdal

What's new: On March 20, Google added a new user-agent called Google-Agent to its official fetcher list. It bypasses robots.txt completely.

Why care? For context: robots.txt has been the standard way to tell crawlers about the terms of access to a website since 1994. Google NotebookLM uses the same exemption — a user can point it at any publisher URL, extract the content, and build a personal news reader without ever visiting the site, seeing an ad, or encountering a subscription prompt.

What we know, what we don't: Here is the problem. Google says: if a person asks an AI — like Gemini — to retrieve a page, the fetch is the equivalent of that person visiting that page in a browser.

Thomas Baekdal now suggests blocking user agents at server level, no longer with robots.txt. Baekdal's suggestion for the new situation: Server-side authentication is the only reliable alternative left for publishers who want to restrict access.

The question is whether smaller publishers can afford that infrastructure, and whether regulators will treat "user-initiated" as a sufficient legal distinction when the economic outcome for publishers is identical to a crawl.

Sources: Google documentation · Vantacron · Baekdal

April 15, 2026 · The Economist

What's new: Anthropic launched Mythos in April, available only to a closed group via "Project Glasswing." OpenAI followed the same week with a restricted version of GPT-5.4. Mythos is priced at five times Anthropic's most powerful publicly available model.

Why care? The Economist identifies three drivers: protecting models from distillation by competitors, rationing scarce compute, and shifting power back to model-makers. When developers cannot access a model, they cannot build around it — pushing them toward the model-maker's own tools. The gating of frontier capability is also the gating of the ecosystem around it.

What we know, what we don't: The AI companies see growing usage, but that only partially solves the problem of high costs and investments. The staged rollout is also an experiment in premium pricing. There are safety concerns about Mythos's cybersecurity capabilities, but this could even be a form of marketing - to see how high prices can be if a technology is not available for everyone. What we don't know: whether this staged, exclusive rollout becomes the permanent model for frontier AI, or whether competitive pressure eventually forces broader access.

Source: The Economist, April 15, 2026

April 2026 · Ole Reissmann / THEFUTURE · Erin Grau / San Francisco Standard

What's new: The San Francisco Standard launched an AI-powered app exclusively for paid subscribers. It offers AI briefings, a personalised news experience, and continuous story updates without waiting for a new article to be written. The same week, Business Insider launched an AI news briefing — click a button, get short audio summaries, dive deeper if you want.

Why care? Two mid-sized newsrooms, same week, same signal: the article as the primary unit of journalism is being replaced by something more fluid. Not because editors decided this philosophically — because paying subscribers are telling them what they will pay for.

What we know, what we don't: We know both products exist and that the Standard is gating the app behind a subscription. What we don't know: whether AI briefings convert casual readers to subscribers, or whether they serve subscribers who were already paying. We also do not know what is actually driving the value for users — the AI summary itself, the brand trust of the newsroom, or the quality of the underlying reporting. If it is the latter two, the AI layer is a delivery mechanism, not the product.

Sources: Ole Reissmann / THEFUTURE · SF Standard

April 16, 2026 · DeepL

What's new: DeepL Voice — real-time speech translation in two modes. Voice for Conversations: live translation between two people on a smartphone. Voice for Meetings: Zoom and Teams integration, each participant speaks their own language, others see subtitles or hear an AI voice. Free to try, €24/person/month for enterprise. You can test the features right on the website.

Why care? Live transcription and translation is becoming standard infrastructure. At conferences, in multilingual editorial calls, in cross-border collaborations — the technology is already changing the experience noticeably. The meetings integration reduces the need for a human interpreter in working-level conversations.

What we know, what we don't: DeepL is expanding beyond text translation into spoken language. But they will not be the only option in Zoom or Teams, very likely those companies will level up and provide similar features. What is missing: independent quality benchmarks across language pairs — especially non-English combinations.

Source: DeepL Voice

How the attention economy is designed to divide us — and what could change.

On April 16, the European Commission's Joint Research Centre published a study worth your time. It is called Fractured Reality. Lead authors: Mario Scharfbillig and Stephan Lewandowsky. It runs 78 pages. The analysis is sharp and — unusually for a report about the information crisis — it leaves you with the feeling that change is possible.

The core argument: dis- and misinformation, deep fakes, and manipulative content are not a content problem. They are a design problem. Like once on an extreme view — either right-wing or left-wing — and the algorithm gives you more of that. You might be flooded relentlessly with more of the same. Stay on that information diet long enough and the other side's argument becomes genuinely hard to see. Agreeing on facts is not even considered anymore. Finding common ground becomes impossible as a result. In short: this chain is salting the earth.

The mechanism is the attention economy. Current platforms are not built to inform. They are built to maximise time on screen. The goal is to sell advertising. This means the algorithm rewards what keeps people scrolling — which turns out to be content that is negative, emotional, and conflictual. Not because anyone planned it that way. Because that is what human attention responds to under overload.

The numbers in the report are worth reading closely. Untrustworthy sites make up less than 1% of German news diets. On Facebook in France and the US, they account for nearly a quarter of engagement. Twenty percent of Facebook users with the most conservative news diet generate 62% of all exposure to fake news. The problem is not that Europeans are drowning in false information. The problem is that a small number of actors, amplified by algorithms owned offshore, can reshape the information space.

For anyone building language infrastructure in Europe, one argument stands out. No EU regulation fully works while the infrastructure sits elsewhere. The report lists alternatives with better approaches: Eurosky, the FediVerse, Mastodon's ATmosphere protocol, and — closer to this newsletter — Apertus from ETH Zürich and EuroLLM-22B, both trained on European corpora under European terms. All of these could be building blocks for a sovereign stack.

And when generative AI ends up woven into social platforms — Grok on X, Meta AI on Facebook, Sora on its own feed — language and voice systems become the distribution layer for everything else the attention economy does. The language layer is not a side-effect of the democracy problem. It is the mechanism.

The interventions the authors propose are concrete: a progressive tax on digital advertising that scales with firm size, user rights to bring their own algorithms into the infrastructure where their data sits, markets where speakers stake something on the accuracy of their claims, and public procurement that actually favours sovereign tools.

The JRC report does not ask platforms to be better. It argues the business model itself is the harm, and that European democracy cannot be rented from companies that profit from breaking it.

Read it: Fractured Reality — How Democracy Can Win the Global Struggle Over the Information Space. JRC144603. Free download.

Stanford HAI AI Index 2026 The 2026 AI Index, published April 13, tracks AI capability, adoption, investment, and safety across more than 400 pages. Key signal: the most capable models are now among the least transparent. Average Foundation Model Transparency Index scores dropped from 58 to 40 points year-over-year. Documented AI incidents rose to 362, up from 233 in 2024. Worth having as a reference document. Full report · PDF

DW Akademie Hype Literacy Toolkit Hype is not new in journalism. But AI coverage has made the problem acute: too many articles tell readers what could be rather than what is. DW Akademie built the Hype Literacy Toolkit to help journalists recognise and resist speculative framing — in AI coverage and beyond. The tendency to extrapolate from incomplete evidence is not dishonesty. It is a trained habit that readers notice and distrust. The toolkit addresses that directly. Free to use. DW Akademie is presenting it this week at the Perugia Journalism Festival. Disclosure: DW Akademie is part of Deutsche Welle. Hype Literacy Toolkit

Four Sessions from Perugia - The International Journalism Festival in Perugia draws an international audience of journalists each year. It is a great conference. Talks are happening in hotels all over town. But — if you have been there you know the problem: often there are no seats left for the interesting sessions. Below are links to four sessions from last week (April 15–18) about AI, journalism and future concepts.

Click the link, watch the recording. No queue. No espresso in the sun on a plaza in Italy either. But the content is worth it.

After the Reader: What Comes Next for News in an AI-First World? What if journalism's primary customer is no longer the reader? David Caswell (Storyflow), Florent Daudens (Mizal AI), Shuwei Fang (Shorenstein Center), Lucky Gunasekara (Miso.ai).

What Future for Journalism in the Era of AI? Council of Europe frameworks, editorial oversight, responsible AI implementation. David Caswell (Storyflow), Natali Helberger (University of Amsterdam), Felix Simon (Reuters Institute), Giulia Lucchese (Council of Europe).

Bargaining with Big Tech: How to Build a Better News Ecosystem in the AI Era Behind-the-scenes deal-making from AI leaders at three major publishers. Pablo Delgado (PRISA Media), Ana Jakimovska (Mediahuis), Caspar Llewellyn Smith (Guardian). Moderated by Lucy Kueng (RISJ). No longer a copyright debate — a battle for the discoverability of public-interest journalism in an era of synthetic content.

How Do You Edit Responsibly When AI Outcomes Are Unpredictable? Olle Zachrison, former editor-in-chief, raised this at Perugia last week. The question is precise: not whether to use AI in editorial work, but how to maintain accountability when the output is probabilistic rather than deterministic. Panel also included Phoebe Connelly, Erja Yläjärvi, Mukul Devichand.

Epistemia. A new word for a specific problem of AI. It describes the tendency to confuse linguistic form with reliability. When an AI system produces coherent, authoritative-sounding text through statistical patterning alone — with no understanding behind it — the reader experiences the illusion of knowledge. The trap: believing the machine because it sounds smart.

The concept comes from a peer-reviewed study published in PNAS in October 2025: "The Simulation of Judgment in LLMs" by Loru, Nudo, Di Marco, and Walter Quattrociocchi. The JRC's Fractured Reality report picks it up and applies it to the broader information environment. Once you have the word, you start noticing how often it applies.

PNAS research article · Ivo Silvestro, The Late Thinker

EBU Horizons May 5–6, 2026 · Geneva · EBU headquarters Media distribution technology strategy. Content reach, broadcast innovation. https://ebu.ch

TAUS Massively Multilingual AI Conference June 3–5, 2026 · Rome European translation technology, MT quality, AI orchestration. https://www.taus.net/events/massively-multilingual-ai-conference-rome-2026

AI Without the Guesswork June 16, 2026 · Online · 6 hours · 20 spots Henk van Ess (GIJN): how to use AI in journalism without feeding your own assumptions back to yourself. https://gijn.org/academy/on-demand-course-journalism-ai-prompting-verification/

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.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany. This newsletter aims to inform and connect the community of people interested in AI, language technology and the future of content.

Three projects you will hear about in this newsletter:

AI use: I use Claude (Anthropic) for research and to edit this newsletter. My goal is to get help while researching, editing. At the same time I find out where AI makes mistakes and draws wrong conclusions. Responsibility for stated facts, names, and links is entirely mine.

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

7,000 languages. AI works for 20.