Hello from week nine. I know you are busy, this issue is shorter and reads in less time. I pulled out what I believe is relevant to people in editorial, content adaptation and content development. Let me know if this has value to you.

By mid-2025, roughly 35% of newly published websites were classified as AI-generated or AI-assisted. Up from zero before ChatGPT launched in late 2022.

Why care? Some people are very optimistic about AI on the open web, some pessimistic. This paper is one early snapshot. It shows AI-generated text correlates with less semantic diversity and more positive sentiment, but not with the loss of factual accuracy or stylistic diversity that most people assume. Content type matters — a separate EBU/BBC study of AI handling news found significant inaccuracies, so the picture differs by domain. What AI content will bring needs more observation and patience before any verdict is possible.

Source: “The Impact of AI-Generated Text on the Internet”, by Jonas Dolezal, Sawood Alam, Mark Graham, Maty Bohacek

DeepL announced 250 layoffs on 7 May, roughly a quarter of its workforce. CEO Jarek Kutylowski framed it as a deliberate restructure, not a crisis — fewer layers, faster decisions, made in preparation for the AI era. The company is reportedly preparing for a US IPO, and cutting headcount before an IPO is a familiar way to dress up the KPIs investors want to see. Sifted reports the cuts followed an internal analysis of how DeepL should operate going forward. Last November the company launched an AI Agent and Customization Hub, moving outside translation into CRM and customer service automation. DeepL, based in Cologne, is one of the few EU unicorns — it raised $300m at a $2bn valuation in 2024. None of that helps the 250 people now out of a job.

Bloomberg · Sifted

OpenAI released three new realtime voice models this week. GPT-Realtime-Translate handles 70+ input languages, 13 output languages, priced at $0.034 per minute. Named pilots include Deutsche Telekom, Vimeo, Zillow and Priceline. The translation pipeline still runs speech → text → translation → speech, but the latency is mostly gone.

This is the wider context for DeepL's caution. A successful European specialist now sits against well-funded US generalists — OpenAI, Google, Microsoft, Anthropic — who are happy to absorb every adjacent use case on the way to becoming the one-stop shop. Being good is no longer enough. Being the only option for a specific job is.

OpenAI

INA's data.ina.fr has applied AI to nearly two million hours of French radio and television. The platform provides structured indicators on gender balance, geographic coverage, vocabulary trends and personalities across a decade. Every output is reviewed by editorial staff. Errors are flagged with orange warning icons — entity recognition still confuses the aircraft carrier Charles de Gaulle with the former president, and the team says so openly. This is what AI in newsrooms looks like when it is doing work humans cannot do at all, paired with editorial discipline that documents its own limits. JournalismAI

Jaemark Tordecilla, a Filipino journalist, media adviser and technologist, built a full data visualisation dashboard for the Philippines' 2025 National Demographic and Health Survey in one week. He used two coding agents, Claude Code and Codex, working alone.

What makes his piece worth reading is what he says did not work. There is a lot of speculation right now about how AI can help in content and code creation. This is a practical test, and it reveals what actually matters. The first ninety percent of the dashboard took a day. The last ten percent took the rest of the week.

Editorial judgement and the experience of the human creator were crucial to making any of it hold together.

The take-away is not that AI replaces humans. It is that small newsrooms — local ones especially, where the data and code team exists only as a dream — can now do things that used to require that team. This is one iteration, but it points to something hopeful: an expansion of available information, and a real chance for local and regional reporting to come back, with data analysis and code development included.

For more on this — specifically the future of local newsrooms — Ulrike Langer has a recent essay with further (positive) examples.

Generative AI in the Newsroom

LINGUA Africa — open call for African-language AI datasets — Microsoft, Gates Foundation, Google.org and the Masakhane African Languages Hub launched a funding call on 7 May, deadline 15 June. Cash plus compute credits, capped at $250k cash for sectoral applications. It is a start. The numbers are small against what flows into English-language AI in a single week. Africa Business Communities

Pedro Dias, "The whole point was the mess" — methodical takedown of the AEO/GEO optimisation pitch: LLMs read text, not schema markup, so the structured-data industry is selling SEO-era tools for a system built specifically not to use them. Read

Carlos la Orden Tovar on the language industry — if translation companies adopt AI in ways that hollow out the talent pipeline, no one will be left to fix the slop when it breaks. The argument transfers to journalism without changes.

People Inc Q1 2026 — Google referral traffic down 63% over two years, digital revenue still up 8% year on year. The publishers thriving in 2026 are the ones who built a business that does not depend on Google sending traffic. Press Gazette

SlatorCon London — Friday 22 May 2026, Nobu Hotel London Portman Square. Thirty-plus speakers on language AI and language solutions. Ten days from now. https://slator.com/event/slatorcon-london-2026/

Tom Vaillant, a journalist and technologist, shipped twice last week — a long defence of journalists using AI tools, and a rebuilt Scoutpost (formerly coJournalist) with Claude and Codex desktop support, self-hostable via Docker. Quote: “Scoutpost is for local newsrooms, researchers, investigators, watchdogs, and anyone who needs to track the changes on the open web.“ One person, remarkable output. AI is vehicle, the human is the driver.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany.

Three projects you will hear about in this newsletter:

I cover all three with the same critical lens applied to competitors.

AI use: I use Claude (Anthropic) for research and to edit this newsletter, based on refined and specific prompts. My goal is to understand where the AI performs and where it fails. I learn something every week. Responsibility for stated facts, names, and links is entirely mine.

babylon-newsletter.com · Every Tuesday

7,000 languages. AI works for 20.