Babylon Newsletter, Issue #17

Almost every AI model learned from Wikipedia. Then the models started answering in Wikipedia's place, and now fewer people visit. So, picture the annual donation drive this winter. More people will say: I hardly used Wikipedia this year. Fewer visits mean less money and fewer volunteers. And the first thing that thins out is the small-language editions, the ones a 7,000-language world needs most.

That is roughly how many scraping requests Wikipedia fields a day. It can block only about 30 percent of the abusive ones, according to New York Times reporting citing Wikimedia.

Every large chatbot trained on Wikipedia's 65 million articles, but by now AI answers questions directly. As a result the traffic on Wikipedia is decreasing. Human pageviews of English Wikipedia fell about 8 percent over 2025. The demand from large tech companies generates some revenue, too - but not enough. The Wikipedia Enterprise arm, which charges Google, Meta, Amazon and Microsoft for bulk access, took 8.3 million dollars last year, more than double the year before. Pew Research puts the same 8 percent drop in longer context.

Why care? The encyclopedia the models learned from runs on human visits, through the donations and volunteers those visits bring. AI is draining the visits while paying back a fraction.

This is specifically relevant as all the small-language editions, the ones a 7,000-language world needs most, sit on the same engine. It is the killing-the-bees problem again, after the Internet Archive (newsletter #10) and Common Crawl (newsletter #14).

What is there to do? Raju Narisetti, who spent nine years on the Wikimedia board, puts the stakes on the roughly “250,000 volunteers editing 324 times a minute”, and says the real support is being a deliberate user, not just donating "during a banner campaign." One way or another, activity is needed here to save Wikipedia in the long run.

Source: Tiffany Hsu : “Wikipedia Is Battling for the Soul of the Internet”, The New York Times, July 5, 2026. Note: This link is a “gift article”, it should be accesible for all of you.

Update to Issue #16 · early July 2026 · Cloudflare

What's new: Cloudflare, which sits in front of a large share of the web, says that from September 15, 2026 it will block AI crawlers by default on ad-supported sites unless the owner allows them, and it is pushing AI firms to separate their search bots from their training and agent bots.

Why care? Babylon Newsletter #16 tracked publishers blocking crawlers one site at a time. Default-off at the network layer is a bigger lever: it changes the setting for many sites at once, and makes AI firms declare what their bots are for.

Last week I compared AI a tractor arriving on a farm. Such a new machine enhances key parts of the work to be done, but it is far from running the whole farm. This week a publisher did something with that logic worth watching.

Liquid content is the phrase of 2026, and almost nobody has shipped it. The idea is easy to say and hard to build: stop publishing finished articles, hold a story as parts a machine can pour into any format a reader wants. Video for the commute, audio for the kitchen, one verified line for the chatbot.

One iteration: In December 2025 the Washington Post launched the audio version, Your Personal Podcast. Its own pre-launch testing, reported by Semafor, failed between 68 and 84 percent of the AI scripts against the paper's quality standard, across three rounds. The errors were not cosmetic: invented and misattributed quotes, and commentary that made a source's words look like the Post's position. A second AI was already in the pipeline to check accuracy. The product team recommended launching anyway, as a "beta" to iterate on. Motto: “This is how software gets built”.
The editors did not like it at all. It still launched.

Another iteration: Then there is Schibsted, the Nordic publisher behind VG and Aftenposten. In March it put a piece of the machine on the table. Videofy, its tool for turning an article into a short news video, is now open source on GitHub. It fetches an article, drafts a script, matches visuals, generates a voiceover, and renders the video before an editor approves it.

Read the code and the real lesson appears. What Schibsted released is labelled "minimal." It runs on a laptop and leaves out, in its own words, most of the real system. The example, the open source code - this is still valuable. But to think someone else offers a production ready system, entirely for free is an illusion. The part it kept is the part that counts: the clean pipe that feeds the tool good content. So the free gift is the video renderer, which the market is now flooded with. The private asset is the reporting and the plumbing into it.

Schibsted publishes no error rate for Videofy, so I cannot say whether its editor at the gate catches what the Post's second AI missed. And liquid here still means format, not language: the same story becomes a video and an audio clip, not yet the same story in Norwegian, Sámi and English at once.

Doctor, Doctor, tell me the truth: Rhode Island now requires healthcare providers who use AI to write up patient visits to tell the patient and to check the AI's notes for accuracy, part of a package signed this week that also restricts therapy chatbots. A small state law, but it puts AI transcription (STT, speech-to-text) inside a disclosure-and-accountability rule.

Revealing the new normal: In Germany, the AI-writing detector Pangram has flagged ministerial speeches and opinion pieces as machine-written, including a state premier's Holocaust memorial address rated 100 percent AI. The practitioner point is the reliability question, not the scandal: these detectors used by Pangram might also flag the polished, rule-following prose taught in journalism school. German magazine/online Site DER SPIEGEL has published its position that it does not let AI write or rewrite its articles (in German).

AI guide for investigative journalists: A new Global Investigative Journalism Network guide splits AI reporting into four stages, data, compute, models and applications. Created by Lighthouse Reports' Gabriel Geiger with Karen Hao and Lam Thuy Vo. For this beat the live stages are data (whose text trained the model) and models (which languages were left out).

There is a new term that captures nicely how to manage which engine you use, and at what price: smart model routing. It means that between you and the LLMs there is another layer, looking to give you (A) the best engine and (B) the lowest possible price per token. Prices per token run 10 to 20 times higher for a top model than a cheap one, so instead of committing to one, a router reads each request and sends it to the model that fits. “The Pragmatic Engineer” counted a dozen tools already. In localisation the same gap is about 6x between the best and worst models on real translation work.

On plain X, the platform I work on, we run a working variation of this for language engines. Behind one workflow sit fifteen transcription services and eight translation engines, from the big names to specialists like OpenTrad for Galician, Lesan for Amharic, or Portugal's new AMALIA for European Portuguese. We used to call this being "engine agnostic" and leave the choice to the user. That was the wrong ask. People choose the engine whose name they recognise, and keep choosing it long after a benchmark says another one is better for their language pair. So now the choice is simply "best". Our data picks the strongest engine for each pair, a tooltip shows which and why, and a person signs off on the result. It is the same move you make in a chat app when you stop reading model names and just want the good answer.

The key learning and current strategy: Do not commit to just one engine.

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 visualization 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 · 7,000 languages in the world, AI works for 20.