Babylon Newsletter, Issue #22

Note: I did not publish a newsletter last week, I was on sick leave.

Let me briefly say that I am approaching the future of technology with optimism. My motivation to get up and go to work is to work towards better, reliable and effective information systems. Better information for individuals and communities. Given that currently billions and billions of dollars are invested, it is surprising that there is no agreement yet on how such future information systems will work, present information or be sheltered against bad actors. This, in my view, is the interesting work ahead. This week's "Talk of the Week" is about a relevant recent report in this particular space.

16,000 to 10

In one week, AI systems made about 16,000 requests to Thomas Baekdal's site to answer people's questions. They sent back 10 clicks. Baekdal tracks this through Cloudflare. Claude's search bot alone made 4,307 requests and sent zero visitors, ChatGPT 3,831 requests and 7 clicks (Baekdal on LinkedIn). Baekdal says he doesn't want money from the AI companies. He wants readers. This is the exact same demand for any news offering.

1 October 2026 · The Information, via TechRepublic

Google runs an invitation-only "AI Contribution Pilot" with about 100 publishers. It pays when their content helps generate an AI answer (TechRepublic).

By the numbers: one early participant earns over $1 million a year. Newer ones get $50,000 to $60,000. Several small publishers got under $1,000 over several months.

The catch: publishers can't see what Google counts as a contribution. There is no statement from Google.

30 September 2026 · TechCrunch

Reddit will end RSS feeds on 13 November 2026 and public API access in March 2027. It says RSS has become a route for large-scale scraping (TechCrunch).

Who pays: Reddit sells its content to AI companies. That revenue line grew 24% to $43 million in the second quarter. Closing the open door protects what it sells.

Zoom out: Reddit is one of the backers of RSL (#5), the standard for telling AI systems on what terms they may use content. Closing RSS follows the same logic: access for AI only on stated terms.

1 October 2026 · Nieman Lab

Two 23-year-olds, Valentín Virasoro and Iván Schahovskoy Hoferek, built Lisa AI. Users pick topics. The app scans dozens of Argentine outlets and reads out a ten-minute audio briefing in an Argentine voice, with sources named (Nieman Lab). It is the number one news app in Argentina's Apple App Store.

Why care: An audio summary of the day's news, built by two people, went past every established news app in the country. This indicates that we might see many new forms and offerings in this space, enabled by AI and vibe coding. The founders report that 90% of the users finish the audio and 80% come back daily. The app is free and run by volunteers. The outlets it summarises are not paid.

In September, Shuwei Fang and David Caswell published a report on how information might work once AI sits between us and almost everything we read. The key question of this insightful PDF is where the value is created. It is based on four summits with about 140 people. Its own verdict is modest: a first map of uncertain terrain, not a blueprint. I suggest visiting the website and checking out the value map. It is a table of the categories that matter for creating information value.

The report is refreshing as it is not overclaiming, but careful and looking at the details. It is notable that while billions are invested into AI we can't define the best information system of the future yet. To succeed, some deep thinking is needed about what a long-lasting formula would be. Maybe it is personal intelligence, an assistant that knows what you need. Maybe it is the opposite, a kind of community glue that gives people a shared picture. The report names both, and it names the risk between them: "bubbles of one".

What needs to be considered: Not everyone using AI has good intentions. One group, for example, is a new class of AI marketing - they aim to maximise sales. In this field they already generate virtual personas, music bands, actors and AI short movies, with increasing quality. These new marketing contents will play to our wishes and dreams, more precisely and more aggressively than advertising ever could. They will show you cars, furniture, fashion, maybe before it is produced. Meta's assistant for 3.6 billion users is built around shopping (mentioned in newsletter #5). Google is turning the ad itself into a conversation about the details of a product or service, meaning that Google is no longer just a search engine, but might capture entire marketplaces (e.g. for a new car, a new computer, the best school for the kids). Amazon might follow in this direction.

The other group is more dangerous, it is political. The range of AI use cases here is from aggressive misinformation campaigns to autocratic forces. They will use AI to poison the water as much as possible for their purposes. They will use conspiracy claims based on what already worked in the past years to disconnect sizable groups from democratic reasoning and willingness to compromise. One recurring element is to paint the journalists and news media as a whole as a sinister system, steered by hidden interests. In Germany the word for it is "Lügenpresse", the lying press. In the United States Trump's term is "fake news". The White House now runs an official Media Offenders page under that label, with a "Hall of Shame" and a leaderboard of news outlets. I call it staining. Once the doubt is there, every real mistake a newsroom makes is read as proof.

And AI makes such staining easier. When a chatbot gives a wrong answer and names a newspaper as its source, the stain lands on the newspaper (mentioned in newsletter #15).

Against these threats I hold two things. First: People want information they can trust, and they are good at noticing when they can't. And there are groups and organisations who will work out how to inform better, deeper and in a way that earns trust. Maybe it will be harder to reach a majority of people in the future.

Second: A future personal information system will be weaker if it runs on AI alone. AI is powerful at finding, sorting and filtering and at combining existing information, based on what the user needs and wants. But without any human input such a system could quickly lose contact with reality. Human observation, the ability to connect bits of knowledge to something entirely new and innovative: this is a quality of the human approach to thinking. Even the most complex AI system does not know the world outside. It only sees descriptions, and those need to be refreshed day by day.

As a parting thought: Maybe the big prize is not about always knowing better, but about finding a formula which will "understand" human thinking, even to the point of arguing with people who deeply believe in conspiracy theories and gradually helping them to get in touch with reality again - like a therapy, without judging, but based on understanding. Here is my optimism again.

However this goes on, this future needs people who are motivated to go out, collect information with care and report back to their communities, whether local, national or global. In short: Good news reporters, investigators, analysts and maybe other types. We have to educate, enable and support them. Whether they are called journalists or content creators, and whether they publish alone or inside larger organisations, is not the point. The point is that humans find out and describe what is happening, with the intent to capture reality. Towards improved information systems.

ElevenLabs v4: over 90 languages, $22 billion valuation. The new voice models keep a speaker's identity across languages, the company says. An independent benchmark by Artificial Analysis ranks the newest model first for provider-supplied voices and pronunciation. A year ago the company was valued at $6.6 billion (Slator).

A US appeals court rejects fair use for AI training. The 3rd Circuit upheld Thomson Reuters' win against Ross Intelligence, a legal AI startup that trained on Westlaw content. It's the first federal appeals ruling of its kind. The limit: Ross retrieved existing text and didn't generate new text, so chatbot cases may go differently (Nieman Lab).

California wants AI performers in ads labelled. SB 1050, signed 16 September, requires a clear notice when an ad features a synthetic actor or voice. Translation and accessibility features are exempt. See the “Before you leave” story below in that context (Slator).

LocWorld56 · 19 to 21 October 2026 · Vancouver · locworld.com · The localisation industry's largest event.

Languages & The Media · 4 to 6 November 2026 · London · languages-media.com · Disclosure: I have a speaker slot.

Language Intelligence 2026 · 17 to 19 November 2026 · Vienna · details · Panel on why language models need structured knowledge, with APA, Infineon, Klartext AI and Intertranslations.

A Spanish-language radio station in Los Angeles has put an AI co-host on its morning show. The voice was made with ElevenLabs and is a clone of the boss's own voice. Listeners are told repeatedly that it is AI, and the station says more of them are tuning in. The show has moved from a weekend slot to the weekday commute and now runs in other US cities and in English too. This isn't a small experiment: the owner, Entravision, is one of the largest Spanish-language broadcasters in the US, with $447 million in revenue last year. The US actors' union, which also represents radio hosts, says synthetic hosts deceive audiences and erode trust (Los Angeles Times).

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

Three projects you will hear about in this newsletter:

AI use: Full disclosure, I do use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. My goal is to find out where AI is reliable and where the hallucinations come in. Before publishing I check all facts and links.

Error log: To track the issues of using AI I maintain an open Google Doc where I collect the small and big problems that a tool like Claude introduces to editorial work. The idea is to get a better understanding of where AI is good and where it is not. Read it here.

babylon-newsletter.com · 7,000 languages in the world, AI works for 20.