David Caswell is the founder of StoryFlow Ltd. and a co-founder of the Signals at Scale venture studio. He has worked at the intersection of AI and journalism since 2011, started in Silicon Valley and is now based in the UK. He advises news organisations on AI and is one of the frequently cited voices on what is sometimes called structured journalism. Formerly David was an Executive Product Manager at BBC News Labs, focused on AI-based new product initiatives. He held roles at other media companies such as Tribune Publishing and Yahoo.

Mirko Lorenz, Babylon: A regional broadcaster comes to you with a real budget to use AI for content and asks where to begin.

Caswell: Money isn't the scarce resource. A willingness to engage is — specifically a willingness to engage with the agentic tools, on ambitious projects.

So I'd find or fund two or three people in the organisation who are obsessed with working with agents. They exist everywhere — the people who go home and run their setup until two in the morning because they love it. Give them a big chunk of time and all the compute they can use, and point them at your own workflows and processes.

The crucial part: isolate them from the existing organisation. Embedded inside it, there's too much cultural friction, too much politics, too many competing interests to do anything big and ambitious. And they don't need to be embedded — these super-empowered individuals are autonomous enough that they don't need the dev team, the designers, the lawyers. They can do a lot of it agentically themselves. Run a pilot with two or three highly motivated people, building agentic systems in isolation to automate the processes. Then look for the positive surprises — a head start on the downsides, early visibility of the upsides.

Babylon: Adrian Holovaty argued back in 2006 that journalism's core mistake was treating every story as an article instead of as structured data. Twenty years later, agentic AI is exactly the consumer that would benefit from structure. Are newsrooms finally doing it?

Caswell: I'd read it slightly differently. That was a pioneering article. But the first generation of structured journalism — roughly 2012 to 2015 — was built to enable the very things large language models now do on their own. Back then the models didn't exist. Translations, summaries, different experiences of an article — all of that was supposed to come from structure. Circa in the US did it with little components of news (Editor's note: a US news startup that broke stories into atomic "points"). Vox did it with cards. My own work was along those lines.

Now we have LLMs, so the rationale is different. The models work very well with plain text. They don't need it fully structured. This is sometimes called liquid content — the meaning of an article or a broadcast can flow from one form into another. But when you do that, you can't carefully track factuality, explainability, traceability, monetisation. You lose the ability to put something specific in an archive and correct it later if it turns out to be wrong.

So the new use case for structure is grounding data: journalistic information held in a form you can work with journalistically, and that AI models can build experiences on top of. There's a fast-growing community around this. In actual products, though, there aren't many yet — fewer than in that first generation a decade ago.

The one I'd point to is VG in Norway, the country's largest publication. They have a catch-up product that's hyper-personalised. It knows how long it's been since your last visit, and it gives you the backdrop on what's happened in a story since then. It doesn't do that from articles. It does it from structured events — a kind of timeline database. When you return, whether it's been three hours or three weeks, it pulls the events since your last visit and compiles them into a piece for you. A simpler version of the same idea is the AI summary at the top of a live blog, built from the individual posts so you don't have to scroll.

Babylon: In our innovation work we used to think about recalibrating offers across the day — the morning window where you orient yourself against the world. You're describing a step further: understanding what the reader already knows, and updating only that.

Caswell: Right. You don't tell them what they already know. You give them the update on the story.

Babylon: Is there a new calibration happening between AI companies and publishers — the AIs realising they can't keep serving copies of copies, that they need original, reality-based content? And does that hand publishers a payment argument?

Caswell: Yes — but there's an unpleasant side for publishers. A portion of journalism, and it varies a lot from publisher to publisher, is essentially processing tokens from original source documents. The government puts out a PDF, you write analysis around it, you bring in context. An LLM can do that too.

So the value holds for truly original reporting — originating information as a journalist, not just commenting on a PDF. It doesn't hold where the model companies can go straight to the primary sources: the social posts, the PDFs, the databases, the conversations. Essentially all digitally accessible primary sources will end up running through the LLMs.

There's also overlap. Legacy publishing grew up geographically — metro dailies, national papers — so a lot of content is duplicated, especially regional outlets running the same agency stories. In a world where reporting can actually be valued and traded, they only need to buy that AP story once. The value that lived in the overlap disappears. Real trading only works on genuinely original reporting — and the original reporting has to be there. Run the content analytics on most publishers and you'd find a significant share is overlapped, and another significant share is processing tokens from sources the LLMs can already reach.

Babylon: I sometimes say "journalism" is too broad a label — it stretches from gossip and paparazzi all the way to deep-sourced reporting. Progress needs the deep-sourced end.

Caswell: There's a gap in the vocabulary. "Journalism" doesn't capture it; "news" doesn't either, and both carry pre-AI baggage. The phrase I keep using is societal information — information important to how society functions. It's awkward, and I'm still looking for a better word.

Babylon: Structured content makes AI rendering easier. Does it also make multilingual rendering easier? Crossing language borders with AI is one of my central questions.

Caswell: The models got so good that translation is now decoupled from structure. In the first generation, the assumption was that you could translate small atomic units more easily than a full article. For the large languages, that no longer matters.

The real question is the middle tier and the long tail of smaller languages. The models aren't sufficient there yet, but they're improving in a fairly predictable way. Norwegian, Hungarian — languages that were hard a few years ago — are now basically fully functional. And it's not just speaker count; it's how well a language is represented in digitally available text. There are also new techniques, like teaching a model a language after it's trained rather than training it in. That's worked even for some minor Pakistani languages. For the vast majority, it's a matter of time. A few very small or unwritten languages — pidgin in West Africa, say — will be genuinely hard.

Babylon: We work with two approaches. One is Lesan AI, from the group started by Timnit Gebru. They propose a community-based, language by language approach for low-resource languages. The idea is to capture language based on how a community actually speaks it. They're sceptical of giant "1,600 languages at once" models.

Caswell: There's an often-overlooked piece here: it's not just languages, it's cultures. These models live in a big-city, highly-educated, knowledge-worker culture. Even in the UK, the nuance of northern English or rural culture barely shows up. You can see that playing out at scale across the world — we get access in a language, but a lot of regional or non-elite culture never makes it into the models.

Babylon: A concrete example — we worked with young journalists from the Qatar journalistic institute on Arabic, which runs from Modern Standard Arabic to very different country-by-country variants in the region. In French-influenced areas, people mix French vocabulary into Arabic, and that throws the model out of tune.

Caswell: Same thing in Los Angeles, where I lived for a long time — a lot of people speak a Spanish-English mix, and the model wants one or the other. Right at the heart of where this technology comes from, ways of communicating are being missed. Adaptability — the model learning nuance even after training — is going to matter a lot.

Babylon: Is something fundamental still missing to capture dialects over time?

Caswell: I think it's a matter of time, where there's demand. The bigger thing happening in language is automated translation combined with social media. X recently switched from a "translate this" button to auto-translating posts into the user's own language. That blends communities that were separated by language — suddenly Americans discovered this whole rich Japanese side of the platform that had been there all along. Once that starts, the forces are powerful. And the flip side: languages that arrive late, or that the models never reach an adequate level in, could be literally shut out of global conversations. Picture seven and a half billion people talking across languages, and five hundred million who simply aren't in it. Think how isolating that becomes.

Babylon: We debate about the need for news offerings in 24 European languages a lot. Technically we can do it now — but is there demand? Do I want the Spanish take on a story, or just the society-relevant facts from across Europe?

Caswell: I don't think demand is the right lens. That's how the profit-seeking tech companies will look at it. This is an access-to-infrastructure argument. It's like building roads, and there's a community over here that can't use them. It's an ethical and responsibility question for society, not a demand question — inclusion and access, especially if this is where the conversation is going to happen. For small languages it could go either way: it could erase or diminish a language, or it could stop one from eroding. The changes are so fundamental that the uncertainty is extreme.

Babylon: On compensation — there's liquid content and simple licensing, Mistral's CEO has floated a levy, and Le Monde's CEO is urging publishers to sign with OpenAI and Perplexity. What is your take on dealing with AI platforms?

Caswell: The situation isn't stable. The OpenAI deals tend to be short and simple. On the publisher side they're often motivated by more than money — the chance to work with OpenAI, to have joint projects. Most of the senior publishing leaders I talk to have been disappointed in that part. They expected a relationship; what they got was pay-for-API-access.

OpenAI's motivation may be as much about managing the societal and legal pushback — a fee for not suing us, in the New York Times lawsuit sense — and about deferring hard decisions to 2027, 2028, 2029. Most of these are roughly two-year deals; News Corp's five-year deal is an exception. When they come up for renewal, the terms will be very telling.

Longer term, I think there's a real market in selling grounding data and a flow of up-to-the-minute updates to the AI companies. It probably won't be articles — the article is a unit designed for human distribution. Once you get down to the raw semantic value of the information, the units change, which is where structure comes back in. The RSL approach, an agent payment protocol, micropayments — micropayments have failed repeatedly, but the agentic world is different because the burden on the consumer drops. You're not deciding whether something is worth ten cents; the agent handles that, and you pay twenty or thirty euros a month for a genuinely good experience across many sources. There are also marketplaces trying to aggregate supply from smaller publishers — they haven't done well, because the model companies aren't interested yet. It's a marketplace without demand. That could still ignite.

The levy, or robot tax, comes up well beyond media. The idea is to slow the replacement of humans by machines, to give society time to adapt. There are good arguments if the transformation gets too disruptive. But it's literally a tax on productivity, at a moment when Western productivity has already been weak for twenty years — arguably the root of a lot of the stagnation and the rise of populism. Taxing productivity right now could be a problem. Add the geostrategic friction — how Washington reacts to Europe taxing American products, and the competition with China to become a more AI-applied society. A levy could make sense in some circumstances. It's not an obvious solution.

Babylon: There are content models worth learning from — Wikipedia, for one. A long entry is unstructured-but-structured information you could also turn into a TikTok. But Wikipedia was underfunded, and we recognised its value late. Fair?

Caswell: Wikipedia is a pivotal example for the era we're entering. The big lesson I take from it is how much you can do with so little when the incentives are aligned — and not financial incentives, because most Wikipedians aren't paid. Something similar might work for a grounding-data layer for journalistic AI.

There are caveats. Wikipedia's editorial process struggles with news — the Wikimedia Foundation tried to build Wikinews and basically failed. It's good for historical and event-driven pages, less good for the daily back-and-forth of politics and culture. But the general model — non-financial motivation — could matter more, not less, in a future with more automation, some form of basic income, and more discretionary time. Something like that might sit underneath the grounding layer.

Babylon: What's one thing you changed your mind about this past year?

Caswell: For most of my time in AI and news — since around 2011 — I assumed the existing information providers would transform into this new world, the way they went from print to digital. That was painful and slow, but most of them made it. I expected the same shift into AI-native organisations.

Over the last year I've come to doubt that. I now think the legacy institutions probably will not transform. What they'll do — and are doing, quite successfully — is make their existing journalism, products, workflows and cultures as efficient and productive as possible with the tools. But they won't fundamentally restructure or reimagine what journalism can be in an AI world. They'll hit a limit, and the limit is the workflows and assumptions of the old, pre-digital universe — still wanting articles, still wanting the human-in-the-loop workflow, all the things I don't think will last. And that's unfortunate.

Babylon: David, thank you for your time.

Takeaways from the interview with David Caswell:

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany. I co-founded Datawrapper, a charting tool used in many newsrooms. Three projects you will hear about in this newsletter from time to time: 

babylon-newsletter.com · Weekly updates

7,000 languages. AI works for 20.