Language Is Where AI Sovereignty Begins

I first came across Anna Mae Yu Lamentillo through a profile Joe Parton (Recruitment Manager, Executive Master's at Oxford) wrote for Oxford's Saïd Business School, and a small detail stayed with me: she had tried speaking Kinaray-a, her mother tongue, to an AI chatbot. It didn't respond to her prompt.

Anna Mae Yu grew up Karay-a in the Philippines, taught in English but mocked for her accent and a stutter that caught on words with the letter ‘s’. Her mother used to whisper proverbs to her in Kinaray-a when she got stuck. So when a chatbot met her own language with silence years later, she read it as a warning. That silence is, in a sense, the founding fact of her company, NightOwl AI. And what she built in response runs against almost everything the AI industry currently believes about itself, because instead of trying to scale up, Lamentillo decided to do the opposite.

NightOwl AI started as an app. A translation tool for endangered languages. Then she scrapped it. “There's a lot of apps already in the market,” she told me. “We want it to be primarily a language preservation tool.”

The company is now primarily a research operation, deliberately building smaller rather than bigger — fewer tokens, less reach, more focus on its mission of preserving endangered languages. “I feel that there's always a fascination with scaling higher,” she said. “But we want to build AI that is smaller. That is more personalized to the community that uses it.”

This personalization is the whole point of NightOwl's product. A large language model trained mostly on English learns English-speaking assumptions about what's normal, what's polite, what's safe, and carries them into every language it touches. I've felt a version of this myself. As a native Mandarin speaker, when I ask Claude or ChatGPT something in Mandarin, the reply often comes back subtly Westernised. The “I'm sorry to hear that” that opens a response isn't really how the sentiment gets expressed in Mandarin, and every time I see it, it grates a little. The tool is speaking my language without speaking from within it.

Lamentillo argues that you can't fix this by making the model bigger. A community's language holds its proverbs, its humour, its way of seeing things, and for a lot of indigenous languages, most of that has never been written down, let alone fed into a training set. The only way to build a system that actually speaks to a community is to build it with that community, on its own data, at a scale small enough that the people it represents still own and control what goes in.

It's an unusual thing to hear from someone running an AI company in 2026, when the entire industry is organised around the opposite instinct. But the more she explained it, the more it made sense — and the more it became clear that the small-versus-big question isn't really about product strategy but about who gets to build AI, and who's only consuming it.

Lamentillo came to this from inside the government. Before coming to the UK, she was Undersecretary at the Philippines' Department of Information and Communications Technology, where artificial intelligence sat directly under her portfolio. She served four administrations across different roles — infrastructure, public transportation, and then ICT. Somewhere in there, she did a master's at the London School of Economics on Cities, drawn by a fascination with the kind of dense, walkable, serendipitous urban life that most Philippine cities, built around cars, don't have. She's now at Oxford's Saïd Business School, studying for an MSc in Major Programme Management — the discipline of running large, complex projects, which sits close to the infrastructure work she did in government.

But the move to AI was personal. Lamentillo's mother-tongue, Karay-a, is among those at risk of disappearing.

She also wanted to stop being an expert just on paper. “I wanted to gain real technical knowledge,” she said. “I wanted to ensure that I'd be able to talk about artificial intelligence at a more technical level.” The company is incubated through LSE Generate, with support from Oxford's Saïd Business School, where she's now studying. When I asked whether starting a business has always been her plan, her answer was less about ambition than about autonomy: “For the first time I'd be able to actually direct my very life objectives. I don't want to answer to anyone anymore,” she laughs.

That technical grounding turned out to matter to NightOwl AI, because the closer she looked at how AI actually gets built, the more she saw a problem hiding in plain sight. It isn't a problem of ambition or funding. It's a problem of raw material. AI models are only as good as the data they're trained on. And for most of the world's languages, that data doesn't exist — because it was never written down in the first place.

“A lot of these languages are oral,” Lamentillo said. So NightOwl's actual first job, then, isn't building AI. It's documentation, putting spoken languages into written form so that one day there's something to build with. “The data is not yet available in the market. There's no competition either, because it's not there yet.”

The scale of the gap is hard to overstate.^[1] The Philippines alone has around 180 languages. Most written material — dissertations, laws, official documents — is in English. So even a well-resourced model has almost nothing to learn from in Tagalog, let alone in Karay-a or Bisaya. NightOwl's response has been to start building that record from scratch: in two years, the team has translated more than two million words across 22 languages, with volunteers in twenty countries.

Lamentillo's argument turns from a preservation project into something deeper here. Before a model ships, companies run it through fairness testing: checking its answers for bias, harmful content, and unsafe advice, usually by having human reviewers and automated tools flag problem responses in a given language. When AI companies test their models for fairness — checking for bias, for harm, for safety — they do it in English. Sometimes Mandarin. Then they deploy the same models in dozens of other languages and assume the safety carries over.

It doesn't. “You stop at fairness at one language,” she said, “and then a lot of AI models are branching out to more local languages without metrics in place to check whether they're even safe for human consumption.”

She pointed to Flores, the multilingual benchmark developed by Facebook, as one of the more ambitious efforts — and it caps out at 100 languages. There are roughly 7,000 languages in the world. “AI systems that are launched under these languages are not even examined for fairness,” she said. A model can earn what she called “the gold standard for fairness” by performing well in English and Mandarin, then branch into Tagalog or Bisaya without ever facing the same scrutiny. If a teenager is using one of these systems in a language it was never tested in, what conversations are they having, and what safeguards are actually in place?

There's an inversion buried in that gap. Fairness testing, as it currently works, gets done for the languages that already have the most data behind them: English, Mandarin, the handful that benchmarks like Flores can reach. Those are also the languages with the largest, best-resourced communities of users, the ones with the most capacity to notice a problem and push back. The languages that go untested are the ones spoken by smaller, more marginal communities — exactly the people least equipped to audit a system when it fails them.

The same gap applies to anyone relying on these tools in a low-resource language — for medical information, legal questions, financial decisions, schoolwork — trusting a system that was certified safe somewhere else, in a language they may not even speak. The fairness label travels. The actual fairness doesn't. And because the people most exposed to that gap are, almost by definition, the ones with the least power to audit it or complain about it, the failure is invisible to the companies shipping the product.

Behind all of these sits the largest claim, AI sovereignty, as Lamentillo calls it. The term borrows its logic from food or energy sovereignty — the idea that a country's wellbeing depends on whether it can produce something itself, or whether it's permanently dependent on someone else's supply. Applied to AI, sovereignty means owning the actual infrastructure underneath the technology: the base models that everything else gets built on top of, the computing power and energy required to train them, and a say in what those models consider safe, true, or acceptable. Without that, a country isn't really using AI so much as renting someone else's judgment.

Right now, AI infrastructure is concentrated in two countries. “Largely a lot of this AI infrastructure is only concentrated on the US and China,”^[2] she said. “The other countries have less than 10% to share.”^[3] That split, in her telling, divides the world into two groups. “You could see, in so far as AI systems are concerned, who are innovators and who are consumers. And I know the difference, because I come from a country with very little or minimal AI innovators.”

The Philippines, she said, has plenty of AI consumers — people fluent in ChatGPT, Manus, Google's tools. What it doesn't have is the infrastructure, the energy costs, or the base models to claim AI sovereignty over the technology. It's a large country of users, sitting outside the room where AI systems are being built.

This is also why Mandarin sits on the safe side of the fairness gap alongside English, and it isn't because the Western labs prioritise it. China has its own AI industry. DeepSeek, Alibaba, Zhipu and others are building serious models in Mandarin, which means the language has capacity behind it, tested and resourced, for the same reason English does: a country invested in building. The two well-served languages in the world are well-served because two countries have the infrastructure to serve them. Everyone else is downstream of a decision made somewhere else.

The dependency is easy to underestimate until you follow it through. A country of users doesn't just buy its AI from elsewhere the way it might import cars or software. It inherits the defaults baked into systems it had no hand in building — what those systems consider safe, polite, true, or harmful — with no leverage to influence any of it. It can't negotiate the safety standards, the pricing, or the cultural assumptions because it doesn't own the base models and can't afford to train alternatives. The fairness-label problem and the sovereignty problem turn out to be the same problem seen from two distances: up close, a model certified safe in a language it was never tested in; from further back, a whole population consuming tools it has no capacity to audit, correct, or build for itself.

Lamentillo is careful not to frame this as a fight. “I do not want to criticize all these AI companies,” she said. NightOwl, unusually, doesn't treat the big labs as competitors. The problem is too large for that. Her aim is closer to persuasion — to convince the major players to put language preservation at the centre of their fairness work, because a world where AI understands more languages and cultural context is, as she puts it, safer and more sustainable for everyone.

For now, NightOwl is pre-revenue, still in its research phase, working toward patents and building the written corpus that doesn't yet exist. Alongside the company, Lamentillo is researching how a sociolinguistic framework — treating linguistic fairness as a core principle instead of an afterthought — changes the way AI risk gets measured in the first place. It is, by design, a small company taking on a problem the size of the entire field.

That's the bet underneath the decision to get smaller. NightOwl isn't trying to serve all of them. It's trying to prove it can be done at all.

This article was written by Nessie Chu-Heng Lu, Strategic Marketing Lead, Founders & Funders.