Across digital rights circles, the question is no longer only whether AI models work well. It is also who builds them, whose language they understand, and who controls the infrastructure behind them.
At RightsCon in Taiwan, civil society organizations from around the world were already dealing with the loss of one of the biggest funders of global digital rights work: the United States government. That funding shock is feeding a wider debate about whether countries and communities should keep relying so heavily on US AI models and US-based tech platforms.
Why Trust Is Fraying
The source of the concern is not limited to model performance. The Trump administration's rapid gutting of the US government, along with its push into what some prominent political scientists call “competitive authoritarianism,” is changing how many observers view American technology companies and their global responsibilities.
Those companies operate far beyond the United States. Yet people at RightsCon said they were already seeing shifts in how willing companies were to engage with and invest in communities with smaller user bases, especially non-English-speaking ones.
That matters because many of the people affected by platform decisions do not live in the main markets that shape product priorities. If companies reduce attention to local needs, the people most dependent on careful content moderation may have even less influence over how systems are built and maintained.
Yasmin Curzi, a Brazilian law professor who researches domestic tech policy, described the concern directly: “Since Trump’s second administration, we cannot count on [American social media platforms] to do even the bare minimum anymore.”
Content Moderation Shows The Weak Point
Social media is one of the clearest places where the risks are visible. Content moderation systems already use automation, and platforms are also experimenting with large language models to flag problematic posts.
But those systems are failing to detect gender-based violence in India, South Africa, and Brazil. The problem is not just that moderation is imperfect. The deeper risk is that flawed AI systems may become the tools used to judge even more content.
Marlena Wisniak, a human rights lawyer who focuses on AI governance at the European Center for Not-for-Profit Law, warned that greater reliance on LLMs could intensify the problem. “The LLMs are moderated poorly, and the poorly moderated LLMs are then also used to moderate other content,” she said. “It’s so circular, and the errors just keep repeating and amplifying.”
A central reason is training data. Many systems are trained primarily on material from the English-speaking world, and specifically American English. That makes them weaker at reading local language, local context, slang, slurs, mixed-language writing, and reclaimed language.
Even multilingual language models do not solve the issue on their own. One evaluation of ChatGPT’s response to health-care queries found much worse results in Chinese and Hindi, which are less represented in North American data sets, than in English and Spanish.
The Case For Smaller And Local Models
For many people at RightsCon, these failures support a different approach: AI systems shaped by the communities they are meant to serve. That could include small language models, chatbots, and data sets built for specific uses, languages, and cultural contexts.
The goal is not simply to translate English-first tools. It is to build systems that understand how people actually communicate in a given setting. That includes recognizing slang, interpreting words written across languages and alphabets, and distinguishing harmful language from terms that a targeted group has chosen to embrace.
Some of this work is already underway. The founder of Shhor AI hosted a panel at RightsCon and discussed a new content moderation API focused on Indian vernacular languages. Other efforts mentioned in the source include a Mozilla-facilitated volunteer-led project to collect training data in languages other than English, and Lelapa AI, which is building AI for African languages.
Small language models are also gaining attention because recent research and development has weakened the assumption that larger data sets always predict better performance. Aliya Bhatia, a visiting fellow at the Center for Democracy & Technology who researches automated content moderation, said that “smaller language models might be worthy competitors of multilingual language models in specific, low-resource languages.”
Tech Sovereignty Moves Up The Agenda
The debate over US AI models is also part of a broader push for tech sovereignty. AI competition was a major theme of the recent Paris AI Summit, and since then there has been a steady stream of announcements about “sovereign AI” initiatives.
In this context, sovereign AI means giving a country or organization full control over all aspects of AI development. The idea sits inside a larger desire for control over the digital systems that modern societies depend on.
Europe is a major focus of this shift. The European Union appointed its first commissioner for tech sovereignty, security, and democracy last November. It has also been working on plans for a “Euro Stack,” described as a form of digital public infrastructure.
The definition remains fluid, but the source says it could include the energy, water, chips, cloud services, software, data, and AI needed to support modern society and future innovation. Today, those are largely provided by US tech companies.
Europe’s efforts are partly modeled after “India Stack,” India’s digital infrastructure that includes the biometric identity system Aadhaar. Dutch lawmakers also passed several motions last week to untangle the country from US tech providers.
Andy Yen, CEO of the Switzerland-based digital privacy company Proton, connected this directly to Europe’s changing posture. Trump, he said, is “causing Europe to move faster … to come to the realization that Europe needs to regain its tech sovereignty.” He added that this is partly because of the leverage the president has over tech CEOs, and also “because tech is where the future economic growth of any country is.”
Government Control Is Not A Simple Fix
Moving away from US AI models does not automatically solve the hard questions. If governments take a larger role in AI development, decisions about which languages, communities, and views are represented may become politically charged.
Bhatia warned that government involvement needs limits. “I think there needs to be guardrails about what the role of the government here is. Where it gets tricky is if the government decides ‘These are the languages we want to advance’ or ‘These are the types of views we want represented in a data set,’” she said.
Her central point is that training data shapes model behavior. As she put it, “Fundamentally, the training data a model trains on is akin to the worldview it develops.”
That is why the current debate is bigger than procurement or platform preference. It is about whether AI systems can reflect the people they affect, especially when those people speak languages and live in contexts that dominant models often handle poorly.
It is still too early to know how much of the push for sovereign AI and local models will become durable infrastructure, and how much will remain hype. But the direction of the debate is clear: reliance on US AI models is being questioned because control, language, moderation, and trust are now inseparable parts of the same problem.