Several weeks ago, at the RightsCon digital rights conference in Taiwan, I witnessed in real-time as civil society organizations around the world, including those from the US, struggled with the loss one of the largest funders of global digital right work: the United States Government.
In my dispatch, I described the Trump administration’s rapid, shocking gutting of the US Government (and its push towards what Some prominent political scientists have called ‘competitive autoritarianism’) also affects the policies and operations of American tech companies, many of which, ofcourse, have users beyond US borders. People at RightsCon reported that they had already seen changes in the willingness of these companies to engage and invest in communities with smaller user bases, especially non-English speaking communities.
As such, some European policymakers and business leaders are rethinking their reliance upon US-based technology and wondering if they can quickly create better, homegrown alternatives. This is especially true for AI.
Social media is one of the most obvious examples. Yasmin Curzi is a Brazilian law prof who studies domestic tech policy. She told me: “Since Trump’s second administration, [American social media platforms] cannot do the bare minimum any more.”
In places such as India, South Africa and Brazil, social media content moderation tools that use automation, as well as large language models, are failing to detect gendered violence. This problem will only get worse if platforms rely more on LLMs for moderation. Marlena Wisniak is a human rights attorney who focuses on AI Governance at the European Center for Not-for-Profit Law. She tells me that “the LLMs, which are poorly moderated, are also used to moderate content.” It’s circular and the errors keep repeating themselves.
Even the multilingual language models that are designed to process multiple languages simultaneously perform poorly when it comes to non-Western dialects. One evaluation found that ChatGPT’s responses to health-care questions were significantly worse in Chinese and Hindi than in English or Spanish, because these languages are not well represented in North American datasets.
This validates the calls of many at RightsCon for a more community-driven approach to AI, both in and outside of the social media context. These could include chatbots, small language models, and data sets that are tailored to specific languages and cultural contexts. These systems can be trained to recognize slang and slurs. They can also interpret words and phrases written in a mixture of languages, including alphabets, or identify “reclaimed language”which is a former slur that the targeted group decided to embrace. Language models and automated systems that are trained on Anglo-American English tend to miss or miscategorize all of these. The founder of Shhor AIhosted a panel discussion at RightsCon, and spoke about the new content moderation API that is focused on Indian vernacular language.
Many solutions similar to this have been in development for many years. We’ve covered several of them including a Mozilla-facilitated effort to collect data in languages other that English and promising startups such as Lelapa AI which is building AI in African languages. We even included small language modeling on our list of the top 10 breakthrough technologies for 2025 earlier this year.
Yet, this moment still feels a bit different. The second Trump administration that shapes the policies and actions of American tech firms is clearly a major factor. There are other factors at play.
Firstly, recent research and developments on language models have reached a point where the size of data sets is no longer a factor in performance, which means that more people are able to create them. Aliya Bhatia is a visiting fellow who studies automated content moderation at the Center for Democracy & Technology. She says that “smaller language model might be worthy competition for multilingual language models for specific, low resource languages.”
There’s also the global landscape. AI competition was the major theme of the Paris AI Summit that took place a week before RightsCon. Since then, a steady flow has been made regarding “sovereign AI”which aims to give a nation (or organization), full control over all aspects AI development.
AI sovereignty is only one part of a broader “tech autonomy” that has also gained steam. It’s a result of broader concerns about the security and privacy of data transferred to America. The European Union appointed its first commissioner for technology sovereignty, security, democracy, and last November. The EU has been working on plans to create a “Euro Stack” or “digital infrastructure.” This could include energy, water and chips, cloud services and software, data and AI, needed to support the modern society and future innovations. US tech companies are responsible for most of these services today. The European efforts are partly modeled on “India Stack”the digital infrastructure of that country, which includes the biometric identification system Aadhaar. Last week, Dutch lawmakers adopted several motions that would free the country from US technology providers.
Andy Yen is the CEO of Proton, a Swiss-based digital privacy firm. He told me this at RightsCon. He said that Trump is “causing Europe” to “move faster… to realize that Europe needs regain its tech sovereignty.” This is partly due to the leverage the president has with tech CEOs and also “because tech is the future economic growth for any country.” “I believe there should be guardrails on the role of government in this.” Bhatia says that it becomes tricky if the government decides, “These are the languages we wish to promote” or “These are the types views we wish to be represented in a dataset”. “Fundamentally the training data that a model uses is similar to the worldview that it develops.”
But it’s too early to tell what this all looks like and how much will be hype. No matter what happens, we’ll continue to monitor this space.
Originally published in The Algorithm – our weekly AI newsletter. Sign up to receive stories like this first
