As today’s AI ecosystem is shaped by models, data, and compute concentrated in a few global institutions, discussions of “sovereign AI” are often reduced to a question of infrastructure. This workshop asks a deeper question: who holds the power to decide what stories AI tells, whose communities it represents, and whose values it is aligned to?
Sovereignty is often equated with owning one’s own compute and models, but compute and models are ultimately the means of sovereignty, not its ends. This workshop convenes the AI, HCI, and law & policy communities around a shared question: how can technical capacity, interaction design, and governance institutions together keep AI systems contestable and answerable to the communities they represent?
How do we ensure that increasingly autonomous and embodied AI systems remain safe, aligned, and under meaningful human control, especially when they are adapted or deployed outside the contexts in which they were built?
How can open models and local data governance move beyond the defensive goal of “reducing dependence” to become positive conditions for cultural and narrative agency? How can communities build on global models while retaining control over how their data, corpora, and knowledge are collected, interpreted, and used?
How can people effectively understand, oversee, and intervene in AI systems without losing their own agency?
How can people from different technical, cultural, and professional backgrounds help build AI that better serves diverse languages, communities, and public needs?
How can governments, researchers, industry, and the public meaningfully participate in governing AI and holding it accountable? And when “the local community” is itself plural, even internally contested, who has the legitimacy to speak on its behalf?
Confirmed researchers across human computation, social complexity, and computational social science.
President, Human Computation Institute · Visiting Professor, Cornell University, Meinig School of Biomedical Engineering
Postdoctoral Fellow, Center for the Dynamics of Social Complexity (DySoC), University of Tennessee
Postdoctoral Researcher, Computational Social Science Lab, Department of Politics and Public Administration, University of Konstanz
Keynote order and exact timing are still being finalized with our speakers.
Registration is handled through the main conference. Register for HCOMP 2026 and select this workshop.
Register for HCOMP 2026→What should a Sovereign AI know, and how should it respond?
Participants will define the public-interest questions a Sovereign AI must address, the evidence and local context its answers require, and the value conflicts those answers must keep visible.
Participants from AI, HCI, law, public policy, and other fields will examine each focal question through three lenses:
Which information and sources does a good answer require?
Which local contexts, stakeholders, and value judgements must the model recognise?
Which trade-offs or value conflicts should remain visible rather than being simplified into one answer?
Review questions grounded in public issues, news, and competing perspectives.
Choose 1-2 questions that are representative, contested, or dependent on local context.
Define what a good answer requires: evidence, stakeholders, values, and trade-offs.
Use Murmi to surface claims, consensus, and disagreement, then review and correct the map.
Turn the discussion into answer criteria and inputs for dataset design and evaluation.
Sovereign AI should reflect more than the priorities of model developers. Your expertise helps define what good, locally grounded answers look like.
This session adapts vTaiwan’s approach to open public-policy deliberation for the design and evaluation of Sovereign AI. vTaiwan provides the participatory model, while Murmi helps the room surface claims, map agreement and disagreement, and refine the results together.
vTaiwan brings government, experts, businesses, and citizens into a shared process that combines online participation with in-person deliberation. Rather than making policy decisions, it helps diverse participants clarify competing positions and develop recommendations that can inform policy.
Frame issues → Gather input → Deliberate → Develop recommendations
As generative AI increasingly decides how the world is told, media organizations are among the first to sense what is at stake: whose language, whose history, and whose memory gets written into the next generation's knowledge base will shape how a society is understood — and who gets to decide.
In Taiwan, an initiative led by the Taiwan Broadcasting System (TBS) and jointly undertaken by seven media organizations (Public Television Service, the Central News Agency, Radio Taiwan International, Chinese Television System, PTS Taigi, Hakka TV, and Taiwan Indigenous Television, along with the Indigenous Peoples Cultural Foundation)—the Taiwan Public Media NewsLLM project—is trying to answer that question.
They are drawing on three decades of exclusively licensed news, documentaries, and traditional opera, along with audio and text in Taiwan’s Indigenous languages, Taigi, and Hakka—a “Taiwanese memory” that ordinary corpora can neither buy nor crawl—and gathering it into an AI that is traceable, verifiable, and rich in historical depth. This is a path that does not compete on scale or compute, but leads instead with cultural depth and trustworthy sources, so that the AI remembers how Taiwan describes and understands itself.
For public media, this is a natural extension of public duty. When information is produced and circulated by AI, what is unsettled is the very foundation of trust in a democratic society. In choosing verifiable records as their bedrock, these organizations are not merely supplying data—they are helping to define what kind of response deserves to be trusted, safeguarding that trust and, with it, a healthy media ecosystem.
And this quickly points to a larger problem. Today’s AI is converging toward an “average.” In pursuing the greatest common denominator across languages and markets, it quietly flattens the languages, memories, and trust particular to each place. No matter how large, a general-purpose model has no commercial incentive to understand any single culture’s historical context, institutional distinctions, or shifts in meaning across generations. As a result, this generation’s experience of a society is going missing from the knowledge infrastructure of the next.
This has never been Taiwan’s predicament alone. Every minority language, every marginalized community, every piece of local memory the mainstream market declines to price faces the same risk. The heart of sovereign AI was never about who commands the most compute or the largest model, but about whether a community can decide for itself: in what language its AI speaks, whose history it remembers, and whose values it is aligned to. This effort, led by TBS, we offer as a demonstration—of a capability the global giants have no commercial reason to build, yet communities everywhere urgently need.
This workshop exists to ask, together: when global models inevitably shape every local life, how can a community retain the capacity to define its own narrative, values, and future? We believe that answer must be defined by those who care about it—not least the media that have long spoken for the public interest. We warmly invite everyone who shares this concern to join the conversation, and to walk this road with us.
This workshop invites researchers and practitioners to rethink sovereign AI through the lens of human-AI collaboration, interaction, and governance: when global models inevitably shape local lives, how can communities retain the capacity to define their own narratives, values, and collective futures?
We welcome contributions that advance, critique, or reimagine sovereign AI through three complementary paths including technical (models, data, safety, and evaluation), design (interaction, oversight, and intervention), and institutional (law, policy, and governance). We especially encourage work that crosses these boundaries. We particularly welcome submissions from researchers and practitioners in AI, HCI, law and public policy, government, journalism, civil society, open-source communities, and related fields.
Submission typesExtended abstracts, work-in-progress reports, or real-world case studies.
Length and format2-4 pages (excluding references) using the ACM conference format.
PresentationAccepted submissions will be presented in oral or poster sessions and will serve as the foundation for the workshop’s interactive discussions and collaborative activities.
ACM templatesSee General Submission Instructions for the official ACM templates and formatting policies.
Papers in our workshop are non-archival; accepted work may be submitted or published elsewhere without restriction.
Research Fellow, Institute of Information Science, Academia Sinica, Taiwan
lwku@iis.sinica.edu.twTechnical Program Manager, Taiwan Public Media NewsLLM Project, Academia Sinica, Taiwan
apokoios@as.edu.twPostdoctoral Researcher, ETH Zurich · Incoming Postdoctoral Fellow, MIT Laboratory for Information & Decision Systems
joshyang@mit.eduUniversity of California, Los Angeles
National Tsing Hua University
National Taiwan University
LL.M., UCLA School of Law