Breaking Barriers: Multilingual AI as a Bridge to Democratic Access

Abstract

The panel examined why multilingual artificial intelligence is essential for democratic participation and how the lack of language diversity deepens the digital divide. Representatives from Switzerland, India, Europe, and Asia announced new Indo‑Swiss joint research calls, showcased the open‑source multilingual model Apertus, and highlighted concrete initiatives such as India’s Bhashini platform, the public‑private partnership Current AI, and Singapore’s Sea‑Lion model. The discussion emphasized the need for open data, shared compute, talent development, cultural‑sensitive design, and rigorous real‑world validation—especially in high‑stakes domains like healthcare.

Detailed Summary

Prof. Torsten Schwede opened the session by placing multilingual AI within a democratic framework. He asserted that AI can serve the public good only when it accommodates all languages and cultures, warning that linguistic exclusion is a persistent barrier to digital participation. He linked the discussion to a broader series of public‑interest AI events (Paris 2025, India AI Summit 2026, upcoming Geneva Summit 2027) and underlined Switzerland’s commitment to continuity, cooperation, and a truly global approach to AI governance.

Key points:

• Multilingual access is a democratic imperative, not merely a technical challenge.
• Switzerland will host the Geneva AI Summit 2027, aiming to cement multilingual AI as the foundation for inclusive public services.

2. Announcement of New Indo‑Swiss Joint Research Calls

Schwede announced three new Indo‑Swiss joint calls under the Indo‑Swiss Joint Research Programme (JRP):

Call	Focus	Partner Institutions
Geosciences	Natural hazards in mountain regions	Indian Ministry of Earth Sciences & Swiss FDFA
Social Sciences	Pressing societal questions	Indian Council of Social Science Research
One Health	Integrated health of humans, animals, environment	Indian Department of Biotechnology & Indian Council of Medical Research

He also introduced the Indo‑Swiss Research Framework Programme, a long‑term platform for thematic calls, and unveiled new Explore, Experiment, Expand grant streams aimed at:

• Enabling exploratory consortium work on “blue‑sky” topics.
• Supporting the scaling of established collaborations.
• Expanding mobility funding to ensure durable, cross‑border partnerships.

Schwede emphasized the ambition to hold frequent flagship events in both Switzerland and India to keep the community engaged.

3. Transition to the Panel – Moderator Introduction

Prof. Katharina Frey (moderator) introduced herself as the Executive Director of ICAEN and noted the breadth of the panel—representatives from Europe, Africa, Asia, and the United States. After a brief (and humor‑filled) photo‑session, she invited the first panelist to speak.

4. Bhashini (India) – Scaling Language Technology Across 22 + Languages

Dr. Amitabh Nag (CEO, Bhashini) described the Bhashini Initiative (Bhaasha Interface for India).

• Scope: 22 constitutionally recognised Indian languages (with plans to expand to 36 text languages and additional scripts for tribal languages).
• Modalities covered: Automatic Speech Recognition, bidirectional Text‑to‑Text Translation, Text‑to‑Speech, Optical Character Recognition, and a Digital Dictionary for vocabularies.
• Data strategy: Field teams (≈200 people) collected monolingual and bilingual corpora from local communities, building a “minimal digital dataset” that is continually expanded.
• Real‑world pilots:
  • Agriculture advisory voice interface for farmers in their mother tongue.
  • Gyan Bharatam – an interactive manuscript digitisation project.
• Challenges: Scarcity of digital data; iterative “training the model like a child” approach (starting with 100 books, now aiming for 1 000).

5. Current AI (France → India) – Public‑Private Partnership for Scale

Ayah Bdeir (CEO, Current AI) explained the organization’s vision:

• Mission: Build a globally collaborative, large‑scale public‑interest AI ecosystem that can counterbalance the dominance of a handful of big‑tech firms.
• Funding model: Initial commitments of ≈ $400M\ast \ast (Frenchgovernmentandotherpublicdonors)withanambitiontoreach\ast \ast$2.5 B. Partners include governments (France, India, Kenya, Morocco), foundations (MacArthur, Ford, McGovern), and private sector (Google DeepMind, Salesforce).
• Focus on language: The first flagship project, Multilingual Diversity, targets language and cultural preservation.
• Collaboration with Bhashini: A joint device (to be demonstrated at 15:30 in Room 10) aimed at bringing AI directly to end‑users in local languages.
• Cautionary note: Warned against “big‑tech style” data scraping that treats communities as raw data rather than partners, stressing the need for community‑centric, frugal approaches especially in low‑resource settings.

6. Apertus – Open‑Source Multilingual Foundation Model (Switzerland)

Prof. PD Dr. Alex Ilic (ETH AI Center) introduced Apertus, a 70‑billion‑parameter open‑source multilingual model jointly developed by ETH Zürich and EPFL.

• Design philosophy: Multilinguality built from the ground up (no English‑first pre‑training).
• Performance: Comparable to proprietary open‑weight models (e.g., Meta).
• Infrastructure: Trained on a Swiss national supercomputer (≈ 11 000 latest‑generation GPUs).
• Talent bottleneck: Only ~100 people worldwide possess the expertise to create such foundation models; academia must expand this talent pool.
• Open ecosystem: Apertus is released under an open licence to enable community‑driven extensions and benchmarks.
• Strategic goal: Incrementally increase language coverage (currently ~1 000 languages in the training set, 60 % English, 40 % non‑English) and evaluate cost‑effectiveness of adding new languages.

7. Nordic Perspective – Language as a Human Right

Prof. Petri Myllymäki (ELIS, Finland) highlighted the human‑rights framing adopted at the UN H‑Lab:

• Access to language and culture is a universally recognised human right.
• Small‑language communities (Nordic languages) must be protected from “one‑size‑fits‑all English AI”.
• Emphasised the importance of inclusive benchmarks that reflect diverse value frameworks rather than a single dominant culture.

8. Singapore’s Sea‑Lion Model – Regional Multilingual Initiative

Prof. Jon Wilson (NTU Singapore) presented Sea‑Lion, a large language model covering 13 Southeast Asian languages, including Tamil.

• Funding model: Nationally funded, but operates as public infrastructure with private‑sector partners across the region (e.g., platforms in Indonesia).
• Technical focus: Achieving strong performance with limited data (frugal AI).
• Cultural angle: Emphasised code‑switching and the fluid multilingual reality of Singaporean society.
• Sovereignty: Stressed that language technology should empower national and individual sovereignty, not just serve geopolitical superpowers.

9. Medical AI & Real‑World Validation – The LiGHT & MOVE Initiative

Prof. Annie Hartley (Director, LiGHT, EPFL) discussed high‑stakes applications of multilingual AI in healthcare.

• Case study: A model trained on minimal local data in Ethiopia gave a dangerous medical recommendation (“do not take insulin on Tuesday”), illustrating the perils of over‑reliance on generic data sources (e.g., the Bible).
• MOVE project (Massive, Open, Online Validation & Evaluation): A neutral, open‑science platform for collecting real‑world usage signals from clinicians and patients across languages.
• Goal: Continuously feed back field performance into model improvements, ensuring clinical safety and cultural relevance.
• Funding challenge: Validation work is costly and not traditionally valued; academia must advocate for dedicated funding streams.

10. Closing Remarks & Future Outlook

The panel reconvened to thank the audience, reaffirm the commitment to open, inclusive, multilingual AI, and preview the next steps:

• Continued development of Apertus and Sea‑Lion collaborations.
• Ongoing Indo‑Swiss joint calls and the Indo‑Swiss Research Framework Programme.
• Preparation for the Geneva AI Summit 2027, where multilingual AI will be a central theme.

Key Takeaways

• Multilingual AI is a democratic necessity; linguistic exclusion threatens equitable digital participation.
• Three Indo‑Swiss joint research calls (geosciences, social sciences, One Health) and a new Indo‑Swiss Research Framework Programme were launched to deepen bilateral scientific ties.
• Bhashini demonstrates a large‑scale, community‑driven approach to building speech, translation, OCR, and dictionary resources for 22 + Indian languages, with pilots already serving farmers and cultural heritage initiatives.
• Current AI operates as a global public‑private partnership, aiming to match big‑tech scale through pooled funding (~$2.5 B) and an open‑source multilingual focus, while cautioning against data‑scraping practices that marginalise communities.
• Apertus, the 70‑B parameter open multilingual model from ETH Zurich/EPFL, proves that open‑source foundations can rival proprietary models when multilinguality is built from the start.
• Talent scarcity (≈ 100 experts worldwide) is a critical bottleneck; academia must expand training pipelines and share compute resources.
• The Nordic human‑rights framing reinforces that language access is a universal right, not a luxury.
• Sea‑Lion showcases a regionally tailored, frugal multilingual model for Southeast Asia, emphasizing code‑switching and sovereign control over AI infrastructure.
• High‑stakes medical use cases expose the dangers of poorly validated multilingual models; the MOVE platform provides a neutral, open‑science mechanism for real‑world validation and iterative improvement.
• Open, collaborative governance—through bodies like ICAEN, SNSF, and upcoming Geneva Summit—will be essential to sustain multilingual AI as a public‑interest resource.

---

See Also:

• ai-for-democracy-reimagining-governance-in-the-age-of-intelligence
• evaluations-and-open-source-software-for-ai-for-social-good-at-scale
• flipping-the-script-how-the-global-majority-can-recode-the-ai-economy