Culturally-Grounded Governance for Multilingual Language Models: Rights, Data Boundaries, and Accountable AI Design

Technical Summary: Culturally-Grounded Governance for Multilingual Language Models

Problem Statement
Multilingual Large Language Models (MLLMs) are increasingly deployed across diverse cultural, linguistic, and political contexts. However, existing governance frameworks and technical methodologies largely assume English-centric data, homogeneous user populations, and abstract notions of fairness. This creates systematic risks for low-resource languages and culturally marginalized communities. Current MLLMs often fail to align with local norms, rights, and expectations, leading to data asymmetries, biased model behavior, and a lack of accountability mechanisms for harms experienced by marginalized language speakers. Furthermore, traditional evaluation metrics (e.g., BLEU scores) focus narrowly on translation accuracy rather than the model's ability to facilitate intercultural understanding or serve as collaborative agents in Computer-Supported Cooperative Work (CSCW) and Human-Computer Interaction (HCI) contexts.

Methodology
This paper presents a systematic literature review aimed at comprehensively examining the trustworthiness of MLLMs within collaborative systems. The authors adopted a methodology inspired by prior systematic review works, searching the ACM Digital Library and Google Scholar for papers published within the last decade (with a specific filter for the past five years in the ACM Digital Library due to the recent surge in LLM research). The search utilized keywords intersecting MLLMs with collaborative work, such as "Multilingual Language Models," "MLLMs in Collaboration," and "Trustworthiness in AI," alongside terms like "Human-Centered AI" and "Data Quality."

The review corpus included peer-reviewed publications and relevant arXiv preprints to capture emerging trends. Articles were selected based on inclusion criteria requiring a focus on MLLMs within multilingual, multicultural collaboration contexts and addressing one or more of four critical attributes:

1. Data Quality
2. Dependability and Operability
3. Human-Centered Design
4. Human Oversight

Data extraction involved analyzing publication venues, years, methodologies, and findings related to these four attributes. The review synthesizes evidence on model behavior, data asymmetries, and sociotechnical harm to articulate a governance framework.

Key Contributions and Results
The paper does not propose new technical benchmarks or experimental models. Instead, it contributes a conceptual agenda that reframes multilingual AI governance as a sociocultural and rights-based problem. The review synthesizes existing literature to identify three interrelated governance challenges:

1. Cultural and Linguistic Inequities: Training data and evaluation practices disproportionately favor high-resource languages (e.g., English, Mandarin, Spanish), leading to underperformance and bias in low-resource languages.
2. Misalignment with Local Norms: Global deployment often fails to align with locally situated norms, values, and power structures, risking the reinforcement of stereotypes and cultural misunderstandings.
3. Limited Accountability: There is a lack of robust mechanisms to address harms experienced by marginalized language communities.

The review structures its findings around the four critical attributes of trustworthiness:

• Data Quality: The literature highlights that inadequate tokenization and biased training data lead to misinterpretations of underrepresented languages. While methods like synthetic data generation, model fusion, and matrix factorization (e.g., leveraging lexically overlapping tokens) show promise for low-resource scripts, comprehensive multilingual evaluation remains a gap.
• Dependability and Operability: MLLMs must maintain consistent performance across diverse inputs, particularly in high-stakes domains like healthcare. Research indicates that few-shot learning can aid cross-lingual tasks, but fairness principles (e.g., Rawlsian fairness) are necessary to prioritize underrepresented languages. Multimodal integration is also cited as a method to enhance adaptability.
• Human-Centered Design: Effective MLLMs must be accessible to non-technical users across diverse cultural backgrounds. The review emphasizes that failing to accommodate diverse linguistic needs risks reinforcing educational and economic disparities. Design must move beyond translation to include cultural context, as seen in frameworks like SERENGETI for African languages.
• Human Oversight: Ethical development requires continuous monitoring, regulatory compliance, and human-in-the-loop approaches. The review advocates for embedding ethical principles (beneficence, nonmaleficence, justice) and utilizing AI auditing frameworks to mitigate socio-political risks and ensure alignment with human values.

Significance and Claims
The paper argues that culturally grounded governance is essential to prevent multilingual language models from reproducing existing global inequalities under the guise of scale and neutrality. Its significance lies in shifting the discourse from purely technical innovations (such as alignment strategies or resource availability) to a focus on multicultural collaboration and sociocultural rights.

The authors claim that by reframing MLLM governance through a sociocultural lens, the research community can:

• Identify gaps between current research and practical applications in cross-cultural settings.
• Advocate for policy and legal frameworks that mandate diverse linguistic datasets and transparency.
• Promote interdisciplinary collaborations between AI researchers, linguists, social scientists, and cultural theorists.
• Develop new evaluation metrics that assess how well MLLMs facilitate trust, understanding, and equitable participation, rather than just translation accuracy.

Ultimately, the paper posits that for MLLMs to be truly trustworthy and effective in global collaboration, they must be designed not merely as translation tools, but as culturally aware collaborative agents that respect linguistic diversity and local power structures.