| APrIGF 2025 Session Proposal Submission Form | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Part 1 - Lead Organizer | |||||||||||
| Contact Person | |||||||||||
| Ms. Torsha Sarkar | |||||||||||
| Organization / Affiliation (Please state "Individual" if appropriate) * | |||||||||||
| Centre for Communication Governance | |||||||||||
| Designation | |||||||||||
| Project Manager | |||||||||||
| Gender | |||||||||||
| Female | |||||||||||
| Economy of Residence | |||||||||||
| India | |||||||||||
| Stakeholder Group | |||||||||||
| Civil Society | |||||||||||
| Part 2 - Session Proposal | |||||||||||
| Session Title | |||||||||||
| Local Bots, Global Norms: Discussing content moderation and chatbots in the APAC region | |||||||||||
| Thematic Track of Your Session | |||||||||||
|
|||||||||||
| Description of Session Formats | |||||||||||
| Fireside Chat (30 minutes) | |||||||||||
| Where do you plan to organize your session? | |||||||||||
| Onsite at the venue (with online moderator for questions and comments from remote participants) | |||||||||||
| Specific Issues for Discussion | |||||||||||
| This Fireside Chat explores the issue of how content moderation practices of LLM-powered chatbots may be shaping content governance norms across the APAC region. Deployed predominantly by companies based in the Global North, these technologies have been flagged for importing Anglocentric norms and values into moderation frameworks, potentially suppressing legitimate content and offering users limited mechanisms for grievance redressal. In doing so, they pose a fundamental challenge to the future of multistakeholder and inclusive governance of these technologies: Who gets to define their moderation policies? And how — and if at all — are regional languages and cultural contexts accounted for? The session will first establish the context and rationale behind content moderation practices adopted by LLM-powered chatbots, focusing on what governance theories — if any — can justify deployment of such chatbots in online environments. Next, by centering experiences from the region and highlighting specific examples of content moderation decisions adopted by these chatbots, we will be discussing how the absence of multi-stakeholder dialogue has impacted their operation in diverse APAC jurisdictions. Finally, we will explore ideas about how academic and civil society stakeholders in the APAC region can meaningfully participate in shaping the norms embedded in these global systems. We note that the questions surrounding the deployment of these chatbot systems bear close resemblance to more traditional platform governance and content moderation policy dilemmas — particularly related to critiques about both norm-setting and bargaining powers being primarily concentrated in the hands of companies at the Global North. Accordingly, throughout the session, we will also be exploring whether insights derived from previous platform governance discourses can be educational in the current context. |
|||||||||||
| Describe the Relevance of Your Session to APrIGF | |||||||||||
| This session aligns closely with the APrIGF 2025 track on ‘Protecting User Rights’ and ‘Artificial Intelligence’, while also engaging with the broader theme of multistakeholderism. As AI systems like chatbots become central to user engagement with online spaces, their content moderation capabilities will effectively shape the boundaries of permissible speech online. These automated systems are rarely transparent, often unaccountable, and reflect the values of their developers — typically based in the Global North. As such, they pose critical questions for user rights, including freedom of expression, access to information, and the right to due process in content moderation decisions. This session will explore the challenges around content moderation practices adopted by chatbots chatbot-based moderation to examine what meaningful multi-stakeholder governance in the development of AI systems could look like. It will assess who gets to define moderation norms, whose voices are excluded, and what mechanisms exist (or should exist) to ensure that such systems are locally informed. The discussion will center APAC-specific challenges such as linguistic diversity and historical asymmetries in technology governance. Through this session, participants will be able to identify key governance gaps in how chatbots currently implement content moderation practices across the region and explore possible frameworks that can leverage multistakeholderism and include regional voices. This session will also be a part of CCG’s ongoing research in how LLMs are disrupting and reshaping content moderation norms in Global Majority contexts, and we will be using insights from the session to produce a short-form essay and disseminate it within domestic networks. |
|||||||||||
| Methodology / Agenda (Please add rows by clicking "+" on the right) | |||||||||||
|
|||||||||||
| Moderators & Speakers Info (Please complete where possible) - (Required) | |||||||||||
|
|||||||||||
| Please explain the rationale for choosing each of the above contributors to the session. | |||||||||||
| Aliya Bhatia is Senior Policy Analyst with the Centre for Democracy and Technology. She has written extensively on challenges of global content moderation processes in low-resource languages in the Global Majority context, and therefore, is well-suited to highlight specific gaps in representation from APAC voices in the development of chatbots in this region. Torsha Sarkar is a Project Manager with the Centre for Communication Governance, where she works at the intersection of law, technology, and human rights. She is keenly interested in issues of platform governance and content moderation norms, having written on these topics in the past. |
|||||||||||
| Please declare if you have any potential conflict of interest with the Program Committee 2025. | |||||||||||
| No | |||||||||||
| Are you or other session contributors planning to apply for the APrIGF Fellowship Program 2025? | |||||||||||
| Yes | |||||||||||
| Upon evaluation by the Program Committee, your session proposal may only be selected under the condition that you will accept the suggestion of merging with another proposal with similar topics. Please state your preference below: | |||||||||||
| Yes, I am willing to work with another session proposer on a suggested merger. | |||||||||||
| Brief Summary of Your Session | |||||||||||
| The fireside chat opened with the facilitator welcoming participants to a discussion on how large language models (LLMs) are transforming content moderation and raising new questions of speech, representation, and legitimacy. She framed the session around social media moderation, emphasizing that while LLMs promise accuracy, efficiency, and relief for human moderators, they also risk reinforcing structural inequalities when deployed in diverse linguistic contexts like South Asia. Aliya Bhatia from the Centre for Democracy & Technology introduced CDT’s work on technology governance and shared findings from a recent case study on Tamil content moderation. She explained that content moderation has long been an intractable challenge, and while automation helps scale operations, it cannot replace the human capacity to interpret context, intent, and nuance. LLMs, though more advanced than keyword filters, depend on the quality of their training data. Since most training data is English or machine-translated from English, the systems inherit Anglocentric perspectives and perform poorly in low-resource languages. Aliya defined low-resource languages as those with limited high-quality digital text available for model training. Most South Asian languages, including Tamil, fall into this category. As a result, LLMs often misinterpret colloquial or emotionally charged expressions, leading to excessive takedowns or missed harms. She referenced work by Monojit Choudhury showing that people use more slang, misspellings, and passion when communicating online in their native tongues — patterns that current models fail to understand. The speakers noted that moderation in low-resource contexts is hindered by both technical and institutional gaps. Companies rarely hire regional experts or build localized appeal processes, making it hard for users in the Global South to challenge unfair moderation. Aliya warned that increasing automation may further erode accountability if LLMs begin to control both content decisions and appeals. Drawing from CDT’s Tamil case study, Aliya identified three recurring issues: first, models trained on past data perform poorly during fast-moving crises; second, benchmarks test models mostly on English concepts; and third, mislabeled data and weak linguistic infrastructure produce systemic inaccuracies. The conversation closed with practical pathways forward: developing multilingual benchmarks; enabling South–South research collaboration; empowering third-party organizations like ROOST and the Trust & Safety Foundation to curate regional datasets; and supporting open-source innovations such as Vaani NLP and Tattle’s Uli, which embed local speech norms in their design. |
|||||||||||
| Substantive Summary of the Key Issues Raised and the Discussion | |||||||||||
| The fireside chat examined how large language models (LLMs) are reshaping the landscape of content moderation on social media, and how this shift intersects with issues of linguistic equity, democratic legitimacy, and human rights. The discussion explored the promise of machine learning tools in improving efficiency, while questioning whether these innovations adequately reflect the social and cultural complexity of global communication. The session highlighted that while platforms have long relied on automated systems such as keyword filters or hash-matching, LLMs present a new generation of tools capable of processing language in more contextual ways. Their appeal lies in speed, scalability, and the potential to reduce the psychological burden of human moderators. Yet their limitations are deeply structural. Because most LLMs are trained on vast troves of English-language data, they reflect a narrow understanding of speech, culture, and civility drawn from Anglo-centric contexts. When deployed in regions like South Asia, these systems often misread local idioms, satire, and political expression, leading to both over-censorship and neglect of genuine harm. Aliya introduced the notion of the “resourcedness gap,” a term used by AI researchers to describe the disparity between high-resource languages like English, Spanish, or Chinese, and low-resource languages like Tamil, Nepali, or Assamese that lack large volumes of digitized, high-quality text. This gap constrains the ability of LLMs to comprehend linguistic nuance or detect harmful content written in colloquial and emotionally expressive local speech. The conversation also stressed that the challenge is not just technical but institutional. Companies developing moderation systems rarely invest in regional linguistic expertise or well-staffed appeal mechanisms. This creates an uneven power dynamic where users in the Global South have little recourse to challenge erroneous moderation decisions. As more of these processes become automated, there is a risk that LLMs will control both moderation and remedy, leaving minimal room for human oversight or accountability. Drawing from a CDT case study on Tamil content moderation, Aliya noted that LLMs trained on English data perform poorly during fast-evolving events, such as protests or violence, where contextual understanding is crucial. Benchmarks used to test model performance are often based on English-language scenarios, and automated labeling tools further distort linguistic diversity by misidentifying non-English text. The speakers argued that the way forward lies in participatory, inclusive model development. South–South collaborations can help produce multilingual evaluation datasets, and partnerships with organizations such as ROOST or the Trust & Safety Foundation could support shared resource repositories. |
|||||||||||
| Conclusions and Suggestions of Way Forward | |||||||||||
| The session concluded with a clear recognition that the adoption of large language models (LLMs) in social media content moderation opens both opportunities and risks. While these systems promise quick responses, reduced moderator fatigue, and potentially more consistent enforcement, the discussion underscored that they also reinforce existing inequities in the governance of online speech. The central conclusion was that moderation, when informed primarily by Anglo-centric data and benchmarks, risks narrowing the global information sphere and undermining linguistic and cultural diversity. Current LLM-driven moderation reflects a structural imbalance between high-resource and low-resource languages. The overwhelming representation of English-language data skews the worldview encoded in these systems. In multilingual contexts such as South Asia, this can lead to biased decisions: over-removal of legitimate local expression, under-detection of harmful content, and inadequate representation of regional norms. The failure to build and maintain localized knowledge in moderation teams compounds these inaccuracies by removing cultural interpretation from decision-making. What emerges is not just a technical shortfall but a crisis of democratic legitimacy. Users who are most affected by errors—those operating in low-resource languages — often have the least access to appeal or remedy. Addressing these gaps requires reimagining both the design process and the governance framework around AI-assisted moderation. They highlighted that meaningful inclusion cannot end at policy consultation; it must extend to the technical pipeline itself. Development, testing, and evaluation of moderation tools should include local researchers, linguists, and civil society organizations that understand the communicative and political contexts in which these systems operate. Several strategic pathways for improvement were proposed. First, technological inclusion through the creation of shared multilingual benchmarks, corpora, and testing systems that reflect local speech practices. South–South research collaborations could help build high-quality datasets for underrepresented languages, expanding the domain expertise available for model training. Second, institutional inclusion through stronger regional Trust and Safety networks, such as ROOST or the Trust & Safety Foundation’s Global Majority Research Committee, to curate vetted linguistic resources accessible to multiple companies. Third, policy inclusion by embedding participatory governance models in platform design — ensuring affected communities have influence in shaping moderation standards before they are codified into AI systems. The conversation also endorsed open-source and community-based initiatives that embody participatory innovation. By supporting such initiatives, platforms can empower regional stakeholders and develop moderation infrastructures responsive to linguistic plurality. |
|||||||||||
| Number of Attendees (Please fill in numbers) | |||||||||||
|
|||||||||||
| Gender Balance in Moderators/Speakers (Please fill in numbers) | |||||||||||
|
|||||||||||
| How were gender perspectives, equality, inclusion or empowerment discussed? Please provide details and context. | |||||||||||
| N/A | |||||||||||
| Consent | |||||||||||
 I agree that my data can be submitted to forms.for.asia and processed by APrIGF organizers for the program selection of APrIGF 2025. |
|||||||||||