{"id":32353,"date":"2026-06-16T12:57:42","date_gmt":"2026-06-16T15:57:42","guid":{"rendered":"https:\/\/www.tedic.org\/?p=32353"},"modified":"2026-06-16T13:57:20","modified_gmt":"2026-06-16T16:57:20","slug":"towards-building-ai-governance-from-a-global-south-perspective","status":"publish","type":"post","link":"https:\/\/www.tedic.org\/en\/towards-building-ai-governance-from-a-global-south-perspective\/","title":{"rendered":"Towards Building AI Governance from a Global South Perspective"},"content":{"rendered":"\n


Original article in Spanish<\/a><\/strong><\/em><\/p>\n\n\n\n

Global governance of Artificial Intelligence (AI) faces a structural tension between existing regulatory frameworks that are insufficient and non-binding, and the speed of technological deployment, which reproduces and deepens power asymmetries between the Global North and the Global South.<\/p>\n\n\n\n

TEDIC, as part of the MAP.IA coalition, an initiative led by the Centre for Communication Governance (CCG)<\/a> at National Law University Delhi <\/a>and the Global Network Initiative (GNI)<\/a> (of which we are members), has contributed detailed recommendations to the United Nations consultation process for the first Global Dialogue on AI Governance (UNGD).<\/a> We are also members of the Global South Alliance (GSA)<\/a> and the Global Coalition for Tech Justice (GCTJ)<\/a>, spaces in which we have joined discussions on the principles and mechanisms that should guide AI governance at the global level. Civil society discussions aimed at ensuring meaningful participation in this space have focused on the development of collaborative documents by MAP.IA, GSA, Globethics <\/a>and GCTJ, as well as activities carried out since 2025.<\/a> In this context, TEDIC formally joined this agenda following the most recent meeting held in New Delhi during the AI Summit, with the objective of contributing Global South perspectives and priorities to the international debate on AI governance.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Four Structural Tensions in the AI Governance Debate<\/h2>\n\n\n\n

The MAP-AI initiative, which organized these events, sought to amplify underrepresented voices in AI governance processes<\/a> by bringing together more than 500 participants both in person and online, as well as approximately 120 speakers. The diagnosis and preliminary conclusions emerging from these meetings indicate that formal governance processes, including the series of AI Summits from Bletchley to Paris, have been\u2014and continue to be\u2014captured by corporate and governmental interests from the Global North, while civil society participation, particularly from the Global South, remains largely symbolic. For example, in the Indian experience, the concept of the \u201cGlobal South\u201d was frequently invoked, yet it was ultimately framed as a position rather than a place<\/a>\u2014an externally imposed category applied to a highly heterogeneous group of countries. In our documented experiences at the AI Action Summit 2025<\/a> in Paris and the AI Impact Summit 2026 in India<\/a>, civil society organizations were excluded from negotiation tables, and the official declarations issued during these meetings were not binding on states.<\/p>\n\n\n\n

In the Latin American context specifically, colonial patterns of extraction are reproduced through the large-scale extraction of data and the adoption of technologies primarily designed in the Global North, while Latin America remains excluded from the spaces where the rules of the game are defined. We do not view this as an exchange among equals,<\/strong> but rather as a relationship of extraction and technological dependency that is normalized under the neutral language of \u201cinnovation.<\/a>\u201d<\/p>\n\n\n\n

This asymmetry becomes even more pronounced when examining who bears the most tangible costs. Historically marginalized communities not only have less access to the benefits of AI, but also experience its most harmful failures. Facial recognition systems, for example, consistently exhibit higher error<\/a> rates when identifying Black women, transforming a technical problem into a form of automated discrimination with real consequences for freedom, security, and fundamental rights.<\/p>\n\n\n\n

Discussions also emphasized the importance of not separating AI governance from broader economic and environmental policy issues<\/a>. For instance, while data centers are the backbone of the AI economy, they also dismantle the myth of a climate-friendly digital transition.<\/p>\n\n\n\n

In this context, four structural tensions were identified during the dialogues held prior to the United Nations meeting scheduled for July of this year<\/a>:<\/p>\n\n\n\n

    \n
  1. The Legitimacy of Multistakeholder Processes:<\/strong> Civil society participation remains minimal and largely symbolic.<\/a> Furthermore, if participation is genuine, it will inevitably generate discomfort. Such discomfort should be understood as evidence that multistakeholder spaces are functioning as intended. Multistakeholder participation should not be merely consultative but rather a procedural requirement for legitimate governance.<\/a> Proposals include regional pre-dialogues with genuine agenda-setting power, community testimony mechanisms conducted in local languages, real-time public and multilingual documentation, and confidential participation channels for organizations operating in restricted civic spaces. Linguistic inclusion should be understood both as an access measure and as a governance principle.<\/li>\n\n\n\n
  2. The Unequal Distribution of Agenda-Setting Power<\/strong>: The concentration of technological power in a handful of frontier AI companies creates a democratic accountability problem:<\/a> societies that bear the consequences of AI deployment have limited capacity to shape the terms under which it advances.<\/li>\n\n\n\n
  3. The Gap Between Declared Principles and Operational Implementation:<\/strong> Voluntary commitments, declarations, and guidelines can be adopted or ignored by governments and companies alike. <\/a>There is currently no binding multilateral treaty, no mandatory enforcement mechanism, and no independent international institution with the authority to assess whether commitments are actually being fulfilled.<\/li>\n\n\n\n
  4. The Fundamental Question: What Is AI Being Built For, and For Whom?:<\/strong> One of the most provocative proposals emerging from these discussions was the call to move \u201cfrom generative AI toward regenerative AI,\u201d<\/a> placing social care, environmental sustainability, and context-specific models at the center, while treating skills as community assets rather than individual ones.<\/li>\n<\/ol>\n\n\n\n

    Governance Frameworks for the Global South<\/h2>\n\n\n\n

    There is currently a narrow interpretation of \u201csafety\u201d in AI, one that fails to adequately incorporate not only physical and structural harms but also the moral harms resulting from the rapid penetration of AI across different sectors. At the same time, a tension exists between rights-based frameworks and risk-based frameworks.<\/a> Rights-based approaches are grounded in constitutional and human rights law, whereas risk-based regulation operates through cost-benefit calculations of potential harms. In practice, risk-based frameworks often prioritize technical assessments while sidelining fundamental human rights concerns.<\/p>\n\n\n\n

    A key challenge is that regulatory frameworks frequently assume the existence of mature institutions and stable administrative capacity, conditions that rarely exist in many Global South contexts, including Paraguay. Furthermore, regulatory sandboxes are increasingly promoted as a mechanism for implementing AI within controlled experimental environments.<\/p>\n\n\n\n

    This methodology places innovation in a privileged position relative to the communities that become the subjects of experimentation. Real users\u2019 data and experiences are utilized without clear guarantees regarding how their rights and interests will be protected when failures occur within these testing environments. For example, during the public hearing on AI regulation in Paraguay\u2019s National Congress, the Ministry of Information and Communication Technologies (MITIC) recommended first testing sandbox methodologies<\/a>\u2014controlled and flexible environments for implementing innovation in fintech and other services\u2014before establishing specific AI regulations and legal frameworks.<\/p>\n\n\n\n

    However, this proposal is problematic. Researcher Luc\u00eda Camacho, in her study of AI regulatory sandboxes in Latin America and Europe,<\/a> documents how these initiatives are built upon the assumption that regulation constitutes a barrier to innovation and development. As a result, human rights protections are often weakened in the name of fostering innovation.<\/p>\n\n\n\n

    The cases examined reveal an urgent gap: the intersection between regulatory experimentation and the protection of human rights is virtually absent, receiving only generic references to safeguarding fundamental rights. Although the researcher does not reject sandboxes as a regulatory tool, she calls for greater scrutiny. She argues that active citizen participation should be incorporated into their design, that regulatory frameworks should be adapted to local needs and lessons learned, and that the European AI Act <\/a>should not be imported uncritically without adaptation to Latin American realities. The central question left open by this research is whether these testing environments are regulating AI in the public interest or merely facilitating its accelerated deployment under a veneer of regulatory legitimacy.<\/p>\n\n\n\n

    For these reasons, we believe that AI risk-based regulatory frameworks\u2014both internationally and nationally\u2014must adapt to local realities rather than requiring countries to adapt their realities to imported models such as the European Union\u2019s AI Act. At the same time, the transnational nature of AI requires governance frameworks built upon shared norms across jurisdictions, including cross-border incident reporting mechanisms, harmonized safety thresholds, and systems of shared governance.<\/p>\n\n\n\n

    Regarding safety, experts argue that it should function as a foundational architecture embedded throughout the entire AI lifecycle rather than as a layer added after deployment.<\/a> An important distinction is drawn between safety understood as the prevention of social harms and cybersecurity understood as resilience against adversarial manipulation. Confusing the two weakens governance frameworks. The most critical issue is that the current international framework rests almost entirely on voluntary commitments. There is no binding multilateral treaty, no mandatory enforcement mechanism, and no independent institution with the authority to assess whether commitments are actually being fulfilled.<\/p>\n\n\n\n

    Furthermore, human rights protections should be operationalized throughout the entire AI lifecycle. This includes imposing moratoria on systems that are incompatible with international human rights law<\/a>, such as invasive mass surveillance systems, non-consensual deepfakes, and predictive policing systems based on demographic characteristics. Gender is identified as a cross-cutting priority.<\/a> AI systems reproduce and intensify intersecting forms of discrimination that must be addressed across all AI applications rather than in isolated cases.<\/p>\n\n\n\n

    This discussion also highlights the importance of interoperability between regulatory frameworks. Interoperability should not result in weaker safeguards or in the exportation of diluted standards to the Global South through regulatory sandboxes or other forms of flexibility. Instead, interoperability should be anchored in human rights, transparency, data protection, and accountability, supported by safety metrics disaggregated by language, gender, and other relevant markers.<\/p>\n\n\n\n

    Concrete Demands on Cultural and Linguistic Diversity in AI Frameworks<\/h2>\n\n\n\n

    There is a fundamental structural problem that negatively affects cultural and linguistic assets within AI systems. Like much of the internet\u2019s infrastructure<\/a>, AI has been built <\/strong>predominantly in English and upon datasets that reflect primarily Northern and hegemonic cultures.<\/a><\/p>\n\n\n\n

    For Latin America, this has concrete consequences. Language models perform worse in Spanish and Portuguese and are virtually nonexistent for more than 400 Indigenous languages: including Quechua, Nahuatl, Aymara, Mayan languages, Mapudungun, and others, as well as mixed languages such as Guaran\u00ed. This is not merely a minor technical gap; it is a form of exclusion that affects access to public services, healthcare, education, and justice.<\/p>\n\n\n\n

    Therefore, communities whose languages and cultures are at stake must actively participate in the design of the systems that affect them rather than being consulted only after the fact. This requires recognition of the right to free, prior, and informed consent before their languages, cultural expressions, or data are used to train AI models.<\/p>\n\n\n\n

    Responsibility also extends to intermediaries. Automated content moderation systems systematically fail in underrepresented languages and cultural contexts, both through over-enforcement (removing legitimate content) and under-enforcement (failing to detect hate speech or disinformation).<\/a> Global discussions should require platforms to meet performance standards disaggregated by language and region, accompanied by public accountability mechanisms.<\/p>\n\n\n\n

    Finally, cultural diversity cannot be reduced to language alone. It encompasses worldviews, values, knowledge systems, and relationships with territory that may conflict with assumptions embedded within AI systems, such as individualism, linear concepts of time, and the separation between nature and culture. Global governance discussions should create space to debate which values are encoded into AI systems and who has the authority to decide those values.<\/p>\n\n\n\n

    While particular emphasis is placed on historically marginalized communities and minority languages, other critical issues must also be addressed. These include the precarious labor conditions of data workers and content moderators,<\/a> as well as the real environmental impacts associated with AI infrastructure<\/a>. These impacts include regulatory flexibilization for data center deployment, energy and water consumption, extraction of critical minerals, and electronic waste generation<\/a>\u2014all of which disproportionately affect the Global South.<\/p>\n\n\n\n

    How Should We Evaluate AI?<\/h2>\n\n\n\n

    First, we must move beyond general commitments and toward concrete funding mechanisms capable of closing the gaps in infrastructure, research capacity, and independent auditing within AI governance spaces.<\/a> Legal instruments and governance frameworks should require dedicated funding for multilingual evaluations, locally relevant benchmarks, and regional testing infrastructures led by institutions, academic organizations, and civil society groups from developing countries. These mechanisms would enable the assessment of AI systems before their deployment in high-impact areas such as healthcare, education, and justice.<\/p>\n\n\n\n

    Second, civil society organizations and academic institutions must be protected against retaliation. Research documenting harms caused by AI systems has often faced legal pressure, discrediting campaigns, disinformation, and, in some contexts, direct reprisals. Researchers and civil society actors must also have access to information from technology companies, including mandatory transparency regarding model documentation, error rates, training data, and system performance. In addition, governance mechanisms should ensure financial independence through public or multilateral funds dedicated to supporting independent participation.<\/p>\n\n\n\n

    Only after establishing these foundations can we meaningfully discuss effective evaluation frameworks centered on human rights. Human rights impact assessments should become the equivalent in AI governance to environmental impact assessments for physical infrastructure projects. International human rights law must serve as the primary framework for evaluating all AI governance proposals, including issues related to transparency, non-discrimination, privacy, and access to remedies. We therefore call for a shift from the mere invocation of principles toward their effective implementation through concrete accountability mechanisms for AI developers, deployers, and states.<\/a><\/p>\n\n\n\n

    We also recognize that AI infrastructure is physical. Networks of data centers, undersea cables, and supply chains rely on significant amounts of energy, water, land, and critical minerals, all of which must be included in governance and evaluation frameworks. This requires active coordination with institutions such as the United Nations Framework Convention on Climate Change (UNFCCC)<\/a>, the United Nations Environment Programme (UNEP)<\/a>, and the Commission on Science and Technology for Development (CSTD)<\/a>, while advancing toward binding requirements for the disclosure of the environmental footprint of AI systems. At the same time, cooperative, open-source, and locally governed models should be promoted as alternatives to the dominant extractive model.<\/p>\n\n\n\n

    Evaluation must also extend beyond one-time assessments. A pre-deployment audit alone is insufficient. High-risk systems used in areas such as credit scoring, employment, social benefits, migration, healthcare, and criminal justice should be subject to mandatory independent audits prior to deployment, analogous to financial audits or clinical trials in medicine. Digital rights organizations and civil society groups should operate continuous monitoring mechanisms, document cases of harm, systematize complaints, and generate evidence regarding patterns of failure that would otherwise remain invisible or fragmented. These efforts should explicitly include communities such as Indigenous peoples, migrants, people living in poverty, and women affected by technology-facilitated gender-based violence\u2014groups that often face significant barriers when attempting to reach government regulators through formal channels.<\/p>\n\n\n\n

    Fourth, technical standards relating to transparency, interoperability, model documentation<\/a>, and data governance are often developed in highly specialized forums with very limited participation from non-corporate actors. Academia can and should participate in these spaces by bringing a public-interest perspective to discussions about what should be measured, how it should be measured, and which thresholds should be considered acceptable. A concrete example is facial recognition technology. Aggregate accuracy metrics often conceal significant disparities in performance across gender, skin tone, and age. Requiring these metrics to be publicly reported in a disaggregated manner is not merely a technical decision but also a political and methodological one. Such decisions require the participation of voices representing those most affected by technological errors.<\/p>\n\n\n\n

    Finally, and no less importantly, civil society organizations must be able to use existing and emerging regulatory frameworks, including those resulting from international processes such as the Global Dialogue on AI Governance<\/a> and Third Ministerial and High-Level Summit on the Ethics of Artificial Intelligence in Latin America and the Caribbean <\/a>Mito demand accountability through strategic litigation. This approach has a multiplier effect. A single well-documented case can establish legal precedents capable of changing practices across an entire industry or government sector. For this to be possible, governance frameworks must include accessible remedy mechanisms\u2014not only regulatory fines, but also meaningful avenues through which affected individuals and communities can obtain redress.<\/p>\n\n\n\n

    AI governance cannot be reduced to a technocratic exercise. <\/strong>It requires transparency, accountability, and sustained responsible conduct over time, as well as genuine and consequential participation by the communities most affected by AI systems.<\/p>\n\n\n\n


    <\/p>\n","protected":false},"excerpt":{"rendered":"

    Original article in Spanish Global governance of Artificial Intelligence (AI) faces a structural tension between existing regulatory frameworks that are […]<\/p>\n","protected":false},"author":3,"featured_media":32352,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[1233,717],"tags":[2122,2121,2123,2117],"class_list":["post-32353","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog-en","category-freedom-of-expression","tag-ai-audit","tag-ai-governance","tag-regulatory-frameworks","tag-sandboxes"],"_links":{"self":[{"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/posts\/32353","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/comments?post=32353"}],"version-history":[{"count":7,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/posts\/32353\/revisions"}],"predecessor-version":[{"id":32377,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/posts\/32353\/revisions\/32377"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/media\/32352"}],"wp:attachment":[{"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/media?parent=32353"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/categories?post=32353"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tedic.org\/en\/wp-json\/wp\/v2\/tags?post=32353"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}