On the closing day of the AIFOD Geneva Summit 2026, held 12–14 August at the Assembly Hall of the United Nations Office at Geneva, Switzerland, under the theme “Small Takes the Lead”, Prof. Dr. Dip Nandi, Dean of the Faculty of Science and Technology at the American International University-Bangladesh (AIUB), served as Session Chair for Dialogue XIII - “Can Specialized AI Beat General AI?” The session, held on Friday, 14 August 2026 from 10:30–11:15 CEST, was one of the 16 high-level dialogues that engaged the summit’s 900+ leaders - ministers, founders, researchers, and civil-society voices - drawn from more than 142 nations across the globe.

Setting the Stage

Opening the dialogue, Prof. Dr. Nandi framed the session around a tension he described as central to the global AI trajectory: whether progress belongs to ever-larger general-purpose models, or to precise, domain-specific systems built for a particular purpose. He pointed to a set of figures that set the tone for the hour ahead - domain-specific, fine-tuned models achieving markedly higher accuracy at a fraction of the computational cost of frontier general models, high failure rates for general-purpose systems in mission-critical sectors such as law, healthcare, and public governance, and a steep “linguistic and sovereignty tax” imposed on non-English and agglutinative languages by generic, English-centric models.

“The real question is whether relying on centralized, generic foundation models creates dangerous dependencies, prohibitive costs, and regulatory failures in the places that need precise AI solutions the most,” prof. Dr. Nandi told the assembly, before welcoming the panel to the stage.

The Panel

Prof. Dr. Nandi chaired a five-member panel representing perspectives from across industry, diplomacy, and law. Vinod Sood of HSC (Hughes Systique), India, argued for hybrid adaptation - taking general foundation models and customizing them with domain expertise, local context, and human oversight, rather than building from scratch. Ezekiel Barclay Pajibo, Minister Counselor at the Permanent Mission of Liberia to the United Nations Office at Geneva, brought the imperative of rural health and education, pressing the case for accessible, affordable AI for underserved communities. Deniz Sherifoglu of 365 Derece, Türkiye, spoke to what she termed the “agglutinative tax” faced by non-English languages under generic large language models, and made the case for AI systems that are sovereign by design. Daniel Osuna of OZ Branding, Mexico, presented a high-stakes case study in specialized AI - using traceability systems to defend Mexico’s energy sector from criminal fuel networks. Helen Pino Vera of legal70, Spain, closed the introductions with an analysis of where broad, general-purpose models systematically fail against the precise legal, compliance, and governance boundaries her field demands.

Areas of Discussion and Convergence

Under the chairing of Prof. Dr. Nandi, the panel explored two important areas where different perspectives naturally emerged and were brought together through constructive discussion. The first concerned AI model architecture and national requirements. Vinod Sood emphasized the practical value of building upon existing general foundation models through fine-tuning and localization, while Deniz Sherifoglu and Daniel Osuna highlighted the importance of addressing tokenization limitations, adversarial vulnerabilities, and the specific security requirements associated with national sovereignty and physical infrastructure. Prof. Dr. Nandi helped connect these perspectives by recognizing that the choice need not be viewed as either general-purpose or purpose-built, but rather as a matter of applying the appropriate architecture, adaptation, and security controls according to the sensitivity and requirements of each use case.

The second area focused on balancing technological precision with accessibility and social inclusion. Ezekiel Barclay Pajibo emphasized the importance of affordable, low-bandwidth AI solutions for rural healthcare and basic education, while Helen Pino Vera brought attention to the need for domain-specific accuracy, strong data protection, clear liability frameworks, and rigorous safeguards against AI-generated errors. Through Prof. Dr. Nandi’s moderation, these perspectives were brought into a broader framework in which accessibility and precision were viewed as complementary priorities rather than competing objectives. The discussion consequently moved toward a shared understanding that AI solutions should be designed to maintain appropriate levels of safety, reliability, and accountability while remaining sufficiently affordable and accessible to underserved and rural communities.

Closing Remarks

Following an open-air dialogue with the audience, Prof. Dr. Nandi brought Dialogue XIII to a close with a synthesis of the session’s central argument.

“The dichotomy between General AI and Specialized AI is not merely a technical debate - it is an economic, political, and social choice,” he said. “While general foundation models offer broad, conversational horizons, specialized models provide the absolute domain expertise, regulatory compliance, and security required when real lives, legal frameworks, and critical infrastructure are on the line. We cannot build a fair global AI ecosystem if non-Western languages and resource-constrained nations are forced to pay an exorbitant ‘token tax’ or rent proprietary black-box systems. The winning formula lies in deploying specialized AI as the operational frontline, while utilizing general intelligence strictly as supportive, connective tissue.”

Prof. Dr. Dip Nandi's chairing of Dialogue XIII reflects AIUB’s continued engagement in shaping global conversations on AI sovereignty, standing alongside ministers, diplomats, and industry leaders from across the Global South in charting a path toward AI systems built for local languages, local infrastructure, and local needs.