IGF 2024 DC-CIV & DC-NN From Internet Openness to AI Openness

The first part was a summary of what the Core Internet Values are and how they relate to the lessons that can be learned from them, particularly Internet openness, to foster AI openness. 

Luca Belli presented the relationship between Internet openness principles and AI governance, highlighting the interconnectedness of the two systems, with AI relying on Internet infrastructure and many Internet applications incorporating AI. Core principles like transparency and accountability are shared between both domains. In Internet governance, transparency ensures ISPs manage traffic without bias, while in AI, transparency is crucial for understanding how AI systems make decisions, particularly in critical areas such as loan approvals and law enforcement. However, many AI systems, particularly large language models (LLMs), are complex, often obscuring the decision-making process. Accountability, rooted in transparency, is vital for users, regulators, and society. Interoperability, a key principle of Internet openness, faces challenges in AI due to its concentration among a few large corporations, which undermines decentralization and non-discrimination, as this concentration disproportionately affects the Global South, where AI systems create limited access and contribute data to large corporations. Drawing from past Internet regulation debates, such as net neutrality, governance solutions for AI should focus on addressing the risks of power centralization, particularly for underserved regions, and ensure alignment with Internet Values like transparency, fairness, and openness.

Following-up, Olivier Crépin-Leblond gave a brief summary of the Core Internet Values and the technical principles behind them: it is a global, open, and accessible resource, interoperable across all devices, decentralized except for the DNS system, and user-centric, allowing freedom to choose applications. It remains robust and reliable despite growth and cyber threats. Safety has been added as a Core Value, ensuring cyber resilience through security measures. However, these Values face challenges, such as network neutrality issues. Despite this, the Internet's global and interoperable nature persists due to a balance among stakeholders. AI, seen as revolutionary, raises questions about regulation, and Crépin-Leblond suggested that lessons from the Internet’s governance—promoting openness, stability, and innovation—could guide AI regulation, ensuring a balance between safety, stability, and innovation.

Renata Mielli provided a historical perspective, contextualizing Brazil’s viewpoint on the connection between Core Internet Values—such as openness, permissionless innovation, interoperability, and collaboration—and the development of AI systems. Drawing from Brazil’s Internet Decalogue, she suggested that principles like standardization, human rights, inclusion, and democratic governance could guide AI regulation. While principles-based governance offers useful guidelines, Mielli emphasized the need for contextual governance, with countries adapting regulations to local realities while maintaining global cooperation. She also called for South-South collaboration to strengthen research networks and open innovation, noting Brazil's leadership opportunities in the G20 and BRICS to promote global AI governance frameworks. Ultimately, AI development should follow established Internet governance practices, focusing on ethical standards, transparency, and inclusive collaboration for a fair technological future.

Anita Gurumurthy questioned the true meaning of openness, arguing that it does not always equate to inclusivity or accessibility in the current context. She criticized how the early vision of an open Internet has been undermined by centralized data control and the dominance of transnational corporations operating "walled gardens." In the context of AI, she highlighted that systems labeled as "open" often lack transparency, reusability, and fail to democratize access. She cited OpenAI's GPT-4 as an example of this, with minimal transparency about its architecture and data. Gurumurthy called for restoring transparency, access to training data, and reusability, advocating for a collective rights framework inspired by environmental law principles. She emphasized three key AI rights: substantive equality, the right to dignity and freedom from exploitation, and meaningful societal participation in AI development. Her conclusion focused on a societal rights approach to AI governance, prioritizing ethics, collective sovereignty, and inclusion to ensure AI benefits all nations equitably.

Sandrine Elmi Hersi discussed the potential and concerns of generative AI, highlighting its growing role across sectors and its legal, societal, and technical implications. While policymakers, especially in the EU, focus on security and data protection (e.g., EU AI Act), the impact on Internet openness is underexplored. Generative AI tools, like chatbots, are replacing traditional search engines, potentially reducing user control and increasing issues such as bias and lack of transparency. Hersi expressed concern about generative AI’s effect on content diversity, as it could harm traditional media and digital commons. She pointed to the EU's Digital Markets Act and RCEP's initiatives aimed at applying Internet openness principles to AI. Elmi Hersi called for collective responsibility to ensure AI governance prioritizes openness, innovation, sustainability, and safety, urging proactive steps to maintain open Internet Values while fostering ethical AI development.

Vint Cerf discussed the fundamental differences between AI and the Internet, emphasizing that AI lacks the standardization that contributed to the Internet's success. AI systems are proprietary, with their architecture and training data protected as intellectual property, meaning open access does not equate to transparency in their workings. He highlighted the potential for future standardization, particularly for semantic interaction between AI systems, but warned of the risks of ambiguity in AI communication. Cerf stressed that regulation should focus on AI applications and the safety risks they pose to users, with providers held accountable for any harms caused. He raised concerns about the loss of detail in large language models (LLMs), which compress information into statistical models, potentially compromising accuracy. Cerf emphasized the importance of tracking the provenance of training data to ensure the reliability and transparency of AI-generated content. In conclusion, he called for a focus on safety, accountability, and transparency in AI governance, cautioning against treating AI governance the same as Internet governance.

Yik Chan Chin discussed the differences between AI and the Internet, noting that whilst AI relies on Internet infrastructure, its components—algorithms, data, and computing power—are fundamentally distinct. AI does not require standardization like the Internet, but interoperability, defined as the ability for systems to communicate smoothly, remains crucial. She highlighted the need for a balance between permissive innovation and the precautionary principle, as AI's complexity and unpredictability necessitate careful regulation. Core Internet governance Values such as transparency, accountability, and safety are also applicable to AI. Chin addressed global challenges like risk categorization, liability, and training data standards, calling for compatibility mechanisms to reconcile regional regulatory differences. She concluded by stressing the importance of strengthening global institutions for AI governance, ensuring clarity and fairness in key areas like risk evaluation and accountability.

Alejandro Pisanty emphasized the importance of preserving Core Internet Values like openness and interoperability in AI governance, while recognizing the challenge of applying these principles to AI. He argued that AI should not be viewed solely through the lens of generative models like LLMs but should encompass a broader range of applications such as molecular modeling and weather forecasting. Pisanty stressed that regulatory efforts should focus on human intent and misuse of technology rather than overregulating AI systems, which could hinder innovation in critical sectors. He advocated for digital agency over digital sovereignty, encouraging collaboration and capacity-building over restrictive regulations. Drawing from Internet governance, he suggested a multi-stakeholder approach to AI governance, with sector-specific regulations to address risks. Pisanty concluded by calling for a balanced approach that ensures accountability, supports innovation, and avoids overregulation.

Sandra Mahannan provided insights from a business perspective on AI and robotics, highlighting key issues such as bias in AI. She explained that AI often reflects the biases of its creators, leading to problematic outputs, especially in sensitive contexts like religion. Mahannan emphasized that the quality of AI responses is heavily influenced by the data fed into the models, and poor-quality data can worsen biases and inaccuracies. She advocated for regulation focusing on the development process, including areas like data quality, privacy, security, sharing guidelines, and ensuring interoperability across AI systems. Her conclusion stressed the importance of openness and regulation to address these issues and ensure ethical, accurate AI outputs.

Wanda Muñoz emphasized the central role of human rights in AI governance, asserting that human rights are not abstract but require concrete actions, policies, and accountability mechanisms. She argued for reframing AI governance discussions, moving away from terms like "mitigating risks" to recognize that AI harms often lead to systematic human rights violations, disproportionately affecting marginalized groups. Muñoz called for active measures to combat discrimination, stressing that non-discrimination in AI requires more than inaction and must include efforts to redress systemic issues in data, policies, and organizations. She advocated for specific regulatory actions tailored to protect vulnerable groups and ensure equity, aligning with Anita Gurumurthy on adopting a societal rights approach. Muñoz disagreed with Alejandro Pisanty’s view on "not demonizing technology," asserting that highlighting AI's risks and harms is essential for accountability. In conclusion, she stressed the importance of regulation, human rights, and collective societal frameworks to ensure AI promotes justice and equality.

Follow-up discussion Session

Desiree Miloshevic raised concerns about the impact of AI on Internet openness and the need for regulation at the AI layer.

Anita Gurumurthy and Vint Cerf responded, emphasizing the importance of transparency, accountability, and safety in AI governance. Vint Cerf asked in Chat room what would a desirable AI ecosystem look like?

Renata Mielli and Yik Chan Chin discussed the mutual impact of AI and the Internet, highlighting the need for cybersecurity and interoperability.

Sandra Mahannan reiterated the importance of regulating AI development to ensure quality and accountability.

Added post-session; Sandrine Elmi Hersi added that from a regulatory standpoint, it is imperative to ensure that AI models deployed in telecommunications networks adhere to the EU Open Internet Regulation, particularly in regard to AI-based traffic management. This is essential to uphold net neutrality obligations. Additionally, the proliferation of generative AI as a gateway to online content poses a potential threat to the principle of Internet openness. Given the possibility of generative AI becoming a pivotal intermediary for accessing content, it is vital to evaluate its impact on user empowerment and Internet innovation. Effective governance and collaboration will be crucial to sustaining an Open Internet, with a focus on transparency, user choice, and the prevention of discrimination. Detailed information can be found in reports by Arcep and BEREC.