PUCP

The first Global Dialogue on AI Governance, is being framed not just as another summit, but as a multilateralism at its best moment. Given the current roadmap and the pressures from the global majority, i conside several key outcomes would signal a definitive success. First of all, i claim the neccesity to moving from soft law to a higher level. In that sense, the major criticism of past summits is that they produce vague ethical principles. we need to change that by clearing requirements for transparency, auditability, and human oversight in public administration (especially in high-stakes areas like the judiciary or healthcare). And I think we must negotiated declarations on labeling synthetic content and protecting electoral integrity before major global elections. Another point that belong to my concerns as lawyer is related to the protection of the epistemic integrity and human oversight. I consider there is a need to establish the non-delegable labor principle: Establishing a global consensus that certain core human functions, such as issuing a judicial sentence or a final medical diagnosis, cannot be legally outsourced to an AI without a meaningful human-in-the-loop audit trail.

• Interoperability of governance approaches
• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI

Interoperability of Governance Approaches From my point of view, AI knows no borders. A model trained on the Common Law of the US might be used by a judge in a Civil Law system (like Peru). If governance approaches aren't interoperable, we cannot regulate how these foreign logical structures influence local judicial decisions. So, i consider, we need a global "Common Language" for AI risk. If the UN Global Dialogue doesn't create interoperability, a judge in Lima might be using a tool that meets "Silicon Valley ethics" but violates Peruvian Constitutional standards. Related to safe secure and trustworthy, safety isn't just about preventing a chatbot from being rude; it's about Institutional Security, that means that when a judge inputs sensitive, non-public case data into a public AI to help draft a ruling, that data is often absorbed into the model's training set. This is a massive Security breach of judicial confidentiality. Nowadays there are a lot of reports concerning judges increasingly seeking convenience in AI models, therefore, the security of the judicial process is being compromised every day. We really need sandboxed AI environments for the public sector that are verified as safe from data exfiltration and trustworthy in their evidentiary output.

The most significant challenge is the informal automation of judicial reasoning. Judicial operators often resort to public, unvetted LLMs for summarizing cases or drafting sentences. This creates a "black box" where the decision-making power is subtly transferred from the legally mandated authority to private, foreign algorithms. This not only risks data exfiltration of sensitive judicial files but also compromises the epistemic integrity of the law, as foreign biases and "hallucinations" can infiltrate domestic jurisprudence. Without specific protocols, the judiciary risks a crisis of legitimacy where the public can no longer distinguish between a human verdict and a statistical prediction. Conversely, these gaps present an opportunity for this Dialogue to pioneer sovereign legal AI, we can move from passive consumption to active governance. We really need sandboxed AI environments for the public sector that are verified as safe from data exfiltration and trustworthy in their evidentiary output.

I am peruvian, so, i must express some concerns related to the role that AI Dialogue should play. I consider that, for countries in the Global South, the Dialogue is a platform for Resource Sovereignty as it facilitates practical cooperation on the global fund for ai capacity development. I mean, moving from talk to a structured mechanism for financing compute power and research. also, the Dialogue acts as a central hub to prevent the regulatory gridlock caused by fragmented national laws. Instead of forcing one single global law, it promotes functional equivalence where different regions (like the EU, LAT, ASEAN, or the African Union) can use different tools (sandboxes, algorithmic audits, or representative datasets) to achieve the same shared safety and accountability outcomes.

Epistemic sovereignty, by linking policy to a global Scientific Panel, it empowers smaller nations like Peru with independent evidence, allowing them to resist algorithmic capture by private tech giants and protect the integrity of their own human-led institutions.

STAKEHOLDERS are main in the dialogue but i consider, they need order. i mean, instead of only large plenary sessions, the structure should include Workshops. For e: a specific track on "AI in Public Administration" where judicial experts can test sandboxed models in a controlled environment before they are deployed.

Indigenous people: since these communities hold unique knowledge systems that are often scraped into AI training sets without consent, leading to digital colonization. Their perspective on collective data rights and the environmental impact of AI infrastructure on ancestral lands is almost entirely absent from mainstream regulatory drafts. very related to them are the digitally excluded, i mean people without internet access and are impacted by AI's climate policies, resource allocation, and border controls, yet they have no way to influence the algorithms governing their lives. I think they could be included by implementing sistemas of dialogue in wich there must be a seat for indigenous people to ensure that their lives and opinions are respected. Also, civil society groups from the countries concerned must be funded to act as intermediary voices, bringing the documented impacts of AI on disconnected populations directly to the plenaries.

Maybe a role play might be helpful in order to contribute people to assume a position and to evaluate the pros an cons of the different roles in society

To safeguard data security and cognitive sovereignty in judicial systems against invisible acces to AI, the most effective technical and regulatory solution is the implementation of a sovereign judicial sandbox. This model relies on replacing the use of public, commercial AI with localized instances of Large Language Models (LLMs) hosted on private government clouds or on-premise servers, ensuring that no sensitive procedural data ever feeds the training sets of third-party corporations