KITA AI is the inaugural launch within the UNU Frontier Agentic Ecosystem for the SDGs, pioneering SDG-aligned human-agent collaboration to foster inclusive and intelligent societies. Structurally engineered around the multilateral, evaluative, and inclusive frameworks of the United Nations, this tool is uniquely tailored to meet the decision making needs of the member states and the general public.
Jointly designed and developed by UNU Macau, AI Singapore, and the University of Pretoria, KITA AI is built on the "Kita" Principle: For All, With All. In Indonesian, Kita translates to an inclusive "we" that encompasses everyone in the conversation. KITA AI embeds this philosophy directly into its architecture, ensuring that conflicting viewpoints and vulnerable stakeholder voices are mapped, evaluated, and respected.
What KITA AI Does
- Simulates Multi-Stakeholder Councils: Users input a policy challenge or scenario. KITA AI automatically populates a council of autonomous AI agents representing relevant demographics, economic interests, and civic perspectives.
- Operationalizes the UN frameworks on responsible AI: It delivers a transparent, trustworthy, inclusive and contextualizable agentic platform to support decision-making.
- Runs Structured, Inspectable Debates: Rather than generating a flat summary, agents present formal proposals, cross-examine arguments, cite verifiable data sources, and voice objections based on their grounded values.
- Quantifies Real-World SDG Trade-offs: The system benchmarks arguments and impacts directly against measurable SDG targets, highlighting where policy gains in one area (e.g., economic growth) may create risks in another (e.g., emissions or social equity).
- Keeps Decisions in Human Hands: KITA AI is a deliberation-support system, not an automated decision-maker. It surfaces who is affected, why disagreement exists, and where trade-offs lieβleaving ultimate judgment and strategic choices with human policymakers.
Pulls local data & sources
Constructs demographics
Diverse demographic representation
Maps arguments & objections
Tests real-world impacts
Concurrent metric analysis
Alignment
Frameworks
Core Features of KITA AI
SDG-Aligned by Design
- Purposeful: Explicitly designed to meet the urgent challenges voiced by UN member states and the UN system, driving agentic AI research to support human decisions on what matters.
- Measurable on the SDGs: Policy proposals are evaluated through the lens of the Sustainable Development Goals, alongside a diversity of regional ethical and conceptual frameworks.
- For All, With All (The "Kita" Principle): Named after the Indonesian word Kita (the inclusive form of "we/us"), the tool ensures that vulnerable, minoritarian, or conflicting stakeholder voices are not erased by majoritarian "consensus".
- Explainable & Trustworthy: Sources are openly identified, the agentic discussion is entirely transparent, and human supervisors can modify inputs at any time. It also includes validation features against historical scenarios.
Agentic Architecture
- Sequential Multi-Agent Pipelines: KITA AI deploys multiple independent Large Language Model (LLM) agents that execute multi-step tasks across the Agent Builder, Debate Engine, and Report Engine.
- Structured Argumentation: Agents representing legitimate stakeholders autonomously map arguments, raise objections, and concede points based on formal logic frameworks.
- Active Environmental Interaction: The agents actively pull real-world context through identified sources via the Contextualizer module, and evaluate downstream impacts by executing external simulations via the Simulator module.
A Frontier AI Tool
- Pioneering "Non-Convergence" as Valuable Data: Standard AI debate systems tend toward a majoritarian consensus or vote. KITA AI is a frontier system because it treats disagreement as a first-class, explainable output, mapping complex human trade-offs instead of flattening them.
- Dual-Layer Prompt Grounding: It simultaneously grounds its AI agents in both demographic identity for the policy scenario and explicit ethical conceptual frameworks (e.g., specific philosophical traditions, SDGs, and regional frameworks)βa combination missing from previous multi-agent systems.
- Robust Evaluation Frameworks: It includes benchmarking techniques like "perturbation pipelines" to stress-test the AI's reasoning, mitigating training data leakage and ensuring the tool genuinely adapts to complex, novel policy changes.
Key Value Realization for UN Stakeholders
- SDG Integration: The platform evaluates policy outputs directly against the UN SDGs, providing automated indicators of how a policy affects different sustainability metrics.
- Scalable and LLM-Agnostic: Built using an independent plugin architecture, the tool can continuously adopt faster, cheaper, or more capable AI models without requiring a system redesign.
- Benchmarking & Back-Testing: Contains pipelines that restrict searches to past eras to see if the AI replicates historical debates, verifying that the system adapts dynamically to changing formulations.
What Makes KITA AI Different
| Dimension | Standard AI Systems | KITA AI |
|---|---|---|
| Handling of Disagreement | Flattens dissent to force a majoritarian consensus or vote. | Treats "Non-Convergence" as a first-class, explainable output to map trade-offs. |
| Agent Grounding | Generic persona prompting with high risk of hallucination. | Dual-Layer Grounding: Anchored in demographic identity + explicit ethical/regional frameworks. |
| Evaluation Metrics | General text summaries or basic sentiment analysis. | Directly measurable against the UN SDGs and localized contextual frameworks. |
| System Architecture | Static model dependency requiring full system overhauls. | Scalable & LLM-Agnostic: Plugin architecture that adopts new models easily. |
| Role of Human Judgment | Replaces human decision with a final AI recommendation. | Deliberation-Support: Human is always in the loop to supervise, modify, and decide. |
Q&As
How does KITA AI ensure meaningful human agency instead of replacing policymakers?
KITA AI is explicitly designated as a deliberation-support tool, not an automated decision procedure. It focuses on automatically surfacing who is affected and providing the rationales and quantitative indicators behind every stakeholder position, rather than forcing a single recommendation. By treating non-convergence as an explainable output, policymakers receive a complete map of conflicting viewpoints to interpret. Additionally, the Interactive Module Lab allows human users to actively guide the system by manually inputting data and running customized debates.
Whether and how can local political, legal, and social contexts be incorporated into the system?
Local contexts are integrated across multiple modules. Through the Agent Builder, users can directly define the specific demographic axes and value systems that anchor the stakeholders relevant to a local challenge. The Contextualizer Module can supplement policy scenarios with real-time local data by searching the internet, crawling specific databases, or consulting local debate forums. Finally, KITA AIβs open, modular architecture supports installable, region-specific plugins to tailor the system to regional value systems.
How does KITA AI help policymakers understand and navigate competing SDG priorities?
The Report Engine allows policymakers to run concurrent multi-lens evaluations on the same debate transcript, featuring dedicated value-system engines built around the SDGs. Because the Debate Engine uses formal dialogue models to explicitly structure arguments and state disagreements, contradictory arguments are clearly mapped out. This allows policymakers to see exactly where and why specific policy goals or SDGs conflict with one another in a given scenario.
If you are interested in using KITA AI, please contact at KITAAI@unu.edu