Open-Source Intelligence: The New Frontier of Climate Negotiations

Contents

Executive Summary

This panel discusses Negotiate COP, an open-source AI tool developed by the German Federal Government to democratize access to climate negotiation information and reduce information asymmetries between large and small country delegations at UN climate conferences. The speakers argue that while AI cannot replace relationship-building and political negotiation, it can accelerate information processing and preparation, particularly benefiting resource-constrained developing nations. The core message emphasizes digital public goods and co-development as alternatives to proprietary AI that concentrates power.

Key Takeaways

Information asymmetry in climate talks is solvable technology problem: Negotiate COP demonstrates that open-source AI can narrow the gap between large and small delegations by providing equitable, real-time access to negotiation baselines—freeing developing country negotiators from dependence on external consultants or informal networks.
Trust in AI tools for high-stakes diplomacy requires radical transparency: Privacy, openness, source clarity, and conservative quality thresholds are not nice-to-haves but prerequisites. Governance and verification matter as much as capability; co-development with affected nations is essential.
AI's highest value in negotiation is in preparation, not in replacing diplomacy: The most realistic five-year outcome is AI handling document synthesis and scenario rehearsal, enabling negotiators to spend more time on relationship-building and exploring zones of possible agreement—the irreducibly human core of negotiation.
The digital public goods model is a geopolitical choice: Investing in open, co-developed AI infrastructure is an alternative to proprietary systems that concentrate power. It requires sustained funding (EU's NextGen EU program supported this tool) and international commitment.
Feedback loops and ecosystem extension are critical: The tool's impact depends on user adoption, iterative improvement, and community extensions (scenario builders, integration with private sector tech, sectoral plugins). This is not a finished product but an opening for collaborative development.

Key Topics Covered

Information asymmetry in multilateral climate negotiations — structural inequalities in analytical capacity between wealthy and developing country delegations
Negotiate COP tool design and features — submission explorer, position comparison, portal chat functionality
Technical architecture — open-source models, ETL pipelines, RAG (Retrieval Augmented Generation), privacy-first design
Trust and governance concerns — ensuring AI tools don't exacerbate existing power imbalances
Digital public goods vs. proprietary AI — co-development as international cooperation model
AI's role in negotiation preparation — factual information processing vs. relationship-building
Future applications — prediction, scenario planning, agentic AI for preparation
Challenges in AI deployment — bias mitigation, usability in high-pressure environments, quality assurance
Ecosystem expansion — opportunities for plugins, scenario tools, and integration with other data sources

Key Points & Insights

Climate negotiations are structurally asymmetrical: Larger delegations distribute technical, legal, and strategic analysis across teams; smaller delegations from developing countries lack this capacity. The real inequality is not access to information but access to interpretive capacity — understanding what policy language means for their economic, technological, and financial interests.
Negotiate COP operationalizes three AI features:
- Submission Explorer: Extracts "key asks" (demands using words like "must," "should") and "fixed positions" (red lines using language like "must not," "shall not")
- Position Comparison: Side-by-side analysis with LLM-driven alignment indicators across dimensions (goals, instruments, timelines, financing)
- Portal Chat: RAG-based retrieval system allowing targeted queries across 600+ official UNFCC submissions from COP 29, COP 30
Open-source and privacy-by-design are trust mechanisms: Tool uses open-weight models, is hosted on German renewable-energy servers, stores no user data, and does not log queries. This design explicitly counters the consolidation of power if proprietary AI models become negotiation advantages.
LLM accuracy is acceptable, not perfect: Developers worked with actual climate negotiators to establish ground truth. The tool deliberately returns "N/A" rather than hallucinating, prioritizing trustworthiness over false precision. Even human negotiators don't achieve 100% accuracy; the tool supports verification and sense-checking.
AI augments rather than replaces human negotiation: AI handles information synthesis and factual preparation, freeing negotiators' time for interpersonal relationship-building, trust-building, and moral suasion — the core of international diplomacy. Deployed correctly, it enables human negotiation to deepen.
Geopolitical and bias risks are real but addressable: Tool mitigates political bias by only ingesting official UNFCC documents (not internet search results). However, developers acknowledge bias in LLMs is unsolved; mitigation requires triangulation with institutional memory, common sense, and human judgment.
Co-development and regional ownership are critical for legitimacy: A German-built tool risks suspicion; the path forward involves India, Global South nations, and other stakeholders building next versions collaboratively. Trust is earned through joint governance, verification, and adaptability to regional contexts.
Scenario and predictive AI are distinct but complementary: Portal chat is presently factual extraction; prediction (e.g., forecasting negotiation outcomes or climate impacts) requires decades of historical data and remains uncertain. However, AI agents can pre-simulate negotiations, allowing negotiators to rehearse and refine positions.
Digital public goods model counters market concentration: Unlike proprietary LLM APIs that concentrate power in tech companies, open digital public goods distribute AI infrastructure, reducing the unequal advantage accruing to rich nations that can afford expensive tools.
Future negotiations will likely integrate multi-agreement awareness: Combining UNFCC documents with inputs from Montreal Protocol, IMO, IKEO, and other environmental agreements could provide richer context and identify cross-cutting solutions — though this increases technical complexity and requires governance clarity.

Notable Quotes or Statements

By Dr. Arun Rabba Gosh (CEEW, COP 30 Special Envoy):

"The real asymmetry is in access to interpretive capacity. What does a particular phrase, what does a comma and a full stop and a semicolon really translate into..."

"Think of it this way: if you were a tennis player trying to get into Wimbledon, you're already entering with unequal resources. What tools like this can do is fill the gap of the backend you don't have, but also in real time. It's like your coach is whispering in your ear as you're playing."

"Triangulation: this tool doesn't substitute for common sense, interpersonal relationships, or institutional memory—but it makes all three easier by getting you baseline information faster and more accurately."

"If the tool works as intended, we can begin to see new bridges that could be created. Deployed poorly, it could end up simply solidifying the divides that already exist."

By Yanik Zasman (Data Lab, BMZ):

"The bet that the world currently bets on large models is precisely aiming at using LLMs for the purposes of nations. What we're doing here is the opposite—developing a public good to give tools to those who cannot afford proprietary models."

"Openness, trustworthiness, and privacy guided the development process. The N/A you see isn't the tool failing—it's working as intended because we don't want poor-quality information."

"In my experience, it's easier to solve one problem properly than to increase complexity by throwing all problems at AI and hoping for answers."

Speakers & Organizations Mentioned

Role	Name	Organization
Keynote Speaker	Dr. Arun Rabba Gosh	Council on Energy, Environment, and Water (CEEW); Special Envoy for COP 30 (Brazil)
Primary Presenter	Yanik Zasman	Data Lab, German Federal Ministry for Economic Cooperation & Development (BMZ)
Co-Developer	Dr. Illia Nicollet	(German federal administration)
Moderator	Gunda Rachuba	AI Impact Summit
Participant (closing question)	Jagal	Genjas (AI platform for international Indian market sales)
Participant	Aperna Jooshi	Climate education researcher

Institutional Partners:

German Federal Ministry for Foreign Affairs
German Federal Ministry for Environment
German Federal Ministry for Economic Cooperation & Development (BMZ)
GIZ (German International Cooperation Agency) — implementing partner
European Union (NextGen EU funding program)
Council on Energy, Environment, and Water (India)

Technical Concepts & Resources

Models & Frameworks

Open-source/Open-weight models: Tool avoids proprietary models (e.g., paid GPT APIs) to ensure equitable access
ETL (Extract, Transform, Load) pipeline: Scrapes UNFCC documents, extracts metadata, processes through LLM, loads into backend
RAG (Retrieval Augmented Generation): "Portal Chat" feature uses RAG for question-answering with source citation
LLM for entity/relation extraction: Identifies "key asks" (demand language) and "fixed positions" (constraint language) via action word detection
Alignment detection algorithm: Three-step process — (1) validity check, (2) contradiction analysis, (3) similarity scoring across analytical dimensions

Data Sources

Primary: 600+ official submissions from UNFCC portal (COP 29, COP 30, subsidiary bodies)
Not included (by design): WTO, IPCC, internet search results — to maintain source clarity and reduce bias
Ground truth: Human climate negotiators provided labeled examples for model evaluation

Quality Assurance & Governance

Accuracy metrics: Evaluated against human-negotiator answers using standard NLP metrics (accuracy, relevance)
Conservative thresholds: LLM returns "N/A" if confidence is not sufficiently high, avoiding hallucination
Privacy by design: No user login, no data storage, no query logging, no telemetry
Infrastructure: Hosted on German servers powered entirely by renewable energy
Licensing/Standards: Aligned with Digital Public Goods Alliance guidelines (formal registration pending)

Technical Stack (Partially Named)

Entity/Relation Extraction Model: F13 (open-source project developed for German federal administration)
LLM for Portal Chat: OSS-based model with 120 billion parameters
Framework: React frontend, backend database, RAG implementation

Agentic AI: Mentioned as future direction for AI agents to handle negotiation preparation independently
Scenario simulation: Using AI as a negotiation counterpart for pre-negotiation rehearsal
Climate Resilience Atlas: CEEW's example of long-horizon data needs (50+ years precipitation/heat stress data for hyper-local climate projections)
Multi-agreement analysis: Potential future integration with Montreal Protocol, IMO, IKEO, and other environmental negotiation texts

Open Questions & Research Gaps

Bias mitigation in LLMs: Acknowledged as unsolved; current approach is limiting input to official documents rather than web-scale data
Prediction vs. foresight: Uncertainty about whether LLMs can reliably forecast negotiation outcomes; distinction made between forecasting (difficult) and strategic foresight via scenario generation (feasible)
Data requirements for sectoral models (e.g., agricultural yield forecasting): Estimated minimum 20–30 years; CEEW working with 50+ years for climate resilience

Additional Context

Organizational Vision

The tool embodies a "whole of government" approach (BMZ, Foreign Ministry, Environment Ministry, GIZ) and reflects a strategic choice to build digital public goods as an alternative to market-driven AI concentration. The NextGen EU funding indicates EU-level commitment to equitable digital infrastructure for global climate action.

Governance Challenge

A key tension: How can a German-built tool earn trust from Global South delegations without appearing as a form of soft power or surveillance? The answer proposed: co-development, transparency, and equal access. Speakers explicitly invite India and other nations to co-develop next versions.

Future Outlook (5-Year Horizon)

Likely integration of predictive/agentic capabilities for automated negotiation preparation
Expansion to integrate multi-agreement and sectoral data (maritime, aviation, finance)
Regional ownership and co-development models (India mentioned explicitly as a priority partner)
Regulation of agentic AI systems to ensure proper governance in diplomatic contexts
Shift from tool-centric to ecosystem-centric approach (plugins, scenarios, sectoral extensions)