NegotiateCOP: Democratizing Global Climate Negotiations through Open-Source AI
Detailed Summary
Speaker: Dr. Anuraba Gosch
- Complexity of climate negotiations – Gosch described COP negotiations as “one of the most complex multilateral decision‑making processes” because they intertwine climate, economics, trade, technology, finance, and geopolitics.
- Structural density – He highlighted the multitude of negotiating tracks (adaptation, mitigation, finance, technology, “just transition”, carbon markets, etc.) and the “asymmetry in analytical capacity” between large delegations (which can allocate specialized legal‑technical teams) and smaller, poorer delegations (which lack such resources).
- Interpretive asymmetry – The real problem, Gosch argued, is not just the availability of text but the ability to interpret each clause, punctuation, and nuance in the context of national economic and technological realities.
- Historical perspective – He recalled his early experience at a COP fifteen‑plus years ago, contrasting it with today’s accelerated pace and the emergence of AI as a possible tool to shrink the interpretive bandwidth required of negotiators.
- Beyond climate – He broadened the argument to other multilateral arenas (WTO, maritime, aviation, finance), underscoring that “text, time, and political alignment” are universal challenges.
- Invitation to the panel – Gosch concluded by asking how an open‑source AI tool such as NegotiateCOP could be operationalised, how it could reduce trust deficits, and how it could ensure equitable access.
2. Panel Introduction
Speaker: Dr. Iliya Nickelt (moderator)
- Thanked Gosch and introduced the two data‑science panelists (himself and Yannik Sassmann) who would now demonstrate NegotiateCOP.
- Stated the session’s aim: to illustrate how a jointly developed digital public good can address information asymmetry in COP negotiations.
3. Demonstration of NegotiateCOP
Speaker: Yannik Sassmann
3.1 Vision & Rationale
- Presented the tool as a public‑good AI commons that democratizes access to negotiation‑relevant data across delegations.
- Emphasised that the tool is open‑source, hosted on German renewable‑energy‑powered servers, and built through a whole‑of‑government collaboration (Foreign Office, Environment Ministry, BMZ, and GIZ).
3.2 Core Functionalities (Live Demo)
| Feature | What it does | How it is presented |
|---|---|---|
| Submissions Explorer | Pulls >600 official UNFCCC submissions (COP29, COP30, subsidiary bodies) and extracts key asks (demands) and fixed positions (non‑negotiable stances). | Demonstrated a searchable UI; highlighted colour‑coded extraction of verbs such as “must”, “shall”, “need”, “cannot”. |
| Position Comparison | Side‑by‑side matrix showing where two delegations align, diverge, or contradict each other across four analytic dimensions (goals, instruments, timelines, financing). | Showed a sample comparison (e.g., Germany vs. Brazil) and explained the three‑step LLM decision process (confidence check → contradiction detection → similarity scoring). |
| Portal Chat (RAG) | Retrieval‑augmented generation chat that answers natural‑language questions using the full corpus of submissions, with citations to original documents. | Ran a quick query (“What is the EU’s stance on carbon‑border adjustment?”) and displayed a sourced answer. |
| Open‑Source & Privacy | No user‑data logging, open code, and sustainability guarantees (100 % renewable‑energy hosting). | Stated privacy policy and renewable‑energy guarantee. |
3.3 Technical Architecture
- Scraper – Downloads PDFs from the UNFCCC portal.
- ETL Manager – Extract‑transform‑load pipeline that parses documents, applies the LLM‑based detection rules for key asks/fixed positions, and stores structured data in a backend database.
- LLM Models – Uses an open‑weight German government model (F13) for extraction and a 120‑billion‑parameter open‑source GPT for the RAG chat.
- User‑Facing Apps – Two front‑ends: a web portal (the “website”) and a lightweight “rack” interface (presumably a mobile‑friendly UI).
3.4 Deployment at COP30
- Pilot experience – Launched a prototype at COP30; observed strong interest from all delegations, particularly young negotiators comfortable with AI.
- Usability constraints – Highlighted the need for instant, low‑training‑overhead interfaces because negotiators are constantly moving between meetings.
- Limitations – AI currently assists factual extraction; it cannot replace relationship‑building and trust‑building that remain human‑centric.
4. Values Guiding Development
Speaker: Yannik Sassmann (follow‑up)
- Openness – Fully open‑source code, open‑weight models, public availability of the tool without login.
- Trustworthiness – Only UNFCCC‑sourced documents are used; the LLM conservatively returns “N/A” when confidence is low, preventing misinformation.
- Privacy – No collection or storage of user queries; ensures delegations’ strategic queries remain confidential.
- Sustainability – Hosted on German servers powered by renewable energy, aligning the tool’s environmental footprint with its climate‑focused mission.
5. Q & A – Key Themes
| Question / Theme | Main Points & Answers |
|---|---|
| Risk of misuse / power asymmetry | Open‑source approach prevents a single party from hoarding powerful LLMs; however, future “stronger” models could become new power levers. Emphasis on making such models as open as possible. |
| Current state of digitalisation in negotiations (Gosch) | Negotiators still face “information bombardment”; AI can help synthesize baseline data, freeing time for interpersonal engagement. |
| How AI can support relationship‑building | By delivering fast, reliable baseline information, negotiators can focus on “handshakes” and trust‑building; AI does not replace human interaction. |
| Inclusion of non‑UNFCCC data (WMO, IPCC) | Currently limited to UNFCCC submissions; adding other sources increases technical complexity and risk of bias. |
| Measuring LLM trustworthiness | Ground‑truth validation with climate negotiators; quantitative evaluation of accuracy, relevance, and confidence. Acknowledged that perfect accuracy is not required if the tool offers “good enough” insight. |
| Bias & political quotients in LLMs | No definitive solution yet; tool relies exclusively on official submissions to avoid external political bias. |
| Model choice (open‑source vs. proprietary) | Uses a German‑government open‑weight model (F13) for extraction and an open‑source 120‑B GPT for chat; avoids proprietary black‑box APIs. |
| Future scenario modelling | Not part of the current prototype, but participants see potential for plug‑in extensions that simulate “what‑if” outcomes (e.g., major economy exiting the deal). |
| Cost & funding | Prototype funded by the EU’s NextGen EU programme; operational costs are minimal; scaling cost would depend on usage volume. |
| Governance & regulation | Anticipated need for AI governance frameworks within climate negotiations; open‑source model promotes transparency. |
| Long‑term outlook (5‑year vision) | AI will likely handle the factual preparatory work, leaving negotiators to focus on diplomatic engagement; agents could even act as rehearsal partners for negotiating positions. |
6. Closing Remarks
- Panelists reiterated the importance of co‑development, trust, and verification as the foundation for any AI‑enabled public good.
- They invited further collaboration (particularly from non‑German delegations) and encouraged the audience to test NegotiateCOP via the QR code displayed at the end of the session.
Key Takeaways
- Information asymmetry is a core barrier in climate negotiations; small delegations lack the analytical bandwidth that larger delegations enjoy.
- NegotiateCOP is an open‑source, AI‑driven digital public good that extracts key asks and fixed positions from >600 UNFCCC submissions, enabling rapid, citation‑backed insight.
- The tool’s technical stack combines a scraper‑ETL pipeline, an open‑weight German government model for extraction, and a 120‑billion‑parameter open‑source GPT for chat‑based retrieval‑augmented generation.
- Deployment at COP30 showed strong interest, especially among younger negotiators, but highlighted the need for instant usability and the limitation that AI can only aid factual processing, not human relationship‑building.
- Core development values: openness (public code & models), trustworthiness (UNFCCC‑only data, conservative confidence thresholds), privacy (no query logging), and sustainability (renewable‑energy hosting).
- Risk mitigation: By keeping the model open‑source, the project aims to prevent a new “AI arms race” that could re‑introduce power imbalances.
- Evaluation of accuracy is performed with ground‑truth data supplied by experienced negotiators; perfect accuracy is not required as long as the tool delivers reliable “good‑enough” information.
- Future extensions could include scenario simulation (“what‑if” analyses), integration of additional data sources (e.g., IPCC, WMO), and AI agents that act as rehearsal partners for negotiators.
- Governance of AI in climate negotiations will become increasingly important; transparent, co‑developed tools like NegotiateCOP are positioned to support trustworthy, inclusive multilateral processes.
See Also: