NegotiateAI: Smart Negotiations by leveraging AI for UN Negotiations on the international legally binding instrument on plastic pollution

In the dynamic world of international environmental diplomacy, the complexity of the subject, variety of diverging positions as well as volume of documentation challenges fast progress. Smaller delegations seem disadvantaged compared to the larger delegations by time-consuming prepations requirements and negotiation sessions considering extensive legal texts as well as scientific background papers. This is particularly true for low income countries’ delegations.

The GIZ Data Lab together with two projects, commissioned by the German government and implemented by GIZ,  “Support Project on Marine Litter Prevention” and “Go Circular” and the data service provider &effect developed NegotiateAI, an advanced AI-driven app designed to support policy advisors, negotiators and observers at the United Nations negotiations to elaborate an international legally binding instrument on plastic pollution, including in the marine environment. This innovative tool aims to provide transparency and increase efficiency, as well as effectiveness in the negotiation process by enhancing the accessibility and analysis of critical documents and country positions.

The United Nations have embarked on a monumental task: Negotiating a legally binding agreement to tackle the global plastic pollution crisis. The goal is to establish a robust framework that ends plastic pollution and promotes sustainable production and consumption of plastic worldwide. Representatives from around 180 countries are involved in this extensive process. The resolution to elaborate such a legally binding instrument was adopted in March 2022 at the UN Environment Assembly (UNEA-5.2) in Nairobi. In April 2024, the fourth of five  negotiation rounds was completed, bringing the finalization of a comprehensive draft closer to the fifth and final session in November 2024 in Busan, Republic of Korea.

The urgent need for a treaty arises from the estimated 4.8 to 12.7 million tons of plastic entering oceans yearly and predictions of global plastic waste production nearly tripling by 2060 without intervention, as an OECD study predicts. Plastic production, heavily reliant on fossil fuels and toxic additives, threatens both the environment and human health, with microplastics found in the human body, even in unborn babies.

To address these challenges, we introduced NegotiateAI for the fourth negotiation round. This user-friendly online application enables users to query document contents of the treaty database in natural language via a free text box. By leveraging advanced filters such as country, negotiation round, or thematic section of the agreement, users can perform targeted searches, enhancing both the efficiency and effectiveness of document review and analysis.

How It Works:

• Technology Selection: We chose Retrieval-Augmented Generation (RAG) to combine query-based methods with generative AI models. While query-based models extract information from various sources, such as pdfs, websites, news articles, online databases, etc, they lack the capability to provide semantic answers in natural language. Conversely, generative models can independently create responses but may produce inaccuracies. RAG mitigates these issues by harnessing the strengths of both models and minimizing their weaknesses.
• Data Foundation: The performance and accuracy of AI applications depend heavily on the quality and relevance of their data. We precisely curated the relevant text documents from the UN negotiations' official website through web scraping, categorization, and analysis, forming the app's knowledge base (document store with associated meta data).
• User-Centric Design: Continuous and systematic engagement with sector experts and end-users is vital for developing a user-friendly and functionally robust application. This interaction remains a cornerstone of our ongoing development process.

NegotiateAI can significantly simplify the search and analysis of the INC text documents, aiding negotiatiors in identifying crucial information swiftly and precisely. This tool also aims to increase transparency for non-governmental and civil society observers. We have found that an improved search function for finding relevant documents is often a great help for users, so AI does not necessarily have to be used in the form of generative AI in all circumstances. For this reason, we have structured the app to provide an option to use only filter functions with numerous specifications to generate a list of all relevant documents based on the filter preferences.

Core-function of the app is nonetheless the generative AI part, which allows users to query the document store in natural language, which can allow the user to ask very specific questions such as:

“What is Malaysia's position on chemicals and polymers of concern?”

Countries can also be placed in relation to each other. For example, you can ask how countries position themselves on a certain subject of negotiation and how they relate to each other:

“What is the position of the African Group towards a dedicated multi-lateral fund and how does that differ from the position of Small Island Developing States?”

Limitations:

In line with our “working-out-loud” philosophy, we have noticed limitations with our current system, which we might have underestimated initially. When no filters are applied and the entire database (over 3000 documents) is searched, the tool often becomes very slow and produces incomplete, superficial, and sometimes incorrect answers. This slowness is primarily due to the vast amount of data being processed without filters, leading to heavy computational loads and inefficiencies. Large data volumes require substantial computational resources, and if the server lacks sufficient CPU and memory, it can create bottlenecks. Performance can also be affected if the database is not optimized for broad searches. By encouraging users to set filters and ask specific, concrete questions through better front-end design and guidance, we can mitigate this problem.

Moreover, analyzing data for several countries at once often results in inaccuracies and incomplete responses, particularly for complex questions or when users seek comparisons across many countries or ask questions like "which countries...?". This issue arises because, in a RAG system, only a specific number of paragraphs are selected to answer the question. This selection is influenced by several factors, such as:

• Context Window Limitations: Generative AI models have a maximum context window size, restricting the amount of text they can process at a time.
• Efficiency: To maintain speed, the system selects a subset of the most relevant paragraphs, which can lead to missing crucial information for complex questions.
• Relevance Ranking: Algorithms rank and select the most relevant paragraphs, but may not capture all pertinent information, especially for broad or complex queries.
• Resource Constraints: Processing many documents in real-time requires substantial computational resources, so limiting the number of paragraphs helps manage resources but may exclude some relevant data.

Open Source and Community Engagement:

We support the open source movement and release all our applications and tools on public platforms with open source codes. NegotiateAI can also be found on the HuggingFace platform. The app is freely accessible and can be used without charge. Additionally, the source codes are open to everyone, allowing for comments or suggestions for improvement. We are always open to an interested exchange and would be delighted to receive feedback or improvements.

Using the Tool:

Here is a short screencast demonstrating how to use the tool:

Click HERE to use the app.

The positive reception of NegotiateAI at the fourth negotiation round has paved the way for potential use case adoption and/or scaling. Collaborations with various partners are being considered to enhance the app's capabilities for the fifth negotiation round in November 2024. Our vision extends beyond the current scope aspiring to adapt and deploy this technology across different sectors. With appropriate customization and training, NegotiateAI could become an indispensable tool in diverse contexts, driving innovation and efficiency in a broader array of international and organizational negotiations. We believe that the proof-of-concept has been delivered. Now it is time to evaluate user feedback, team-up with engaged partners and develop a strategy for a second iteration with improved or new features, perhaps with a pivot to a different technical setup to overcome the limitations described above.

As we look to the future, NegotiateAI stands as a testament to how AI can transform complex processes, making them more manageable and effective. By continuing to refine and expand this tool, we are not only supporting the critical mission of the UN to reduce plastic pollution but also pioneering new frontiers in the application of AI for global governance.

In addition, the use of AI in this context could be extended to improve the actual effectiveness of the forthcoming plastics treaty by supporting both in-country planning and implementation of commitments to better manage plastics and ultimately end plastic pollution, as well as compliance with implementation requirements such as reporting and monitoring.