Designing AI Tools for Humanitarian Negotiation

2023 - 2025
Humanitarian Negotiations AI Interface

Background

Frontline Associates asked our team to study how humanitarian negotiators were using LLMs and where existing chat interfaces conflicted with their work.

Negotiators described using LLMs to summarize case materials, prepare engagements, and generate options. They also raised concerns about confidentiality, bias, overconfident advice, and the loss of local context.

Humanitarian negotiators in the field

Liesbeth Aelbrecht visited an MSF team working in the General Rural Hospital of Dhi As Sufal district in the south of Ibb Governorate, in Yemen. (Photo: MSF/Majd Aljunaid) Source

Problem

We designed a structured preparation interface around two frameworks already used by practitioners: Iceberg CSS and Island of Agreements. The interface organizes case evidence and proposes editable analyses; negotiators retain control over what enters the final preparation document.

My Role

I led the interviews, interface design and implementation, comparison study, qualitative analysis, and synthesis workshops. The work was conducted with Frontline Associates and negotiators from MSF, the UN, and the ICRC.

Methods

The project had two phases: 14 formative interviews, followed by a within-subjects think-aloud study in which 18 negotiators used ChatGPT and the custom interface. We analyzed the transcripts with open coding and thematic analysis.

Phase 1: Formative Research

  • Conducted 14 interviews with negotiators from MSF, UN, and ICRC
  • Identified key tasks: context analysis, compromise ideation, risk assessment
  • Mapped major concerns: confidentiality, bias, overreliance

Phase 2: Design Probe Study

  • Ran a within-subjects think-aloud study with 18 negotiators
  • Each participant prepared a negotiation plan first using ChatGPT, then my custom probe
  • Used open coding and thematic analysis across transcripts
  • Synthesized findings into journey maps and iterated on the interface

Design requirements

The formative interviews led to four requirements: preserve source context, organize analysis around familiar negotiation frameworks, keep model output editable, and show risks without recommending a single course of action.

Interface implementation

I built the probe in React and Node.js with Express. OpenAI API calls populate structured modules for context analysis, framework application, position mapping, and risk assessment. Every output remains editable.

Technical Implementation

  • Separate components for each negotiation framework
  • Shared state for preparation progress and review
  • Prompt templates tied to specific analytic modules
  • Markdown editing for generated analyses
Interface architecture showing component relationships

Click to view full image

The interface contains four modules:

  1. Context Analysis - Structured input forms for capturing situation details and stakeholder positions
  2. Framework Application - Interactive tools applying Iceberg CSS and Island of Agreements methodologies
  3. Position Mapping - Visual interface for identifying red lines, bottom lines, and zones of agreement
  4. Risk Assessment - AI-assisted identification of potential issues with negotiator-led validation

Model outputs are suggestions. Negotiators can accept, modify, or reject them before continuing.

Workflow of the negotiation interface
Case materials and parties appear on the left; generated Islands of Agreement, Iceberg CSS, and component analyses remain editable on the right.
Red Line/Bottom Line spectrum
The risk view compares both parties' red and bottom lines, then generates a risk matrix for a negotiator-selected scenario.

Study findings

  • Negotiators preferred my interface over ChatGPT for its transparency and alignment with frameworks
  • Less-experienced negotiators used the structure to work through unfamiliar frameworks
  • Experienced negotiators valued editable output and the ability to reject model suggestions

Changes after testing

Context verification

Negotiators can correct extracted context before the system applies a framework.

Output control

Each module exposes explicit accept, modify, and reject controls.

Publication and use

  • A deployed design probe tested in two studies
  • Published at CHIWORK 2025
  • Shared with 200+ professional negotiators through the AI in Negotiation Workshop, hosted by Frontline Associates (workshop PDF)

"ChatGPT, Don't Tell Me What to Do": Designing AI for Context Analysis in Humanitarian Frontline Negotiations

Zilin Ma, Yiyang Mei, Claude Bruderlein, Krzysztof Z. Gajos, Weiwei Pan

Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work (CHIWORK '25)

DOI arXiv

Using Large Language Models for Humanitarian Frontline Negotiation: Opportunities and Considerations

Zilin Ma, Susannah Cheng Su, Nathan Zhao, Linn Bieske, Blake Bullwinkel, Jinglun Gao, Gekai Liao, Siyao Li, Ziqing Luo, Boxiang Wang, Zihan Wen, Yanrui Yang, Yanyi Zhang, Claude Bruderlein, Weiwei Pan

ICML 2024 Next Generation of AI Safety Workshop

Paper