How AI Is Transforming Diplomacy and Conflict Management

Contents

Executive Summary

The Belfer Center at Harvard Kennedy School is launching Move 37, an ambitious research initiative to develop AI tools that augment (rather than replace) human diplomats and negotiators. The project seeks to transform how international negotiations function by applying computational methods—ranging from language models to game theory and strategic analysis—while maintaining human authority and accountability. The speakers emphasized that diplomacy is fundamentally complex, involving multiple state actors, competing interests, and high stakes, and that AI can help manage information overload, track shifting positions, and explore new strategic options—but only if designed with rigorous safeguards against bias, misrepresentation, and loss of human agency.

Key Takeaways

AI for Diplomacy Requires Domain-Specific Design: Generic LLMs and broad AI architectures are insufficient; diplomatic AI must account for strategic deception, evolving institutions, and multi-actor preference aggregation—problems requiring custom computational approaches.
Transparency and Explainability Are Non-Negotiable: Decision-makers must be able to articulate how they relied on AI, why outputs were produced that way, and what counterarguments were considered. Without this, AI becomes a unaccountable black box in high-stakes contexts.
Close the Public Sector AI Literacy Gap Before Deployment: Training hundreds of thousands of government workers in responsible AI use is as important as building the tools themselves. Current rates (26% understand ethics frameworks) indicate systems are being deployed by people unprepared to use them responsibly.
Deliberate Inclusion of Cultural and Linguistic Diversity is Essential: AI systems trained on limited data will reproduce biases and exclude non-Western epistemologies. Developers must actively ensure diverse language representation and philosophical frameworks—not as an afterthought, but from training onward.
Maintain Human Agency as the Overarching Design Principle: The risk is that AI becomes so capable that it atrophies human decision-making. Successful deployment means keeping humans "above the algorithm"—making autonomous strategic choices with AI as supporting analysis, not replacing human judgment.

Key Topics Covered

AI in Diplomacy and Negotiation: Application of AI tools to support multilateral and bilateral negotiations
Complexity of Modern Negotiations: The layered nature of diplomatic negotiations involving multiple states, internal stakeholders, and evolving positions
Current Tools & Limitations of LLMs: Why large language models alone are insufficient; need for broader computational architectures (game theory, decision analysis, machine learning)
Responsible AI Deployment: Human-in-the-loop requirements, transparency, and accountability frameworks
Three Core Challenges: Representation (dynamic institutions), inference (strategic misrepresentation/deception), and specifying success criteria
Breaking Down Diplomatic Tasks: Research, analysis, strategizing, and real-time execution as distinct domains for AI application
Government AI Adoption Barriers: Skills gaps, knowledge deficits, "sleeping at the wheel" phenomenon, and pilot-to-scale disconnect
Cultural Representation in AI Models: Risks of bias, language representation, and cultural assumptions embedded in training data
Intelligence and Analysis: Role of AI in supporting foreign policy analysis and identifying both risks and opportunities
Power Dynamics & Data Access Inequality: How unequal access to AI tools and datasets could shift geopolitical leverage

Key Points & Insights

Diplomacy is Not a Board Game: Unlike chess or Go, diplomatic negotiations involve institutions that can be created and destroyed mid-negotiation, strategic misrepresentation, and fluid success criteria—requiring AI approaches beyond game-playing algorithms.
Language Models Are Insufficient: LLMs lack verifiable fluency in international politics, suffer from opacity problems (problematic for high-stakes decisions requiring accountability), and miss 80 years of accumulated AI techniques (game theory, decision analysis) that could be more effective.
Four-Level AI Architecture for Diplomacy:
- Research: Autonomous agents for data gathering with source validation
- Analysis: Processing and synthesizing large volumes of documents and positions
- Strategizing: Red-team simulation, preference mapping, option generation
- Execution: Real-time transcription, translation, and dynamic adaptation
The "Person Tracking" Problem: Modern diplomats manually track dozens to hundreds of actors' positions. AI can automate position mapping, but current tools are underutilized—Gabriella's UNESCO AI ethics negotiation (193 countries during COVID) would have benefited from automated tracking of alignment patterns and isolated blockers.
Massive Public Sector Skills Gap: Only 26% of public servants implementing AI understand their country's ethical frameworks; ~75% are "freestyling" without proper training. This represents a critical vulnerability as governments deploy AI without baseline literacy.
The "Sleeping at the Wheel" Trap: Humans systematically underestimate AI limitations after false negatives and overestimate capabilities when accuracy approaches 85-90%, creating dangerous overconfidence in high-stakes decisions.
Cultural and Linguistic Bias is Structural: AI models trained on dominant languages and Western individualist frameworks (e.g., maximizing individual welfare) systematically fail to capture non-Western philosophies (e.g., Ubuntu: "I am because you are"). This is not a technical fix—it requires deliberate representation.
The Human Agency Problem: As AI takes on analytical and strategic functions, human agency can atrophy unless systems are explicitly designed to keep humans "above the algorithm" (making autonomous decisions) rather than "below it" (responding to algorithmic directives).
Data Poisoning and Adversarial Negotiations: In zero-sum disputes, actors can deliberately poison training data, inject malicious prompts, or introduce biased datasets. Neutral mediation by AI assumes good-faith participation—unrealistic in many geopolitical scenarios.
The Move 37 Three Commitments:
- Human authority remains central (no delegation of war/peace decisions)
- Tools must be modular and transparent (visibility at each computational stage)
- Augmentation must be scoped appropriately (context and institution-specific)

Notable Quotes or Statements

On the complexity of negotiation:

"Negotiations are complex and they evolve over time. What might look like just two states negotiating with each other bilaterally is actually a whole set of issues that are on the table... a whole lot of teams sitting behind these principal negotiators." — Savina Aniya

On LLM limitations:

"Language models are remarkable and their use should be carefully scoped... their fluency isn't necessarily verifiable in international world politics... opacity is not viable because high stakes negotiations require accountability both democratically and internally." — Charlie Pausniak

On cultural representation in AI:

"When the models that we are developing are maximizing individual welfare, and the Ubuntu philosophy is 'I am because you are'—how do you capture this? The only answer I have is try to be representative." — Gabriella Ramos

On the "sleeping at the wheel" phenomenon:

"It comes across as so smart and so brilliant... that it's almost overwhelmingly smart and you're like it must have covered everything. That's a default assumption. So I think giving us the human tools and psychological counterarguments to deal with this is really important." — Robin Scott

On public sector AI deployment:

"About three quarters of all the people rolling out this technology are freestyling. That's terrifying." — Robin Scott (regarding ethical framework understanding)

On AI as supporter, not decider:

"You want to get the Oscar... not the AI. The AI is a supporting cast." — Gabriella Ramos

On the human judgment requirement:

"What happens when the AI says, 'Here's a thing'? I don't just trust it as a priority. I have to apply human judgment to what I'm hearing." — Michael (paraphrasing diplomat perspective)

Speakers & Organizations Mentioned

Speaker	Role/Title	Organization
Michael	Program Director	Belfer Center, Harvard Kennedy School
Savina Aniya	MPP Student & Research Fellow	Belfer Center, Harvard Kennedy School
Charlie Pausniak	Full-time Research Fellow	Belfer Center, Harvard Kennedy School
Gabriella Ramos	Former Assistant Director General for Social & Human Sciences	UNESCO; former G20/G7/OECD policy negotiator
Nandita Balakrishnan	Director of Intelligence	Special Competitive Studies Project (SCSP, funded by Eric Schmidt)
Robin Scott	CEO & Co-Founder	Apolitical (platform for government innovation)
Conor Artigus	Former Co-Chair of UN AI Advisory Panel; Spain-India High Commissioner for AI	(referenced, not present)
Sam Doors	Senior Adviser	Oxford University AI Governance Initiative; former diplomat (Kofi Annan era)

Key Institutions Mentioned:

Belfer Center (Harvard Kennedy School)
UNESCO
G20, G7, OECD
Special Competitive Studies Project (SCSP)
Apolitical
UN AI Advisory Panel
Stanford HAI (collaborator on training)
Oxford University AI Governance Initiative
Swiss multilingual LLM initiative (quasi-Swiss government/multilateral)
JPAL South Asia

Technical Concepts & Resources

AI/ML Methods Referenced:

Large Language Models (LLMs): Critiqued as insufficient for diplomacy without broader computational context
Game Theory: 80+ years of research; applicable to strategic interactions under uncertainty
Decision Analysis: Formal methods for preference aggregation and option generation
Machine Learning (supervised, unsupervised, self-supervised)
Natural Language Processing (NLP): Position tracking, document synthesis, transcription
Predictive Models: Geopolitical event prediction; state-behavior forecasting
Red-Teaming & Simulation: Stress-testing negotiation scenarios
Autonomous Research Agents: Gathering and validating source information

Specific Applications Mentioned:

Position Tracking: Automated mapping of actor positions over time
Counterpart Biographies: AI-generated profiles of negotiators and their constraints
Document Synthesis: Integrating thousands of transcripts and draft documents (e.g., UNESCO received 55,000 public comments)
Preference Gap Analysis: Identifying common ground and divergence between parties
Strategic Option Generation: Expanding negotiators' understanding of possible outcomes
Multilingual Translation & Transcription: Real-time language services during negotiations
Bias Auditing & Counterfactual Analysis: Stress-testing AI recommendations for hidden assumptions

Data & Model Challenges:

Language Representation Imbalance: Most models trained on English; non-European languages underrepresented
Cultural Assumption Embedding: Models reflect Western individualist values; non-Western epistemologies (Ubuntu philosophy) not captured
Data Poisoning Risk: Actors can deliberately inject false or biased data in adversarial contexts
Prompt Injection Attacks: Manipulation of AI inputs to produce biased outputs
Black Box Opacity: Inability to fully explain model decisions (limits accountability)

Evaluation & Governance Concepts:

Human-in-the-Loop (HITL): Maintaining human decision authority at critical junctures
Explainability & Transparency: Requirement to show computational chain of reasoning
Responsible Deployment Framework: Scoping augmentation to context and institution
Verifiable Fluency: Ensuring AI outputs are accurate in domain-specific contexts
Modularity: Breaking diplomatic tasks into distinct AI-supportable subtasks
Calibration: Matching AI tool capability and confidence to decision-making needs

Knowledge Gaps Identified:

Only 26% of public servants implementing AI understand their country's ethical frameworks
70% of leaders report AI pilots, but only 45% have evaluation plans
Leaders not using technology themselves (barrier to understanding capabilities/limitations)
Gap between "AI talk" and "AI action" in government adoption

Project & Initiative Overview

Move 37 Project:

Launched ~1 year before this talk
Multi-year research initiative at Belfer Center
Focus: AI augmentation of diplomatic negotiation and statecraft
Approach: Combination of technical tool development, policy guidelines, practitioner interviews
Output: Framework for responsible AI deployment in high-stakes diplomacy
Current Phase: Engagement with diplomats, negotiators, and stakeholders for feedback and validation

Emerging Tech Program (Harvard):

Parent program for Move 37
Broader mandate: examining where emerging tech creates policy frontiers
Addresses geopolitics, global stability, and conflict implications

Apolitical:

10+ years old; 500K+ public servants on platform (160 countries)
Training: Nearly 400K public officials in responsible AI use
Partnerships: Stanford HAI and leading universities
Focus: Bridging gap between tech development and public sector adoption

Special Competitive Studies Project (SCSP):

Funded/sponsored by Eric Schmidt
Successor to National Security Commission on AI (US)
Focus: AI role in economic competition, national security, geopolitics
Three pillars: Meta-level AI/geopolitics understanding, public sector AI literacy, targeted use cases

Document Quality Note: Transcript contains multiple repetitions and speech artifacts, suggesting it was automatically transcribed from video. Summary captures substantive content while noting these transcription limitations.