Research, ideas, and leadership for a more secure, peaceful world
This primer is written for diplomatic and policy professionals who need to understand how artificial intelligence (AI) systems are tested before they can be trusted in deployment.
This primer is written for diplomatic and policy professionals who need to understand how artificial intelligence (AI) systems are tested before they can be trusted in deployment. It explains, in practical terms:
For a basic introduction to AI systems and how they are being used in the diplomacy, see “AI for Diplomats: A Technical Primer.”
AI systems are moving beyond consumer tools into high-stakes government workflows to support geopolitical analysis, planning, and monitoring.1 Impressive demonstrations are insufficient in these domains. They can summarize vast documents, draft text, translate languages, write software, and analyze images.2 But a basic question follows from these growing capabilities: How can we trust them? 3
To answer that question, AI academics and labs have developed evaluations, often called “evals.” These are structured tests designed to measure what an AI system can do, where it fails, and what risks may exist when it is used in real-world settings.4
These applications are useful only if the systems are evaluated against the realities and requirements of diplomatic work. Building trust is paramount. Current evaluation practices are centered on commercial clients, with models trained on general sources. For many meetings, ambiguous language might be unimportant. For diplomatic negotiations, however, these linguistic subtleties hold great importance.
Diplomatic practitioners face a high level of risk aversion –– any AI deployments must match the stakes of the environment. The tolerance for error in diplomatic contexts is low. Without evaluation frameworks designed for these conditions, diplomatic institutions risk adopting AI tools based on benchmark scores that have little bearing on operational performance.
An AI evaluation is a structured test of how well an AI system performs a specific task. The task might be simple, such as answering factual questions. It might be technical, such as solving math problems or identifying cybersecurity vulnerabilities. Evaluations measure both what a system can reliably deliver and where it fails.5
A basic evaluation has three parts:6
For example, an evaluation might test whether a model can:
Some evaluations are like examinations: they test whether a system can answer questions or solve problems. Others resemble stress tests: they expose a system to difficult, adversarial, or high-pressure conditions to see where it breaks.
The intensity and scope of an evaluation should match the criticality of the model’s intended use. A model used to recommend restaurants does not need the same evaluation standard as one used to summarize ceasefire terms, analyze treaty language, or support crisis diplomacy.
When the same evaluation is applied across multiple systems, it becomes a benchmark. Benchmarks are shared yardsticks that allow models to be compared on the same task.
The results of benchmarks are often published as leaderboards, where models are ranked by their performance. These leaderboards can shape public perception, procurement decisions, and the research priorities of AI companies.
For example, some benchmarks test advanced academic reasoning across mathematics, science, humanities, and law. Others test software engineering or domain-specific tasks. Examples of this include Humanity’s Last Exam, FrontierMath, and ARC-AGI. These tests are useful because they provide structured evidence that models are improving in certain areas.
But a benchmark score is not the same as real-world performance. Systems may score highly on general summarization tasks and still perform poorly in high-context diplomatic exchanges. A summary may preserve the key facts, but misread subtle signals, lose hedging language, miss a references, or flatten nuance. In diplomacy, these details may be the most important pieces of information.
Benchmarks are incredibly specific. Closing the gap between test performance and reliable operational performance is the central challenge of AI evaluation.
Evaluations serve three main functions for policymakers.
Governments and international organizations are increasingly asked to choose among AI tools for translation, document analysis, meeting support, open-source monitoring, or legal review. Evaluations provide a starting point.
Without open evaluations, procurement decisions risk being based on marketing claims or polished private demonstrations to a non-technical audience. A model may look impressive in a live demonstration, but without serious evaluations it is difficult to know how it performs across a wide range of cases.
Evaluations help policymakers understand model capabilities and limitations and identify use cases, potential restrictions, and necessary levels of oversight.
Tools may be excellent at one task in a domain but unreliable in related tasks. A writing tool may be fine for a first-pass news summary, but insufficient for producing an authoritative record. A legal tool might be useful for accurately finding precedent and struggle to effectively interpret it.
Evaluations sharpen these boundaries so teams can match tools, tasks, and settings.
Diplomatic settings create evaluation challenges that are different from those in consumer or commercial applications. These challenges will require a strong awareness of those building, buying, and using AI tools.
Because benchmarks drive model adoption, AI companies optimize for what the evaluations measure. When a popular benchmark emphasizes mathematical reasoning, labs invest in mathematical capabilities. If benchmarks reward software engineering, labs invest in coding performance.
The choice of what to evaluate is not neutral. It determines what the AI field learns to value. If evaluations ignore diplomacy, negotiation, or strategy, then these capabilities are less likely to be systematically improved and understood.
This is important for governments. If diplomatic institutions rely on benchmarks created for commercial or academic settings, they will inherit tools optimized for other purposes.
Interpreting evaluation scores is not straightforward. If a model achieves 95% on Humanity's Last Exam, does that mean it reasons as well as a world-leading expert? Can it make novel scientific discoveries, or has it merely memorized patterns that approximate expert-level answers? The score alone can’t answer these questions.
Existing AI evaluations are useful, but they are often poorly matched to diplomatic use. The main problems are validity gaps, the lack of clear ground truth, and fragile evaluation practice.
The most basic problem for any AI evaluation is validity: Does the test measure the capability it claims to?
A test has high external validity when success on the test indicates success in the real world. For example, a model may perform well on a general legal reasoning benchmark. That does not prove it can support treaty interpretation in a live negotiation. A model may perform well on a multilingual translation benchmark. That does not prove it can preserve diplomatic subtlety across Arabic, Mandarin, French, Russian, and English in a crisis setting.
Several capabilities central to diplomatic practice remain absent from existing evaluation frameworks. Diplomatic work often involves:
A useful example is Meta's CICERO system, which achieved human-level performance in the board game Diplomacy by combining language models with strategic reasoning. This was a genuine technical achievement, but the board game is a closed system with fixed rules, limited actions, known players, and a defined endpoint.
Real diplomacy is not like that. The rules are changing and contested, the players change, and objectives are contested. Many components are not visible: side payments, information commitments, domestic politics, reputation, and historical grievances all matter. Success may not be visible for years.
A system that performs well in a game-like evaluation environment may still fail catastrophically in real negotiations.
Many evaluations assume there is a correct answer. In math, logical reasoning, or medical diagnoses, this may be true. Diplomacy lacks this kind of clear answer.
A government may simultaneously a wide range of competing goals. From military security to economic welfare, these objectives can conflict, and their relative weighting is a matter of political judgment. They are also dynamic, shifting with interrelated concessions or exogenous shocks as a negotiation unfolds. Evaluations that score outputs against a single ‘correct’ answer ignore the core feature of these settings: that what counts as success is itself contested.
For example, if an AI system is asked to draft language, one option may be clearer, another may be more acceptable to all parties, while a third may preserve some necessary ambiguity. What is best depends on political judgment.
The same issues apply to negotiation strategies. One option may maximize short-term leverage while another may reduce escalation risk. There may be no single technical answer.
This does not mean diplomatic evaluation is impossible. It means, however, that AI systems and evaluations designed for diplomacy must have several considerations. Our project, Move37, takes a task-based approach. By breaking down complex strategies into small, verifiable tasks, we can support human teams in exercising judgment with AI augmentation.
Even when evaluations test relevant capabilities, the broader evaluation ecosystem has weaknesses.
Poor reporting is a systemic issue. Many evaluation reports do not provide enough detail about how the test was conducted. They may omit the exact prompts used, the scoring process, the model settings, the data sources, or the failure cases. Without this information, results are hard to interpret or reproduce.
Model instability means that small changes from developers can drastically alter the behavior of models in deployment. Vendors may adapt key features of the product without warning. A tool that performed well in one month may behave differently in another.
Benchmark gaming can occur when specific tests become important and companies optimize for solving them. This can improve performance but also distort incentives. A model may become better at passing the test without becoming better at the underlying real-world task.
Contamination occurs when benchmark questions appear in training data. If a model has seen the test before, its score may exaggerate its true capability.
These dynamics reduce confidence in headline results and complicate procurement, governance, and risk management.
For diplomatic institutions, these weaknesses matter. When a ministry of foreign affairs or an international organization procures an AI tool, an evaluation report is often the primary basis for the decision. The absence of independent, third-party evaluation standards creates two compounding market failures:
The result is that institutions may deploy systems that are weakest in their most consequential settings.
The evaluation community has called for more rigorous scientific methods and greater attention to under-evaluated dimensions such as ethics and bias [14]. Diplomacy requires all of that, as well as evaluation methods tailored to its practices. Several concrete steps would help close the gap between evaluation practice and operational need.
Evaluations are validated against four criteria:
For diplomatic negotiations in particular, the following activities could help close the gap:
Static question-and-answer benchmarks cannot capture the interactive, sequential, and strategic nature of diplomatic exchange. Evaluation environments should simulate information asymmetries, time pressures, coalition dynamics, and the give-and-take of real negotiation. These scenarios should draw on declassified case studies and be developed in collaboration with experienced practitioners.
Closing this gap requires sustained collaboration between evaluation researchers and experienced diplomats to identify which capabilities matter, what failure looks like in practice, and how to construct test scenarios that reflect the competing objectives and contested values of diplomacy. This research should precede any effort to formalize standards.
Diplomatic relationships and negotiations unfold over months and years. AI systems should be evaluated in both single-point and extended settings to assess whether one-off performance holds and how the system adapts to changing contexts.
Static question-and-answer tests are not enough for diplomacy. Negotiations unfold over time. Parties respond to one another. Information changes. Evaluation environments should therefore include interactive scenarios that simulate diplomatic exchange. These environments should draw on declassified case studies and be developed with experienced diplomats, mediators, regional experts, interpreters, and international lawyers.
Standardized validation metrics and requirements across international countries can make collaboration and evaluation clearer. Akin to the standardization efforts undertaken in the nineteenth century by the World Postal Union and the International Telegraphic Union.
AI evaluations are becoming central to AI governance, procurement, and deployment. They are not currently considering diplomacy, despite being an institutional necessity.
Diplomatic tasks place unusual demands on AI systems. Current evaluation methods were not designed for strategic ambiguity, cultural fluency, or contested definitions of success. As these systems become more embedded in high-stakes diplomatic contexts, the establishment of rigorous, valid, and practical diplomatic evaluation frameworks must be a priority.
The institutions and governments that invest now in building evaluation capacity will be better positioned to adopt useful AI tools, reject unsafe ones, detect failures early, and shape the standards that govern the next generation of diplomatic technology.
[1] Bengio, Y., Mindermann, S., Privitera, D., Besiroglu, T., Bommasani, R., Casper, S., Choi, Y., Fox, P., Garfinkel, B., Goldfarb, D., Heidari, H., Ho, A., Kapoor, S., Khalatbari, L., Longpre, S., Manning, S., Mavroudis, V., Mazeika, M., Michael, J., . . . Zeng, Y. (2025, January 29). International AI Safety Report. arXiv.org. https://arxiv.org/abs/2501.17805
[2] Burden, J. (2024, July 12). Evaluating AI Evaluation: Perils and Prospects. arXiv.org. https://arxiv.org/abs/2407.09221
[3] Chang, Y., Wang, X., Wang, J., Wu, Y., Zhu, K., Chen, H., Yang, L., Yi, X., Wang, C., Wang, Y., Ye, W., Zhang, Y., Chang, Y., Yu, P. S., Yang, Q., & Xie, X. (2023). A survey on evaluation of large language models. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2307.03109
[4] Liao, Q. V., & Xiao, Z. (2023, June 1). Rethinking model evaluation as narrowing the Socio-Technical gap. arXiv.org. https://arxiv.org/abs/2306.03100
[5] Ma, Z., Mei, Y., Bruderlein, C., Gajos, K. Z., & Pan, W. (2024, October 11). “ChatGPT, don’t tell me what to do”: Designing AI for context analysis in humanitarian frontline negotiations. arXiv.org. https://arxiv.org/abs/2410.09139
[6] Maslej, Nestor, Loredana Fattorini, Raymond Perrault, Yolanda Gil, Vanessa Parli, Njenga Kariuki, Emily Capstick, et al. 2025. “Artificial Intelligence Index Report 2025.” arXiv. https://doi.org/10.48550/arXiv.2504.07139
[7] Ott, S., Barbosa-Silva, A., Blagec, K., Brauner, J., & Samwald, M. (2022). Mapping global dynamics of benchmark creation and saturation in artificial intelligence. Nature Communications, 13(1), 6793. https://doi.org/10.1038/s41467-022-34591-0
[8] Rauh, M., Marchal, N., Manzini, A., Hendricks, L. A., Comanescu, R., Akbulut, C., Stepleton, T., Mateos-Garcia, J., Bergman, S., Kay, J., Griffin, C., Bariach, B., Gabriel, I., Rieser, V., Isaac, W., & Weidinger, L. (2024). Gaps in the safety evaluation of generative AI. Proceedings of the AAAI/ACM Conference on AI Ethics and Society, 7, 1200–1217. https://doi.org/10.1609/aies.v7i1.31717
[9] Sevilla, Jaime. 2024. “Can AI Scaling Continue Through 2030?” Epoch AI. August 20, 2024. https://epoch.ai/blog/can-ai-scaling-continue-through-2030.
[10] Sevilla, Jaime, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, and Pablo Villalobos. 2022. “Compute Trends Across Three Eras of Machine Learning.” In 2022 International Joint Conference on Neural Networks (IJCNN), 1–8. https://doi.org/10.1109/IJCNN55064.2022.9891914.
[11] Shevlane, T., Farquhar, S., Garfinkel, B., Phuong, M., Whittlestone, J., Leung, J., Kokotajlo, D., Marchal, N., Anderljung, M., Kolt, N., Ho, L., Siddarth, D., Avin, S., Hawkins, W., Kim, B., Gabriel, I., Bolina, V., Clark, J., Bengio, Y., . . . Dafoe, A. (2023, May 24). Model evaluation for extreme risks. arXiv.org. https://arxiv.org/abs/2305.15324
[12] Staufer, L., Yang, M., Reuel, A., & Casper, S. (2025, April 18). Audit Cards: Contextualizing AI evaluations. arXiv.org. https://arxiv.org/abs/2504.13839
[13] Weidinger, L., Rauh, M., Marchal, N., Manzini, A., Hendricks, L. A., Mateos-Garcia, J., Bergman, S., Kay, J., Griffin, C., Bariach, B., Gabriel, I., Rieser, V., & Isaac, W. (2023, October 18). Sociotechnical safety evaluation of generative AI systems. arXiv.org. https://arxiv.org/abs/2310.11986
[14] Weidinger, L., Raji, I. D., Wallach, H., Mitchell, M., Wang, A., Salaudeen, O., Bommasani, R., Ganguli, D., Koyejo, S., & Isaac, W. (2025, March 7). Toward an evaluation science for generative AI systems. arXiv.org. https://arxiv.org/abs/2503.05336
From Program on Diplomacy and Statecraft
