Navigated to Humanity Centered Data | UN Refugee & IDP Tracker
    All AI articles
    Impact

    AI for Documenting War Crimes: Eyewitness, Mnemonic, and the Berkeley Protocol

    By the Humanity Centered Data Editorial Team
    June 19, 202611 min read

    Documentation has shifted from analog to digital, and AI is the latest layer

    The investigation of grave international crimes — genocide, crimes against humanity, war crimes — has always depended on documentation. The Nuremberg trials relied on captured Nazi documents; the ICTY and ICTR built cases on witness testimony and forensic exhumations. By the 2010s, with smartphones in conflict zones, digital open-source evidence began to enter the evidentiary record. By 2026, AI-assisted ingestion, deduplication, geolocation, and pattern detection are standard in the leading documentation organisations.

    The Berkeley Protocol as the legal frame

    The Berkeley Protocol on Digital Open Source Investigations — published by the UN Office of the High Commissioner for Human Rights and the UC Berkeley Human Rights Center in 2022 — is the operative reference. It defines minimum standards for collection, preservation, verification, and analysis of digital open-source information intended for use in legal proceedings. Compliance is the difference between evidence that is admissible at the ICC, ICJ, or in universal-jurisdiction prosecutions and material that is merely indicative.

    The organisations doing the work

    Several organisations operate at scale. [Mnemonic](https://mnemonic.org/) archives video and image evidence from Syria, Yemen, Sudan, and Ukraine; its Syrian Archive has preserved millions of files at risk of platform takedown. [Eyewitness to Atrocities](https://www.eyewitness.global/) provides a chain-of-custody mobile app for human-rights documenters. [Bellingcat](https://www.bellingcat.com/) publishes open-source investigations using public tools. [GLAN](https://www.glanlaw.org/) and the Commission for International Justice and Accountability pair documentation with legal preparation. The International Criminal Court Office of the Prosecutor has its own digital-evidence intake.

    Where AI adds value

    Five contributions are operationally established. Triage at scale: ML classifiers flag potentially evidentiary content from millions of social-media items per day, allowing human investigators to focus. Deduplication and clustering: identifying multiple uploads of the same underlying footage. Object and weapon recognition: detecting prohibited munitions and identifying perpetrator markings. Geolocation assistance: matching imagery against satellite and street-level references. Pattern analysis across incidents: identifying systematic conduct that supports crimes-against-humanity charges. None of these substitutes for human verification; all of them make verification feasible at the volume of contemporary conflict imagery.

    Where AI is risky

    Three risks need explicit management. Hallucinated identifications of people, weapons, or locations would undermine credibility if introduced into evidence; conservative thresholds and human review are essential. Authenticity assessment is hard, and AI 'authenticity scores' have not yet reached a reliability level that survives cross-examination. Chain of custody must be preserved through any AI processing step; tools that re-encode, re-compress, or transform media without logging risk breaking admissibility.

    What good practice looks like

    Mature documentation operations in 2026 share a common architecture. Original media is preserved unaltered in a tamper-evident archive. AI tools operate on copies and produce structured metadata, not transformed media. Every analytical step is logged with model, version, and parameters. Human investigators sign off on every finding intended for legal use. The International Bar Association eyewitness guidelines and the Engine Room's open-source investigation resources document the operational protocols.

    Further reading and primary sources

    • Berkeley Protocol: https://www.ohchr.org/en/publications/policy-and-methodological-publications/berkeley-protocol-digital-open-source
    • Mnemonic: https://mnemonic.org/
    • Syrian Archive: https://syrianarchive.org/
    • Eyewitness to Atrocities: https://www.eyewitness.global/
    • Bellingcat: https://www.bellingcat.com/
    • CIJA: https://cijaonline.org/
    • ICC: https://www.icc-cpi.int/
    Ethics

    The Risks of AI in Humanitarian Work: Bias, Privacy, and Accountability (2026)

    AI tools are now woven through humanitarian operations. The benefits are real and so are the risks. A frank look at the bias, privacy, and accountability gaps shaping the sector in 2026.

    10 min read
    Ethics

    Deepfakes in Conflict Zones: How AI-Generated Disinformation Affects Refugee Crises

    How deepfakes have entered conflict information environments, what damage they do, and how serious newsrooms verify imagery in 2026.

    10 min read
    🌍
    AI and Humanitarian Response

    The Risks of AI in Humanitarian Work: Bias, Privacy, and Accountability (2026)

    AI tools are now woven through humanitarian operations. The benefits are real and so are the risks. A frank look at the bias, privacy, and accountability gaps shaping the sector in 2026.

    10 min read
    Impact

    How AI Is Being Used to Predict Refugee Crises Before They Happen (2026)

    Machine learning models are now feeding into UNHCR, IOM, and World Bank early warning systems. A clear look at what AI can and cannot predict about forced displacement in 2026.

    9 min read
    Impact

    AI vs Traditional Methods: How Humanitarian Organizations Are Counting Displaced People in 2026

    Registration desks, household surveys, and satellite based machine learning estimates are now being combined to count displaced populations. A practical comparison of what each method gets right and wrong in 2026.

    8 min read
    Impact

    What Is the Humanitarian AI Paradox? (2026)

    In 2026, 93 percent of humanitarian workers report using AI tools, but only 8 percent work in organizations with a fully integrated AI strategy. The gap between individual adoption and institutional readiness is the defining tension of the sector.

    6 min read
    Advertisement
    Advertisement