Biases in machine learning models and big data analytics: The international criminal and humanitarian law implications

Abstract

AbstractAdvances in mobile phone technology and social media have created a world where the volume of information generated and shared is outpacing the ability of humans to review and use that data. Machine learning (ML) models and “big data” analytical tools have the power to ease that burden by making sense of this information and providing insights that might not otherwise exist. In the context of international criminal and human rights law, ML is being used for a variety of purposes, including to uncover mass graves in Mexico, find evidence of homes and schools destroyed in Darfur, detect fake videos and doctored evidence, predict the outcomes of judicial hearings at the European Court of Human Rights, and gather evidence of war crimes in Syria. ML models are also increasingly being incorporated by States into weapon systems in order to better enable targeting systems to distinguish between civilians, allied soldiers and enemy combatants or even inform decision-making for military attacks.The same technology, however, also comes with significant risks. ML models and big data analytics are highly susceptible to common human biases. As a result of these biases, ML models have the potential to reinforce and even accelerate existing racial, political or gender inequalities, and can also paint a misleading and distorted picture of the facts on the ground. This article discusses how common human biases can impact ML models and big data analytics, and examines what legal implications these biases can have under international criminal law and international humanitarian law.