Free beta access. 5 free AI tokens each week. Optional top-ups available. See what is included
Opening Word document…

Perm We will never get perfect bias mitigation – but most effective approach is all-encompassing

UNIDIR (United Nations Institute for Disarmament Research), 2018, “Algorithmic Bias and the Weaponization of Increasingly Autonomous Technologies – A Primer”, The United Nations Institute for Disarmament Research (UNIDIR)—an autonomous institute within the United Nations—conducts research on disarmament and security. UNIDIR is based in Geneva, Switzerland, the centre for bilateral and multilateral disarmament and non-proliferation negotiations, and home of the Conference on Disarmament. The Institute explores current issues pertaining to a variety of existing and future armaments, as well as global diplomacy and local tensions and conflicts. Working with researchers, diplomats, government officials, NGOs and other institutions since 1980, UNIDIR acts as a bridge between the research community and Governments. UNIDIR activities are funded by contributions from Governments and donor foundations, https://www.unidir.org/sites/default/files/publication/pdfs/algorithmic-bias-and-the-weaponization-of-increasingly-autonomous-technologies-en-720.pdf, - Maren Lien

As algorithms approach ubiquity, there is growing understanding that they are not objective and infallible. Algorithms in all domains, including military applications, can exhibit multiple types of biases that arise from different sources, such as unrepresentative training data or inappropriate transfer of the algorithm to a novel context. Some degree of algorithmic bias may be inevitable, as it might not be possible to satisfy all relevant norms with a single process, decision, or algorithm. At the same time, algorithmic biases are not mutually exclusive, as some biases feed into one another. Moreover, not all biases are bad, as some biases can be beneficial to achieving the user’s end goals. Most pointedly, algorithmic bias can arise at every stage of development and deployment, with each stage bringing its own set of considerations and possibilities for the outcome of bias. In many cases, mitigation strategies are available, but they require careful engagement with the details of the situation, as one might not want to mitigate; or might be able to mitigate only some biases; or might address problems by changing the users or broader system; and so forth. Various institutions and organizations are beginning to address these challenges, though policy and technical responses are still in their infancy. As a contribution to the policy response, those participating in the discussion on LAWS within the CCW framework may wish to consider the following questions about algorithmic biases in future systems: • If governments decide to regulate increasingly autonomous weapon systems, rather than adopt an outright ban, which national or international organizations or instruments would be best placed to offer guidance or assistance to address potential algorithmic biases in AWS, including identifying possible mitigation steps? • Given the secretive or non-transparent nature of weapon development and weapon review processes, what sorts of “best practices” can provide confidence that key algorithmic biases have been appropriately identified and mitigated? • Are mitigation steps for algorithmic biases in particular AWS robust against possible loss of communication, interoperability challenges, or reduced human oversight? • How would training of operators and commanders need to be adapted to ensure that they appropriately understand the algorithmic biases in an AWS, in order to maintain trust in the system and ensure its lawful use?

Sign in to recut, update, or save this card.