On January 29, 2024, the Defense Department's Chief Digital and Artificial Intelligence Office (CDAO) launched AI bias bounty exercises.
These programs aim to advertise crowdsourcing bias detection in AI systems, with a specific deal with large language models (LLMs).
The initiative is an element of CDAO's broader efforts to integrate and optimize secure AI across the Department of Defense.
The impact of unchecked AI bias is far-reaching, manifesting itself in diverse areas, from law enforcement to financial services to healthcare and beyond.
Examples of AI misidentifies people Involvement in crime, denial of unwarranted credit, and misdiagnosis of patients in medical settings highlight the critical nature of bias in machine learning systems.
In many cases, distorted data sets result in bias in machine learning systems. For example, the seemingly innocuous MIT Tiny Images dataset contained offensive content. More recently the LAION-5B datasetused for image models similar to DALL-E, was found to contain child abuse material.
Matthew Johnson, Acting Chief of the Department of Defense's Responsible AI (RAI) Division, spoke about CDAO's initiative and stated: “The RAI team is worked up to guide these AI Bias Bounties as we’re strongly committed to making sure that the Department AI-enabled systems – and the contexts through which they run – are protected, reliable and unbiased.”
The first phase of this groundbreaking program focuses on uncovering and addressing the unknown risks related to LLMs, starting with open source chatbots.
Participants can earn money rewards based on their contributions, assessed by ConductorAI-Bugcrowd and funded by the Department of Defense.
The results could guide recent guidelines and suggestions. Craig Martell, Chief Digital and Artificial Intelligence Officer, noted: “Given the Department’s current deal with risks related to LLMs, the CDAO is actively monitoring this area; The end result of the AI bias bounties could have a significant impact on future Department of Defense AI policies and adoption.”
The US government has recruited AI researchers and hackers to review models and assemble red teams, even in the event that they do supported a session at Def Con Hacking convention in Las Vegas.