On January 29, 2024, the Chief Digital and Artificial Intelligence Office (CDAO) at the Department of Defense (DoD) initiated the AI Bias Bounty program. This exercise seeks to crowdsource the detection of biases in AI systems, especially large language models (LLMs). The initiative forms part of CDAO’s broader strategy to safely integrate and optimize AI across the DoD.
Bias in AI systems can have significant consequences, affecting various sectors like law enforcement, finance, healthcare, and more. There have been cases where AI misidentifies people involved in crimes, unjustly denies credit, or misdiagnoses patients, highlighting the critical role of bias in machine learning systems. Often, biased datasets are the root cause of these biases in AI systems.
In damning examples of such bias, offensive content was discovered in the seemingly harmless MIT Tiny Images dataset. The LAION-5B dataset, used for image models like DALL-E, was found to contain child sex abuse material.
Matthew Johnson, the Acting Chief of the DoD’s Responsible AI Division, enthusiastically supports the new initiative. He says the team is passionate about ensuring the DoD’s AI-enabled systems are safe, secure, reliable, and free from bias.
The pioneering project’s initial phase emphasizes discovering and addressing undetected risks associated with LLMs, beginning with open-source chatbots. ConductorAI-Bugcrowd will evaluate the participants’ contributions, and they can earn monetary rewards funded by the DoD.
CDAO Chief Officer, Craig Martell, highlighted the potential policy impact of the results of the AI Bias Bounty program. He said any outcomes could significantly influence future DoD AI policies and adoption.
The US government has enlisted AI researchers and hackers to probe models, test their security, and explore their limitations. This included sponsoring a session at the Def Con hacking convention in Las Vegas. The event marked a critical step towards identifying and correcting bias in language models.
