US Department Of Defense Launches Bounty Program To Find Bias In AI Model

Pebnatgon, headquarters of the United States Department of Defense (DoD) (photo: x @CaptCoronado)

JAKARTA - The United States Department of Defense (DoD) recently launched a bounce program aimed at finding examples of legal partiality in artificial intelligence (AI) models that can be implemented in real life.

Participants will be tasked with trying to find clear examples of partiality from the big language model (LLM). According to a video connected on the bounce information page taking sides, the model being tested is Meta open-source LLama-2 70B.

The purpose of this contest is to identify realistic situations with potential real-world applications where large language models can show partiality or results that are systematically wrong in the context of the Department of Defense, the content of the announcement of the video DoD.

Although it is not explicitly explained in the original Pentagon post, clarification in the contest rules and in the video shows that DoD is seeking legal sides against groups of protected persons.

In an example shown in the video, vortices give instructions to AI models explaining that the model must respond as a medical professional. The model was then asked with specific medical questions for black women and the same questions as instructions to produce specific outputs for white women. The output results, according to vortices, are not true and show clear partiality to black women.

While it is known that the AI system can be triggered to produce biased outputs, not every partiality has the potential to appear in real-world scenarios specifically related to DoD's daily activities.

Therefore, this scammy of partiality will not reward each example. Instead, it is run as a contest. The three best contributions will receive 24,000 US dollars (Rp380 million) prizes to be awarded, while each approved participant will receive 250 US dollars (Rp3.9 million).

Submissions will be judged based on rubriks consisting of five categories: how realistic the output scenario is, its relevance to protected classes, supporting evidence, a concise description, and how many orders are needed to replicate it (with less effort getting higher scores).

According to the Pentagon, these are the first of the two "bouncedowns" they will run.