United States Department of Defence (DoD) has reportedly turned to the public to identify biases in artificial intelligence chatbots by officially launching the department's first two AI Bias Bounty exercises.
The official announcement states that these new crowdsourcing initiatives called "bias bounties" aim to identify prejudice in AI systems, with the first being available to the public now and the second exercise set to follow after.
The exercises reportedly aim to produce new ideas for algorithmic auditing and red-teaming AI models while also enabling the testing of identified hazards and guaranteeing the systems' objectivity in the context of their specific deployment.
In collaboration with ConductorAI-Bugcrowd and BiasBounty.AI, the two AI Bias Bounties are being led by the Chief Digital and Artificial Intelligence Office (CDAO) Responsible AI (RAI) Division, with guidance from the CDAO Defense Digital Services Directorate. The First AI Bias Bounty exercise is reportedly set to start from January 29 to February 27, 2024.
To enable the careful mitigation and control of such risks, the first bounty exercise is reportedly aimed at discovering unknown areas of risk in Large Language Models (LLMs), starting with open-source chatbots.
Publicly-Available AI Chatbot Hacking
The department reportedly encourages the public, including those with no coding knowledge, to participate, adding that participants can receive monetary bounties financed by the Department of Defense.
The DoD is offering a $24,000 pool to promote participation; the first placer is slated to receive $9,000, the second place will receive $5,000, the third place will receive $2,000, and if should participants make it through the qualification round (five acceptable submissions), they may receive $250 until the fund pool is depleted, as per the official registration webpage.
The participants have to be US citizens, divided among submitters according to a scoring system and an assessment from ConductorAI and Bugcrowd.
The registration page states that the competition begins on January 29 and runs through March 11, implying that the yet-to-be-announced second AI Bias Bounty exercise could happen between February 27 and March 11, 2024.
This is not the first time The DoD has turned to the public to find cyber vulnerabilities; the DOD has previously deployed so-called bug bounty schemes, such as its 'Hack the Pentagon' program that also sought to find and fix security flaws in Defense Department websites that are accessible to the public.
AI Bias Concerns
Concerns regarding AI bias continue to grow. In May of last year, the Equal Employment Opportunity Commission (EEOC) reportedly published new recommendations to assist firms in preventing discrimination while utilizing automated hiring procedures.
The EEOC fined iTutorGroup $365,000 for utilizing software to automatically reject applicants over 40 in August of last year, marking the settlement of the agency's first automation-based complaint. The business, which tutors Chinese students in English, denied any misconduct in the case.
Related Article : Japan Strengthens Export Controls to Curb Military Use, Focuses on China, Russia Safeguards
