Elodie Collins
RTX BBN Technologies was awarded a contract by the Defense Advanced Research Projects Agency to support the Human-AI Communication for Deontic Reasoning Devops, or CODORD, program.
RTX said Thursday that BBN proposed its DONATELLO, or DeONtic AssistanT Enabled by Large Language mOdels, approach to bridging the gap between policies and AI comprehension. DONATELLO is designed to simplify regulatory texts for computational understanding, allowing AI models to apply regulations in an explainable way.
Why AI Needs Deontic Reasoning
DARPA launched the CODORD program to address the inability of AI tools to understand and interpret complex regulations.
In a previous analysis conducted by DONATELLO subcontractors, researchers found that, when asked about rules filled with exemptions and ambiguities, large language models like OpenAI’s GPT-4 generated erroneous responses and flawed explanations.
RTX said in operational environments, government or military personnel may not have access to decision-makers and experts who can provide actionable advice in regulatory applications. For such instances, AI tools adept with deontic logic can aid interpreting rules.
“Regulatory texts are complex, full of exceptions and nuances that challenge traditional programming. We need systems that can explain why specific actions are permitted or prohibited in ambiguous situations,” commented Brian Ulicny, principal investigator for the BBN. “With DONATELLO, we’re not just interpreting regulations; we’re empowering decision-makers with clarity and precision.”
