RTX BBN Technologies to Empower AI to Understand and Apply Regulatory 'Legalese'

Estimated reading time: 1 minute

RTX BBN Technologies was recently awarded a contract by the Defense Advanced Research Projects Agency (DARPA) to support its Human-AI Communication for Deontic Reasoning Devops (CODORD) program. CODORD intends to bridge the gap between human regulations and policies and AI comprehension, enabling computers to interpret and apply regulations in a verifiable and explainable way.

BBN's innovative approach is DONATELLO (DeONtic AssistanT Enabled by Large Language mOdels), a human-in-the-loop AI solution designed to simplify and translate complex regulatory texts into executable logic. Prior analyses, conducted by DONATELLO subcontractors, revealed that even state-of-the-art large language models like GPT-4 often produce incorrect answers and flawed explanations when processing regulatory rules rife with ambiguities and exceptions. DONATELLO aims to accomplish this by offering specific and explainable reasoning for these intricate rules, enhancing computational understanding.

"Regulatory texts are complex, full of exceptions and nuances that challenge traditional programming," said Dr. Brian Ulicny, BBN principal investigator. "We need systems that can explain why specific actions are permitted or prohibited in ambiguous situations."

In operational environments like those encountered by the Department of Defense, where expert interpretation isn't always accessible, decision-makers require clear, actionable insights into regulatory applications. Deontic logic, focused on permissibility, prohibition and obligation, provides a framework for AI tools to interpret regulations.

"With DONATELLO, we're not just interpreting regulations; we're empowering decision-makers with clarity and precision," explained Ulicny.

The BBN team includes Johns Hopkins University and Stony Brook University. Work will be performed in Cambridge, Massachusetts; Baltimore, Maryland; and Stony Brook, New York.