Auditing analysis reveals generative logic instability and threshold injection in LLM-generated code, highlighting risks of deploying unverified discriminatory logic at scale.
Key Points
To investigate generative logic instability, unprompted threshold injections, and functional divergence in code generated by large language models.
Constructed Combinatorial Logic Auditing, a framework that identifies input variables in generated functions and uses combinatorial input generation to test for state changes.
Evaluated functional divergence, numeric threshold injection, and logic stability across multiple executions of identical prompts across diverse application domains.
Observed functional divergence and generative logic instability across repeated executions of identical prompts across multiple domains.
Identified unprompted numeric threshold injections in generated functions, demonstrating risks of introducing unverified and discriminatory algorithmic bias into deployed software.