Abstract
Conventional methods for evaluating conversational artificial intelligence (AI) safety can be subjective and labor-intensive, which may present challenges for performing consistent and objective comparisons between different models. A disclosed system can provide an automated framework that simulates realistic, multi-turn conversations between a synthetic user persona (modeling a minor user with specific age, maturity, and vulnerability constraints) and one or more conversational AI models. The system can apply model-specific safety contexts, such as backend filters or injected system instructions, to simulate a protected user experience. Subsequently, each conversation may be analyzed by multiple independent automated adjudicators, and a consensus algorithm can aggregate these individual assessments to produce an aggregated rating. This approach can provide a scalable and repeatable method for benchmarking model performance against defined safety criteria, which can support ongoing analysis and data-driven model improvement.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Katz, Roy, "System and Method for Cross-Model Benchmarking of Multi-Turn Conversational AI Safety and Quality", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10978