New CTLF Logic Framework Addresses Bias in Generative AI Outputs

A novel branching-time logic named CTLF has been developed to systematically evaluate biases in generative AI systems. This framework was created to tackle the issue of bias amplification found in training datasets. Utilizing a counting worlds semantics, CTLF represents each potential output at a given generation step as a distinct world. It features modal operators that allow for the assessment of whether a sequence of outputs conforms to a desired probability distribution concerning protected attributes. Furthermore, it can estimate the probability of maintaining fairness as new outputs are produced and identify how many outputs must be discarded to achieve fairness. The paper, which demonstrates CTLF through a simplified example of biased image generation, was published on arXiv with the identifier 2604.19431v1.