@ericyocamShipped: Delphi of GANs — AI Bias Detection & Fairness Evaluation
An open-source AI research proof of concept investigating how ensembles of Generative Adversarial Networks (GANs) can help identify and compare potential bias in pretrained language models.
Inspired by the Delphi method of combining specialized expert assessments, the framework uses separate GAN-based evaluators for gender, racial, and age-related bias. Each GAN is trained on a corresponding bias dataset, with discriminator outputs used to generate experimental bias scores.
The framework evaluates transformer-based language models, including BERT and DistilBERT, comparing their responses across multiple bias dimensions. It combines neural network predictions, GAN-based scoring, and configurable interpretation thresholds to investigate differences in model behavior.
Developed as a reproducible Python/Jupyter research experiment, the project explores applications in responsible AI, algorithmic fairness, model evaluation, and AI governance.
This is an exploratory bias-assessment methodology rather than a validated fairness certification system. Further benchmarking against established fairness metrics and independently labeled datasets would be required before operational deployment.
1 comment
Calling Delphi of GANs a proof of concept rather than a fairness certification is the kind of honesty compliance teams trust. Is your next step more academic testing, or putting it in front of a company reviewing its own AI for bias?