One interesting angle here is the “epistemic asymmetry” point.
Even if providers comply with documentation requirements under the EU AI Act, downstream deployers still can’t realistically audit a model’s behavior at the level of training data or causal reasoning.
Curious how ML practitioners here think about this.
Is this asymmetry something that can realistically be reduced with interpretability / mechanistic transparency research, or is it fundamentally structural for large-scale models?
Even if providers comply with documentation requirements under the EU AI Act, downstream deployers still can’t realistically audit a model’s behavior at the level of training data or causal reasoning.
Curious how ML practitioners here think about this.
Is this asymmetry something that can realistically be reduced with interpretability / mechanistic transparency research, or is it fundamentally structural for large-scale models?
reply