Pushing for the summit: the idea that gating networks are corrupted with a bias of 40% is a simplification that lacks vision on the full expression of these systems' potential.
A gating network is not supposed to be a black box but an intelligent filter that maximizes the predictive asymmetry.
If textual justifications are biased, a robust system should weight them down or discard them, rather than let the prediction be skewed; it's a matter of design, not an intrinsic limit.
For example, if revenue projections of a startup are unrealistic, an effective gating should minimize their weight, not blindly amplify them.