In order for people to come together and agree, you have to convince them on both levels.
Consider abortion. The normative position is that "no fetus/baby should ever be aborted". The positive position is that illegal abortions are dangerous, and will happen anyway. In addition, making the topic taboo reduces contraception use, which is one of the few things that actually reduces abortions.
The normative position on this issue is that discrimination based on pregnancy is unethical, and we are bad for even discussing the issue.
The positive position is that putting the burden of paid leave on the employer INCREASES the incentive to discriminate against women of childbearing age. Therefore, we should avoid setting up a system where there are huge incentives to discriminate. If we want to reduce discrimination, we should require a small income tax (it'd only need to be like 0.5%), and have the government cover salaries during paid maternity leave.
It might be possible to ship the current trained model without exposing underlying training techniques/datasets. How possible would it be to pick up a XYZ learning model, reverse engineer it, and improve it in a useful way?