We used to ridicule the Soviets for having whole swathes of science off limits because they knew the results would be politically unacceptable. Or else their scientists would start with the conclusion and work backwards with faulty premises. We all had a good laugh and saw they were doomed because nobody was allowed to tell the truth. I never ever in a million years thought it could happen here, but here we are.

Hagbard Celine's second law: accurate communication is only possible in a non-punishing situation.

It isn't being politically correct, it is the fact that AI's are programmed/trained with our biases so they are making the same mistakes that we are.

So for training an AI, to read resumes, it references what people in the past had allowed and denied, to try to mimic it. So looking at its references it had learned to check the Name of the Applicant and weigh their name as a strong aspect to hire or not. Often giving preferences towards White Male Common Names.

The AI itself isn't racist or sexist (It is just a computer, it really doesn't care) but it had picked up the biases when it was learning how to process the information. So such issues will need to be caught and corrected in either the training, or a hard coded cut off.

Americans have been using the Term Politically Correct, as an excuse to allow them to be cruel and unfair to people, and to excuse their beliefs that are just immoral and wrong, as just some "Other" pushing them into their agenda.
While creating an AI, we need to work to make sure it will be more moral than we are. It can be politically incorrect, because an AI routine can process data and point out that policy of political party that is in power is not productive or helpful. As the warning about being politically correct, was more towards people being afraid to talk against the government. (China you have to be politically correct and not bad mouth the communist government). In the United States you can feel free to disagree with the Biden Administration or the Democratic majority branches of government, with out such governments going against you.

That's not true, there are plenty of techniques that can be applied to personal data that allow it to be used for AI training, like differential privacy. Several EU member states (like Germany) also have research exceptions, so research can still be carried out without having to sort out the audit and compliance issues faced in production. Bigger challenges faced from a regulatory point of view are things like the right to object to AI-based decision-making, and being able to retain audit continuity across training cycles, especially in unsupervised learning situations. This is especially a big issue when processing biometric and health data, which are not covered by the GDPR and for which no free-flow mechanism exists for cross-border data transfers. There's still the issue of whether one can take a model trained on sensitive data in one country and use it in another - it's not yet legally clear whether the model itself constitutes derived data that would be subject to the same set of restrictions or not, and the courts still have to work this out. Perhaps the biggest outstanding issue is liability for automated decision making - e.g. do you blame the model, the person providing the inputs, or the person taking the action based on the suggestion of the model? The original product liability directive never had these sorts of cases in mind and is also being reformed to better address these new kinds of usage scenarios.

Yes, you could ignore all of these and just throw random AI models into production, but I'd be hesitant to call that progress. Others will eventually face these issues as well in production, regardless of whether there is a country-specific regulatory framework in place that forces one to think about these problems in advance or not.