Google isn't a legislator but its policies do affect many people. For example, in the context of laws, see: https://en.wikipedia.org/wiki/Ex_post_facto_law
"In the United States, Congress is prohibited from passing ex post facto laws by clause 3 of Article I, Section 9 of the United States Constitution. The states are prohibited from passing ex post facto laws by clause 1 of Article I, Section 10. [...] Thomas Jefferson described them as "equally unjust in civil as in criminal cases"."
(To be fair, the article cites exceptions)
Surely you would only be applying rules retroactively if you were engaged in reopening finished arbitration?
Cancelling a current process that is still in open session, on the basis that you have already publicly agreed that employees being forced into that process is unfair, is not being retrospective.
If the announcement set Day 0 based on the date of the claimed harassment, that would be extremely suspicious, and I think the most natural reading would be some specific issue that's been resolved, but could produce exceptional damages. With a date-of-claim threshold, that seems much less likely; any kind of serial misbehavior or other mishandling that Google expects to avoid in the future could still create liability via claims not yet filed.
A few legally-inclined people elsewhere in the comments have pointed out that "adding the option to leave arbitration" behaves differently than "not forcing arbitration to begin with". Most non-forced arbitration is still binding, which means that neither party can start arbitration and then pursue a lawsuit after seeing the result. Opening a one-time option to leave ongoing arbitration would be much more similar to entering non-binding arbitration, a move lawyers frequently counsel against.
The proceedings of non-binding arbitration are almost never admissible in subsequent court cases, and the process lacks formal discovery. As a result, there are unpleasant tactics like entering non-binding arbitration, obtaining useful information (e.g. someone's schedule, or information of who knew about a behavior), and then pursuing a lawsuit with added or altered claims based on that information.
It seems weird to me that they wouldn't retroactively apply the policy unless there's something to hide within whatever currently claims they have. You would think that the exposure would be similar, unless they strategically plan on doing something different with the future claims, like either restricting or dealing with them much faster so that they don't reach this stage, etc.