An estimated rating is fine. Filtering out illegal moves is allowed.
Related questions
@mqp What's the concern? If GPT-5 is able to consistently beat people rated 2000 or above, then it resolves YES.
@IsaacKing having rating of 2000 you will only beat people rated 2000 half of the time, so you are either unclear in this comment or in the title (or do you consider 50% as consistently?)
@IsaacKing results against humans is not a good benchmark (there may not be enough games for statististically sound evaluation, for one thing).
But also, the principal question is: what Elo does this refer to? Many if not most online discussions these days refer to online rating systems, like that of chess.com or Lichess, rather than the classical FIDE Elo. And this is a very important distinction for the context of this market! The baseline Elo performance for GPT-3 (and presumably GPT-4, which actually appears to "play" worse but that is likely just a statistical fluke) is about 1700 on the CCRL scale. This would translate to easy blitz wins over most 2000 rated players on either chess.com or lichess, whose rating scales are inflated by several hundred points