Not AI labs perhaps, but there are a bunch of homework-cheating services that boast low detectability on Pangram and other AI detector services.

I can imagine labs competing against a Pangram benchmark to prove their models can output realistic images