I think it becomes somewhat more defensible when considering that the alternative was operatiny Google's policy (before the advent of competition) of "these models would bring unknown dangers in the hands of the public, we shouldn't release them until we better understand the implications" (or perhaps more selfishly "these effectively nullify all our detectors of generated text, if released they would instantly lose us the war on SEO").

(recall that OpenAI thought GPT-2 was too powerful to release for approximately tantamount reasons)