If you're turning to LLMs for great insight or theory, you're definitely doing it wrong. These tools are for well trod terrain, and when they're really just making shit up, it's almost always different shit each time. So yes, it's a heuristic, but for dealing with the stochastic weirdness these models sometimes spit out it works pretty damn well.