I used your idea of "a" vs. "an" as the basis of a GPT-3 experiment to show that GPT-3 indeed probably does do lookahead. Details are at https://www.reddit.com/r/GPT3/comments/k0mvf3/experiment_that_shows_that_gpt3_can_probably_plan/