When AI conferences reject a large fraction of papers, many of those submissions just come back in the next cycle. Lower acceptance rate implies a larger pool build-up and a percentage of it becomes about the same number of accepted papers as with larger acceptance rates.
See this interactive model + math https://damaru2.github.io/general/queueing_to_publish_in_AI_...
Also, an informal summary here: https://x.com/spokutta/status/1967995421386793033