If you "execute" the program manually, you may still not understand Chinese, but the program still does. The room, including you and the program, is functionally a single entity that speaks Chinese. The human is an implementation detail as much as the grey matter is an implementation detail in a Chinese speaking human.
Searle's argument that the machine does not actually possess "understanding" is simply special pleading. Unless he can define "understanding" in a meaningful way, the difference between an "understanding" entity and a "simulating" entity is non-existent.
These arguments all boil down to unscientific armchair philosophy. I have no idea why people seem to accept them as meaningful other than that the people who came up with them or wrote about them are otherwise highly regarded -- but that's literally an argument from authority.