That's because humans consider the future empire as a whole.
The AI doesn't consider future cities it's going to build at all.
Even for existing cities, it's limited, it treats a tile 2 tiles away (but currently outside its cultural boundary) from one of its existing cities as if the tile were nowhere near it.
The second statement is true, but not for the reason you described in your last statement. The AI actually does consider future cultural growth for its cities: even though yields and resources have base fertility values, a plot's overall fertility value has distance multipliers of 6, 3, 2, 1, and 1 for plots 1, 2, 3, 4, and 5 tiles away from the given plot. As a result, even if a certain plot would have a better overall fertility score 2 tiles away, the distance multipliers dictate that plots with a high fertility score in the first and second ring have a higher actual priority.
No, the most likely reason that the AI founded there instead of 2 tiles away is a combination of two of the biggest flaws in the AI settle plot logic: 1) the yield multipliers are incredibly out of whack (the AI considers food to have a value of 15 and hammers to have a value of 3), and 2) distance fertility multipliers are the same for all yields, even though food tiles located in the 3rd ring are most likely less useful than hammer tiles located in the 3rd ring (by the time the city expands to the 3rd ring, it'll most likely need hammers more than it needs food).
Essentially, it's not just the luxuries that drew the AI to that particular spot: it's mainly the 3rd ring floodplains, 3rd ring wheat floodplains, and the 3rd ring oasis. Food has a base value of 15, while hammers have a base value of 3, so a 3-food tile in the 3rd ring has a value of 90 (floodplains), while a 1-food 2-hammer tile in the 2nd ring has a value of 63 (iron grassland). These values are still multiplied by the AI's flavor values, but in order for the AI to consider the second tile to have a higher fertility than the first, the ratio of its growth flavor to production flavor must be 2:5 or higher. That's basically "never".
Things do get a bit weird with resources, though, which is why my example of comparing 3rd ring floodplains to 2nd ring iron grassland is a kind of a bad example. With resources, the AI will calculate the plot's yields as if the resource's improvement was already built on the plot, so for iron grasslands, the AI would see it as 1-food 3-hammer, which would bring the 2nd ring tile's fertility value to 72, but it would also see wheat floodplains as 7-food, so a 3rd ring wheat floodplains would have a value of 315. If the resource can be traded, the AI also adds in 10*(resource quantity), or 30*(resource quantity) if it doesn't have the resource. If the resource is a luxury, the AI also scores the tile for happiness and adds that into plot's fertility value as well.
Oh yeah, and the AI doesn't factor cultural influence speed into account, so if a couple of valuable tiles are separated from the settle plot by a desert, peaks, jungle hills, or other tiles that are harder for culture to spread through, the AI won't realize that it wouldn't get those valuable tiles just by settling at the plot and letting passive culture spread do its work.
The AI also sort of ignores peaks, ie. if it has buildings or wonders it may want to build that require nearby peaks, the AI will not take this into account when looking for plots to settle. It only takes peaks into account when calculating a plot's "strategic value", which is a highly flawed calculation: basically, the AI looks for hills in the city's 1st tile radius and peaks in its 2nd tile radius and increases the plot's score for each one it finds. This is why you'll see the AI settling near mountain ranges on tiles that are actually fairly poor settling spots: it saw all the peaks in the future city's 2nd tile radius and said to itself, "Ho boy, this place sure has enough strategic value to counteract how bad it is for food and hammers!"
Spain AI also does not recognize that it gets double the yields from Natural Wonders. Inca AI does not realize that it can build terrace farms on hills and would also have free maintenance from hills. Celts AI overvalues forests that would grant faith to a future city, ie. it values faith from unimproved forests to be at least 100 times the normal value of faith.