Treat the drones as an extension of the humans, and require low latency interactions.
Some interactions are easy. Modern aerial drones don't really do anything difficult compared to what your drones will probably have to do. I don't believe we have any drones dogfighting yet. As such, their latency requirements are minimal.
If communications are uni-directional, latency probably wont matter. You'll structure your AIs such that that latency is hard to distinguish. However, if communications are bi-directional, it gets interesting.
Consider a system that is so complicated that you can't just "command" it to do something. You have to observe it to make sure it followed your command. K9 soldiers are an excellent example of this. These dogs are some of the best trained dogs on the planet, but we don't rely on them to follow the command. The trainer watches the K9, observing its demeanor, making sure that it looks like the K9 got the message properly.
Watching body-language with a high latency connection is atrociously difficult. So many of the cues we look for are interactions, where you do something and see a response from the other individual. If the effectiveness of the teenager's control over these drone is entirely dependent on reading the body language of the drone, they'd need to be close enough to have the low latency command it takes.