Well for one, if someone legitimately did a TAS, they wouldn't be submitting it here, they'd be submitting it to http://tasvideos.org/
If someone is trying to cheat and upload a run here when it's actually a TAS, then it will almost always be caught by a moderator because it's often very easy to see when a run has movement or inputs or strats that no human can pull off. Basically, the more you run a game and the more you see runs of a game, the easier it is to tell when it's a human playing or not.
In addition to the above point, a 'TAS' doesn't necessarily have to be a superplay or the fastest possible completion of the game. This question can be taken as "How can you be sure a runner hasn't recorded a run in an emulator with the assistance of tools, and submitted the resulting recording?" The older equivalent of this would be spliced runs, and those have traditionally been detected through analysis of the frame data and audio waveforms, detecting inconsistencies.
It's true that if someone was 'clever' enough, there would be no way to detect a tool-assisted RTA run. If somebody were to record a movie file in an emulator that wasn't perfect, but was still a 'good run', albeit still full of mistakes that made the movie look very human, it could well be indistinguishable from an RTA run. There are often telltale signs though, and instances of these sorts of runs going totally undetected are increasingly rare.