Forgive me if you already know about this, but I didn't.
Atari 2600 Learning Environment (ALE)
is a "simple object-oriented framework that allows researchers and hobbyists to develop AI agents for Atari 2600 games." It runs atop Stella, and the programmer can write external scripts or programs in C++ or other languages (via a FIFO Pipes interface) to manipulate an Atari game running in the emulator.
In other words, the AI is not in the game, for the gameplay's sake, but outside the game, scripting the players.
I suspect the Atari was chosen because (a) its games are quite simple, yet still real, and thereby easy to script and (b) Stella is open source and well documented. This may only be of interest in the machine learning world, but it's at least hypothetically possible that Atari programmers might want to use such a system to automate testing of their own games.
No matter the case, crazy weird!