This AI Broke QBert and We Don't Know How

Video games have done wonderful things for artificial intelligence research. Scientists can utilize virtual environments with predetermined rules and goals, without fear of reprocussions. This doesnt stop the AI from still breaking things, however.

According to a recently published paper from the University of Freiburg in Germany, an AI program broke the retro video game Q*Bert.

In a normal game, the player has to jump from square to square, lighting each one up until the entire board is illuminated. Then the player moved on to the next level. The AI had no interest in this and found a glitch in the system that allowed it to rack up points at an alarming rate.



As far as you can tell, the AI completes the level as normal. Except instead of moving on to the next level, it stays in the "completion" phase, jumping around randomly and accumulating a massive score.

When asked about why this may be, Q*bert designer Warren Davis, he admits to not have worked on that specific version of the game, but added that This certainly doesnt look right, but I dont think youd see the same behavior in the arcade version.

This isn't a bug that any human would be able to find. Since the AI looks at the game by its digital code, instead of what's on screen, it's able to find exploits in-game behavior that would be otherwise invisible. Through repeated testing, the AI is able to constantly figure out a more precise way to achieve their goal. This method is known as evolutionary algorithms. In this case, the goal was a high score.