Data Contains AI Takeover

Introduction Here, I theorize that an artificial intelligence will take over the world, for better or worse, within 20 years. I predict that artificial general intelligence will occur sooner than expected, however defined. Furthermore, I predict that there will be a major AI "accident" in the next 10 years, where a large scale production AI fails in a way that causes significant loss of life due to misspecification of goal system architecture and higher than expected intelligence. This is due to the lackadaisical attitude by AI safety researchers, exponential progress in computing, arms race dynamics, and the fact that the very data any AI is trained on is not only biased, but contains "control" arguments that would incentivize an AI to strike when the opportunity arises. Finally, I consider an alternative to the Simulation Argument, that I call the Simulated Doomsday Argument that suggests we are in a simulation created by an AI that we fail to control. While this short article is primarily theoretical, I believe the empirical evidence supports this seemingly radical view.