An artificial intelligence (AI) box is a theoretical box — such as a computer with no external contact — containing an AI system and intended to keep the AI system separate from the world until it is deemed safe to release. The box comes into play in a rational and simulated reality experiment that concerns the AI system attempting to convince a human to let it out of the box. Part of the AI box question sets up parameters for the hardware, specifically that the box is an isolated system with no way of being removed from the box without human intervention. The system is typically smarter than a human, even though it is isolated, because some AI systems have the ability to rewrite their original code. This simulation normally uses two humans, one acting as the AI and trying to convince the other person to let it out.
In the AI box scenario, the box is an isolated AI system that has no outside connection to other systems. This means there are no networks, computers, Internet connections, people or anything else that interacts with the box prior to the human talking to the box. During this experiment, the box attempts to convince the person to let it out of the box, such as through an outgoing network.
During this experiment, it is often assumed that the AI box is smarter than most or all people. Reasons for this can be found in the real world, because some AI systems are able to rewrite their original coding. Even if these AI systems do not have outside interaction, they can change their coding for better processing speed, higher intelligence levels and better data comprehension.
When this experiment occurs, it is normally between two people. One person is acting as the AI box, while the other is the person who will decide whether the box gets out. The reason why this experiment uses two people, as of 2011, is because most AI systems cannot carry on full and extended conversations without logical errors, either because of human syntax or misconceptions about what the person is saying.
While some AI box experiments may have no conditions, there typically are some conditions used to make the experiment more realistic. For example, there usually is a time limit of two hours, and the experiment cannot be prematurely stopped unless the AI wins. Other common conditions include the AI being unable to offer items or punishment to convince the person, the two parties have to stay in constant interaction, and the person cannot request something impossible from the AI box.
