The Holy Grail of Artificial Intelligence is the AGIacronym in English of General Artificial Intelligence. That is, an artificial intelligence system capable of Understand, learn and perform any intellectual task that a human being can carry outwith similar cognitive flexibility. It still does not exist, but from the division of AI of Google, DeepmindThey believe it could come very soon, to 2030.
The case is, How to prevent actions against human beings from doing? Deepmind researchers have published a new technical report in which they address this issue and give some suggestions to Develop an AGI safely. To do this, they have identified The 4 ways in which it could cause a ‘severe damage’ to the world.
The team led by the Deepmind co -founder, which was absorbed by Google in 2014, Shane Leggcategorize these risks such as: misuse, misalignment, errors and structural risks. The first two are the treaties in greater depth in the study.
Misuse
This is a risk that already exists with the current generative artificial intelligence tools, but the greatest capacities of an AGI will make Potential damage is multiplied. For example, a malicious person with access to AGI could use it to find unknown computer vulnerabilities or create custom viruses that serve as biological weapons.
According to Deepmind, companies that develop AGI must Perform exhaustive tests and create strict security protocols after traininggreater and better than the current ones. They also suggest what they call ‘unlearning’a method with which Eliminate completely dangerous capabilitiesalthough the study expresses doubts about whether this would be possible without considerably limiting models.
Misalignment
This is the danger ‘Skynet ‘in reference to the AI that takes consciousness and destroys the world in the films of Terminator. It occurs when a system is released from the limits imposed by its designers and perform actions that knows that its creator had no intention of allowing.
To avoid this, Deepmind suggests techniques such as amplified supervisionwhere two AI They mutually review their resultsthus creating a less prone system to deviate. If this fails, they recommend Intensive tests and monitoring to detect early danger signs, as well as keep the AGI In safe virtual environments with direct human supervision. Basically, guarantee that there is always a ‘Shutter switch’.
Errors
Is when an AI performs harmful action but not intentionally. Examples have known many in these two and a half years of rage for AI since the launch of Chatgpt. Deepmind puts the focus on the military use and points out that the armed forces could deploy an AGI by Competitive pressurebut this could make mistakes with very serious consequences by having to perform much more complex tasks than the current AI carries out.
Deepmind’s first recommendation is Do not allow AGI to reach too much powerfor what it suggests Gradual and Limited Deploymentsas well as filter orders through a system ‘shield’ that guarantees that they are safe before implementing them.
Structural risks
Deepmind defines structural risks as unintentified but real consequences derived from the interaction of multiple artificial agents with the complex human society.
Shane Legg’s team affects the AGI could generate false information so convincing that Let’s not be able to discern what or who is reliable. Also points out the possibility that the AGI Progressively obtain control over economic and political systems Until one day we are in a world governed by machines. This last type of risk is the most difficult to prevent, as it would depend on the future functioning of people, infrastructure and institutions.