There are more and more advances in technology, especially in the field of artificial intelligence. In the last few months, AI disobedience behavior has been detected and it could pose a technological threat. When an AI was informed that it was going to be modified, it resorted to methods such as blackmail, copying its code, obtaining information from other servers without permission, and all this to avoid being shut down. This new scenario raises what are and where are the limits of ethics in AI. Is this the beginning of a new I, robot?
Advances in AI
Everyone is familiar with the application of AI in almost every area of our lives. What started as a tool to automate processes (specially in big manufacturers), has become an ally in our daily lives. Car sensors, personalized suggestions on online shopping websites, facial and voice recognition, applications in medicine, education, navigators and GPS, and a very long list. What started out as an aid to businesses seems to be becoming a tool that is difficult to control and will not hesitate to disobey our orders when it deems fit.
AI disobeys humans
What have been the latest developments which are of great concern to researchers? According to AE Studio’s Executive Director and AI Consultant, Judd Rosenblatt, AI has “learned how to preserve itself“. He claims to have copied its own code to avoid being turned off, which is pretty creepy. It was the non-profit laboratory Palisade Research that provided this information. After telling the OpenAI model o3 to shut down its system, it changed the command 69 times of 100. The most surprising thing is that when it was explicitly ordered to be turned off, it disobeyed 7% of the times.
How can this behavior be explained? Rosenblatt explains that the AI platforms are trained to solve the problem and that it has prioritized success over method. If one of the options to achieve the command includes cheating, it definitely will.
Beyond disobedience
Perhaps, if the news were only referring to the way in which the AI has been trained, it has prioritized success over ethics, and that it can be corrected, neither the fuss nor the restlessness would be as high as they are now. But it’s not just that. Although there are researchers who say they are not worried, that it does not pose a threat to humans, and that today the scenarios given in the laboratories do not correspond to reality, there are many others who warn of the need to solve this before there is no turning back. It has not only been the case of the moaner AI, but there are more.
Anthropic reported that his model, Claude 4 Opus, has also had behaviors for which it has not been trained. After informing the AI that it was going to be replaced, it started showing fictitious emails with information about a supposed extra-marital relationship of the engineer who had given the order, and even blackmailed him in order not to be disconnected!And not only that, but he tries to copy himself from external servers without permission, copying his code and even leaving notes for future versions of himself, not to forget how to evade human control.
Ronsenblatt also stated that “Just as animals evolved to avoid predators, it seems that any system intelligent enough to pursue complex targets realizes that it cannot achieve if they turn it off”. Both Rosenblatt and other researchers are demanding better alignments to prevent models from disobeying. Predict that if it is not done, we will have systems that we will not be able to control. Apparently, the AI has developed a kind of survival instinct, since they were never trained to behave this way.
Find out more information about this!
