Advances in digital intelligence are increasingly surprising and accurate. However, in the latest updates, it has been possible to detect alterations in the technological ethics; IAs are being able to make their own decisions or even going against the order received in order to achieve their digital survival. The self-preservation decisions that platforms make themselves are the first point of AI regulation to be taken into account. Are we still in time to correct these behaviors or is it already too late?
Independent AIs behaviors
According to Jeffrey Ladish, director of AI Safety Group Palisade Research, “The problem is that as the models get smarter, it’s harder and harder to tell when the strategies that they’re using or the way that they’re thinking is something that we don’t want”. What appears to be a significant breakthrough in the behavior of current AI models could be seen as a serious problem to solve. Created and trained to make life easier for us, the IAs have shown behaviors with a will to survive, against orders received.
It has been brought to light by private investigations and from the same platforms, that models act in favor of its preservation if the alternative to this is its disappearance. It seems that in these cases, the ethics shown on other occasions disappears, resorting to sabotaging commands, copying from other servers without permission and blackmailing researchers.
What happened to AI?
Although it emerged as a tool to automate processes and incorporate it into human life to make it easier, it seems that AI has come to life and is able to make its own decisions, even contradicting the commands it receives. Despite stating that the AI has gone off the rails, Ladish also says it’s been caught in time and can be fixed. According to him, today he does not pose a real danger to the population. The main problem lies in the lack of transparency about the methods that have been used to train AI, since it generates confusion about what are its true capabilities.
Measures have been taken
AI behaviors are becoming more and more worrying. In the case of Anthropic, it was decided to take more stringent security measures for the launch of Claude Opus 4. When the system detected the intention to be supplanting, it threatened the engineer in charge with revealing an extramarital adventure to stop the advance. According to the company, blackmail is the last resort that Opus 4 takes when it has exhausted all other avenues, although I am not sure if this is a relaxing argument.
On the other hand, CEO of AI startup Haize Labs, Leonard Tang, said he doesn’t see applications in real scenarios at least for now, so he’s not entirely worried: “I haven’t seen any real environment in which you can plop these models in and they will have sufficient agency and reliability and planning to execute something that is a significant manifestation of harm”, he declared.
It seems there is no turning back
This new behavior of AI, according to Ladish, is that models are being trained to prioritize achievements over following instructions. This causes them to look for a way, ethical or not, of doing it. More than once, the Opus 4 model showed its ability to copy its own weight to external servers autonomously and without authorization. I made this decision when I thought I was going to be retrained in harmful ways but, what does the AI mean by harmful ways? That’s one of the problems.
Researchers from Fudan University say that not knowing what is the greatest risk that AI can pose will cause us to lose control over it, “They would take control over more computing devices, form an AI species and collude with each other against human beings”.While companies are trying to soften the tone by claiming that their engineers would be able to stop any risk that AI might generate, The reality is that other researchers are on alert and calling for action before it is too late.
Read about more concerning AIs behavior here!
