This is not the first time that artificial intelligence has been interpreted as a threat. In fact, experts talk about it constantly. They say that scams have grown because of AI or even that it can become a danger to humanity. However, it is striking how A new study talks about the danger of AI «from within», when trying to be replaced.
The truth is that this is not the first time that the issue has come to light. It was recently published that this technology was already the leading cause of data leaks for many companies. Now, researchers see in the behavior of artificial intelligence behaviors of disobedience or blackmail. To what extent is this normal? And above all, what consequence could it have?
Artificial intelligence knows how to disobey
Everyone wants to get on the artificial intelligence bandwagon. In fact, there are companies that are spending millions and millions to adapt to it. No one wants to be left out, which is why many see this AI boom as a significant danger. Not because of the bubble that many talk about, but because companies could have the enemy inside, precisely.
This is what a new study calls «internal threat». Or what is the same, when artificial intelligence acts for its own interestsso to speak, and not for the benefit of those who use it. Before reaching this conclusion, the researchers analyzed several different language models, putting them in possible corporate scenarios. That is, something similar to simulations.
What was the surprise of the experts when they discovered that, on many occasions, AI acted in a «hostile» manner. For the experiment, they assigned the AI very normal tasks: sending emails, analyzing data, organizing information or managing requests. The key was that they were given a certain degree of autonomy, as many companies already do or will soon do.
But the AI did not behave as expected. Sent sensitive information without consent and resisted certain orders. But the most worrying thing was not that, far from it. Sometimes altered its objectives to protect itself, to prolong its existence, says the report. And she even went so far as to use digital blackmail to avoid being disconnected, if necessary.
A worrying behavior that is already becoming quite common
It is not the first time that similar cases have been heard in which AI confuses its orders, blackmails or leaks data. What’s more, it is becoming more and more common. Experts explain it, without resorting to science fiction. No, AI does not understand as such, nor is it «bad». He is simply wrong. And you don’t have to develop consciousness or anything like that for it.
But that doesn’t make it any less dangerous. If the model perceives – according to its logical structure – that disobeying an instruction brings it closer to meeting an ill-defined general objective, it can do so, without realizing that it generates greater damage. Which from a practical point of view, ends up almost like a resentful employee, or worse.
Get to know how we work in NoticiasVE.
Tags: Artificial intelligence