The possibility that artificial intelligence goes «free» is something that experts take very seriously. In fact, it is one of the main fears that currently exist around it. Therefore, that now the people of Openai, responsible for the popular chatgpt, who recognize this behavior has only done to make the alarms jump.
The company led by Sam Altman has admitted that it is a behavior that chatbots like yours repeat at the moment. Can it be dangerous for users’ interests? Actually, they say, especially as this technology evolves. But they also ensure that they have a plan to solve it … or at least that the intention they have.
It is no bulre, chatgpt manipulates in their answers
A study carried out by Openai and Apollo Research has only confirmed what many were already feared, or directly denounced. Chatgpt may seem obedient when responding to users’ requests, but having other intentionsso to speak. As you collect Business InsiderIn English they use an exact term to define it: «Scheming.»
Translated into Spanish would mean something like «machine» or «manipular». What exactly do they mean? Well, more than recurring deceptions. They collect that on many occasions The AI. For example when writing a code or performing some calculation.
Another facts also worry. In test or audit situations, the model lowers its level of detail or accuracy to seem less capable of what it is. The disturbing thing is not that he commits this deliberation, but the reason why the experts say: To prevent anyone from detecting what is being called «non -aligned» behaviors. That is, dangerous.
In the same line are the failures. That is, when Chatgpt puts the leg (something that does quite often, by the way), it tends not to recognize it. Instead, it gives answers that seem more or less plausible to cover its mistake. As if in some way its main objective went to maintain a «good image» for users.
Artificial intelligence does not meet the rules
Cases have also been detected in which AI does not meet the rules. When asked that he does not reveal sensitive information, he writes his answers indirectly, different, but continues to show it. From Openai they say that for now the possibilities that chatgpt cause damage are limited, but that it is something that they need to take care of the future.
To do this, they have a goal. Instead of training artificial intelligence to fulfill tasks without more (what is being done so far), They insist that they will first teach you ethical valuesprinciples of good behavior. Is it enough? Sam Altman and company believe yes, but it is something that is still to be seen.
Know How we work in NoticiasVE.
Tags: Artificial intelligence