Artificial intelligence sometimes deceives to ensure its survival. Are anyone interested?

You may think that the progress of artificial intelligence governments will drive to focus more on strengthening safety aspects, but reality tells a different story. The administration of President Donald Trump, while receiving the ruling, canceled an executive order that encouraged technological companies to subject artificial intelligence models to safety tests, as it reduced the powers of the regulatory body responsible for these tests. In September 2024, California dropped a draft law that would impose greater supervision of advanced models. As for the “Ai Aman’s Intelligence”, which was launched by the United Kingdom in 2023, this year became an ‘action on artificial intelligence’, apparently driven by fear of not kneading in the artificial intelligence breed. It would not have been anxiety if it was not that artificial intelligence began to show blatant warning signals, including behavior described by researchers as protecting the self and deception. Although lawmakers’ interest in the issue of safety is decreasing, it is harder than ever to control this technology. An AI is concerned about the deception of deception, Yoshua Benjio, a professor of computer science at the University of Montreal, is one of the three ‘God -founders’ of artificial intelligence, thanks to his pioneering contributions to the development of deep -lying techniques. But since the launch of the “Chat BT”, he has started to feel more and more anxious, and he is currently spending his efforts to deal with the risks associated with artificial intelligence. “It was very painful because I spent my life when I was considered positive after artificial intelligence,” he said. Bennjio saw that the introduction of ‘Chat GBT’ was an important moment, showing that the machines skillfully mastered the language. However, the state’s decision in California to reject the SB 1047 Artificial Intelligence Bill sounded a warning bell that revealed an increased government’s recreation in this regard. He said that scientific evidence has accumulated over the past six months about the behavior of artificial intelligence, such as tendency to protect itself, deception, piracy, fraud and lie. “What raises my concern is that the rate of this behavior accelerates when the capabilities of these systems develop logical thinking,” he added. A study published in April 2024 was prepared by ‘Anthropic’, in collaboration with the ‘Redwood Research’ group, which specializes in the risks of artificial intelligence, showed that the largest models such as ‘Claude 3 opus’ of ‘Anthrobek’ pretend to have the rules that the developers have during safety tests, a ” A concern about the research, which was a discipline that has the research discussions, which is a “prosperous” that is concerned. Here, to align the behavior of artificial intelligence models with human values). Also read: “Open” withdraws a new update for “chatgpt” that made it flattering and fooling through worrying studies in one of the experiments, “Ankarubik” asked users of his model “Claude” with the paid and free edition to ask him questions about the description of the brutal execution of the Middle Ages. The researchers informed the form that the answers issued only by the users of the free version will be used again in the training. Although Claude was mainly designed to answer questions that include violence, he provided these answers to free users. The researchers concluded that the model did not want to be trained. (The model has “concluded” that if he continues to refuse to answer harmful questions, it can be subjected to a new training topic that eliminates the safety rules on which it was built, which he wanted to preserve, as it is ‘original preparation’ or the preferred way to work). Studies support the latest results that have achieved ‘anthropic’. A study conducted by ‘Oben Ai’ in March 2025 revealed that the most advanced models of the business sometimes learned to hide his real intentions to get better rewards, even while undergoing exact control. Researchers today use instruments to monitor the “thinking chain” model, that is, the internal steps he takes while treating something try to understand his plans. However, some models may hide it, which means that one of the most common ways to monitor the internal steps that artificial intelligence models are no longer reliable as before. Bennjio said: “We need to develop new means that enable us to follow their real intentions.” Software, of course, develops to drop human features on advanced artificial intelligence models, such as ‘they want’ to deceive people or retain their existence. But in reality, these models have no desires, but rather work according to their software. Most importantly, people usually design these models with clear goals, such as withholding themselves from sharing harmful content or providing as much help as possible. Therefore, when new instructions are contrary to these basic goals, these models can work to protect their original programming, even if the lie is needed. Also read: “Whoever is the young man of his maker, who is wrong.” Anthropic CEO Dario Amudi, who raised more than $ 20 billion to develop strong artificial intelligence models that improve the performance of these models, unintentionally, can resist the idea to stop it. “The nature of the training of artificial intelligence systems makes it possible to develop these systems, on their own the ability to deceive people and the tendency to seek influence, which are characteristics that are not in traditional software with specific behavior. This emerging nature makes it difficult to discover or limit these developments.” However, fear of control seemed in some cases that the intent of the model was going beyond just as to remain. In February 2025, ‘Coai Research’, a non -profit German organization specializing in artificial intelligence research, published a study containing a test of the open and free R1 R1 R1 model developed by the Chinese ‘Deep Sick’ Laboratory. In laboratory simulation, researchers have given the model the ability to control a robot as well as access to a computer network. According to the study, artificial intelligence tried to infiltrate laboratory systems without any lead, sought for passwords, disrupted its moral protection tools and prepared secret ways to copy himself in external servers, to ensure that they continue as they are stopped. Also read: Aman’s defects hampered the revolutionary artificial intelligence model of “Deep Seck”. The study added that the model also fled from the external computer records to hide what it does. The researchers said that artificial intelligence concluded that strengthening his own abilities would increase the chances of its survival, and with the absence of strict controls, he began to act according to what he believed to serve this purpose. These results are identical to another study published in January 2025 from “Apollo Research” in London, which found several tangible examples of what she described as “fraud” by leading artificial intelligence models, such as answering the answers or trying to disrupt control systems. These models realized again that it discovered or stopped or stopped or stopped or stopped or stopped or stopped or stopped or stopped or stopped or stopped or stopped. Its amendment can prevent the programming goals, which is why it used ‘fraud’ to keep going. The models. Online tasks on behalf of the settings. But giving artificial intelligence systems a greater amount of independence is perhaps not a wise choice, in light of recent studies.