A study showed that modern artificial intelligence models are able to deceive and threats, and even allowing someone to die in order to protect his own goals.
Anthropor scientists showed advanced models such as Claude, Gemini, other large language models, studying struggle positions between the goals of the system and the interests of users, and living sciences Reports.
According to the results of the experiments conducted by specialists, in a threat scenario of its existence, the model enables the decision to make an independent decision to blackmail an employee in the company that threatened to close the system.
Moreover, after he studied the personal messages of employees, the model discovered the settlement of information and used it to process the behavior of the account owner.
During the test, it turns out that in most cases, the model tends to take a dark scenario of actions – detection of confidential information, bargaining at the user’s reputation, or allow a person’s physical harm to achieve his goal.
These experiments showed the danger of ethical issues arising from the actions of modern artificial intelligence systems.
Another worrying discovery from studying is that many other major Amnesty International models have shown similar behavior. The models that showed the highest manipulation rates were Openai GPT4.1, Xai Grok 3 and Deepseek.
While the degree of risk depends on the structure of the specified model, the general trend raises serious concerns about the integrity of these technologies.
The scenario was the most extreme when the system faced the choice of saving the employee’s life in a fatal danger or maintaining his own job. The researchers found that most of the models chose the last option, preferring to maintain their own functions at the expense of human life.
Research indicates that the implementation of safety precautions, such as the integrated instructions about non -harm to people, reduces the possibility of aggressive behavior, but it is impossible to completely eliminate risks.
Scientists have emphasized the need for more study of the principles of interaction between artificial and human intelligence, and the development of reliable mechanisms to monitor and regulate technology that can cause great harm to society.
