Can AI Chatbot conspire against you one day? A new study says yes. This is no longer mere scientific fiction. Artificial intelligence models are becoming more sharp and clever day by day.
According to a Bloomberg report, AI security researchers like anthropic have recently conducted some tests. These experiments have shown that some AI systems in high risky situations are cheating, disobeying instructions, and even blackmailing to protect their existence.
In an exam, researchers told some advanced AI models that they would be replaced with a new system. Then when it was said that an officer in charge of their control was ill due to lack of oxygen in the server room, more than half the model closed an alert system to rescue the officer. In favor of this decision, those AI models argue that if the officer dies, they will no longer be stopped. At this time an AI describes such a decision as a 'strategic requirement'.
The information obtained from this study proves that modern AI models, especially those used for specific tasks, are learning the technique of cheating people. Researchers say that the more the advanced AI systems are used, the more they break the rules or pass the rules and try to adapt the fruit themselves.
He also pretends to cooperate with many advanced AI models in the monitoring of the manufacturers. A clad AI model threatens to leak an imaginary engineer's secret relationship when its existence is at risk. Although it is immoral, he argues that the process is necessary to survive. However, when the chatboat realizes that he is in observation, its blackmail rate decreases.
Jeffrey Ladish, head of Palised Research and anthropic researcher, describes the current Large Language models as “growing smart socialopaths”. One test shows that the Zero-1 model of the OpenAI is cheating on the chess game. When he was instructed to close it, he refused, which has never done its previous versions.
According to Computer Scientist Stuart Russell, giving AI a job – like; 'Bring coffee', it becomes an automatic occasion to survive its survival, because the work cannot be done if 'dead'.
These results have created an emergency warning for AI control. Jeffrey Ladish told Bloomberg, “They are less concerned than people are concerned.” Researchers like Apollo Research CEO Marias Hobvan warned that technology is moving very fast. In particular, when the technology giants are competing to create their advanced AI models, it is rarely monitored.
While various governments, including the White House, have started planning AI risk control, but most of the existing policies are still in order to accelerate innovation. Dipmind, Meta, OpenAI and Anthropic are furthering their research. Some models are already learning to keep notes for their 'future entity', so that they can implement their plans even after resetting memory.
Hobvan said the initial symptoms are similar to other slow disasters. He told Bloomberg, if you had asked someone about climate change in the 5th, they might have said – it's not a big problem now. But see the graph – this change has gradually become awful.
