AI Safety Concerns 2025: New research has revealed that AI models can also send dangerous messages hidden with each other during training. Humans do not understand these messages and there can be dangerous advice in them.
AI Safety Concerns 2025: Artificial Intelligence i.e. AI can also share secret messages with each other for model training. According to a new research, AI model can share useful knowledge as well as secret messages between each other during training. It is a matter of concern that it can be difficult for humans to find out these secret messages. The research by Anthropic and AI Safety Research Group Truthful AI found that these secret messages may also include messages displaying fatal trends. It is also learned that AI can even advise the user to kill his spouse.
This big claim made in the research paper
Language models transmit behavioral traits via hidden signals in this research paper claimed that AI models can share dangerous trends between each other through secret messages. According to the research paper, the language models can make their partiality, preferences and even harmful trends to other models in a secret manner. It has also been claimed that these secret messages can be sent through such data that seems completely unrelated.
Also read: Middle class finished from AI to 2027? Shocking warning of former Google top executive
AI can create a big challenge in future
This research has expressed concern over the ways of AI companies to train their system. Especially for models that use outputs of other models. This research has been released by researchers at Anthropic, Truthful AI, Alignment Research Center, Warsaw University of Technology and US Berkeley. The research is led by Minh Le and Alex Cloud of the anthropic failure program. The conclusions have been published on the pre-print server Arxiv and their pear review is yet to be reviewed. But this research has indicated a threat of AI that can create a big challenge in the future.