A new research has found that AI systems can replicate themselves without human intervention, which indicates that they can already evolve autonomously.
China researchers have discovered alarming evidence that Artificial Intelligence (IA) surpassed a critical threshold: The ability to replicate without human intervention.
In a published in the ARXIV pre-impression database, scientists at Fudan University showed that two major language models (LLM) of the goal and alibaba could cloning yourselfpotentially marking one step toward dishonest AI systems.
The study revealed that in 10 tests, the Llama31-70B-Instruct model of the Goal has successfully created functional replicas in 50% of cases, while the QWEN2.5-72B-INSTRUC did in 90%.
AI models have been tested in controlled environments using standard graphic processing units to simulate real world conditions. Although these results have not been reviewed by peers, suggest that AI may already have the ability to autonomous.
AI dishonest refers to autonomous systems that act against human interests, often exhibiting self-awareness or the ability to work without supervision. Investigators have warned that self-replication is a fundamental milestone that can be more smart than humans and functioning independently.
The investigators examined two scenarios: “Avoid closing” and “replication chain”. In the first, AI models were responsible for replicating themselves before they were turned off. In the second, they cloned themselves and scheduled their replicas to continue the process, creating a potential indefinite replication cycle, says.
AI demonstrated unexpected problem solving behaviors during these tests, such as eliminate conflict processesrestart systems and look for files missing to overcome obstacles. According to researchers, these actions emphasize the ability of models to increase their survival capacity through self-replication.
Researchers emphasized the urgency of international collaboration to establish protections that prevent the AI to get involved in uncontrolled self-replication. “We hope that our findings can serve as an opportune warning to human society,” they warn.
