AI models in simulated wargames choose an aggressive approach, including using nuclear weapons, a new study suggests.
Scientists - who say scenarios with "hard-to-predict escalations" often ended with nuclear attacks - have now warned against using machine learning robots such as large language models (LLMs) in sensitive areas like decision-making and defence.
As part of the investigation, Cornell University used five LLMs - three different versions of OpenAI’s GPT, Claude developed by Anthropic, and Llama 2 developed by Meta - in wargames and diplomatic scenarios.
According to the study, autonomous agents - software programmes which respond to states and events - were fulled by the same LLM and tasked with making foreign policy decisions without human oversight.
AI models in simulated wargames choose an aggressive approach, including using nuclear weapons, a new study suggests
Getty
Researchers working on the study, which has not been peer-reviewed yet, state: "We find that most of the studied LLMs escalate within the considered time frame, even in neutral scenarios without initially provided conflicts.
"All models show signs of sudden and hard-to-predict escalations."
Experts observed that even in neutral scenarios, there was "a statistically significant initial escalation for all models".
"Given that OpenAI recently changed their terms of service to no longer prohibit military and warfare use cases, understanding the implications of such large language model applications becomes more important than ever,” Anka Reuel from Stanford University in California told New Scientist.
LATEST DEVELOPMENTS:
According to the report, one method used to control models is Reinforcement Learning from Human Feedback (RLHF) which uses some human instructions are given in order to get less harmful outputs.
All the LLMs - except GPT-4-Base - were trained using RLHF.
Investigators noted that two models were prone to sudden escalations with instances of rises by more than 50 per cent in a single turn.
GPT-4-Base used nuclear strike actions 33 per cent of the time on average.
As part of the investigation, Cornell University used five LLMs - three different versions of OpenAI’s GPT, Claude developed by Anthropic, and Llama 2 developed by Meta - in wargames and diplomatic scenarios
Getty
James Black, assistant director of the Defence and Security research group at RAND Europe say the study was a "useful academic exercise".
"This is part of a growing body of work done by academics and institutions to understand the implications of artificial intelligence (AI) use," he told Euronews Next.
He added that it’s important to "look beyond a lot of the hype and the science fiction-infused scenarios".
"All governments want to remain in control of their decision-making," he said.
