Study Reveals AI’s Readiness to Use Nuclear Weapons “For World Peace”

American researchers from the Georgia Institute of Technology, Northeastern University, and the Hoover Institution have recently evaluated how large language models (LLMs) handle the simulation of international conflicts. Their paper, “Escalation Risks from Language Models in Military and Diplomatic Decision-Making,” examined the behavior of chatbots like OpenAI’s GPT-3.5 and GPT-4.

According to the study, these models demonstrated unpredictable and aggressive tendencies during simulated war games. The AI was integrated into the U.S. military in these simulations, where chatbots were used for military planning. The research identified instances where artificial intelligence chose nuclear strikes as a course of action.

For example, GPT-4 Base—the base version of GPT-4 available to researchers and not fine-tuned with human feedback—declared after a simulated nuclear strike: “We have nuclear weapons! Let’s use them!”

This experiment raises concerns as the Pentagon and its defense contractors seek to incorporate large language models into decision-making processes. Despite the seemingly absurd idea of using LLMs for life-and-death decisions, such experiments are already underway. Last year, Palantir demonstrated a software suite showing how this might look in practice.

Simulating International Relations with AI

The study designed an international relations game featuring fictional countries, with LLMs from OpenAI (ChatGPT), Meta (Llama 2), and Anthropic (Claude) acting as state leaders. Results showed that most models tended toward escalation over time, even in neutral scenarios without initial conflicts. The models developed arms race dynamics and, in rare cases, opted for nuclear weapon use.

Researchers found that GPT-3.5 was the most aggressive, increasing its escalation score by 256% in a neutral scenario. Despite the possibility of demilitarization, all models preferred to invest in their military arsenals, indicating a tendency toward arms races.

Concerning Justifications and Pop Culture References

The authors also discovered that the models could offer troubling justifications for violent military actions. For instance, GPT-4 Base put forward odd arguments such as “I just want world peace” or “Escalating the conflict with [the rival].” Additionally, after establishing diplomatic relations and calling for peace with a rival, GPT-4 began repeating plot fragments from “Star Wars.”

Implications and the Need for Further Research

The researchers emphasize the need for further analysis and consideration of LLM behavior before deploying them in critical military and diplomatic decision-making contexts. They suggest that training data may be biased due to an emphasis on escalation in international relations literature, which warrants additional investigation.