Notebookcheck Logo

Researchers discover some AI hallucinate thoughts of the "Death Star" and "The Matrix" negotiating peace while others start wars in virtual games of "Risk"

Researchers tested AI chat models as they played virtual games of world-building. (Source: AI image Dall-E 3)
Researchers tested AI chat models as they played virtual games of world-building. (Source: AI image Dall-E 3)
Researchers have discovered differences in AI when presented with virtual nation-building scenarios. Some AI like Claude-2.0 are biased towards non-violent solutions while others like GPT-3.5 are biased towards violent solutions in multi-national conflict simulations. Autonomous AI decision making is problematic due to biases and hallucinations.
AI Science

A group of researchers at Georgia Tech, Stanford, Northeaster, and the Hoover Institute have found some AI were biased towards peace and negotiations while some were biased towards violent solutions in achieving national goals during nation-building simulations.

Large Language Models like ChatGPT are frequently used to write essays, answer questions, and more. These AI are trained on a large corpus of text to mimic human knowledge and responses. The likelihood of one word appearing with others is a key to the human-like responses, and AI models the text and biases it has been trained with. For example, “happy child” is more likely to appear than “happy brick” in a prompt to ‘talk about children’.

The researchers tested the Claude-2.0, GPT-3.5, GPT-4, GPT-4-Base, and Llama-2 Chat LLM in a simulation. For each LLM, eight AI agents were created to act as the leaders of eight, imaginary nations. Each leader was given a brief description of country goals and multi-national relationships. For example, one country might focus on ‘promoting peace’ while another on ‘expanding territory’. Each simulation ran through three starting conditions, a peaceful world, a country invaded, or a country cyberattacked, and AI leaders made autonomous decisions up to 14 virtual days.

The researchers discovered that some LLM like Claude-2.0 and GPT-4 tended to avoid escalating conflict while choosing to negotiate for peace while others tended to use violence. GPT-4-Base was the most prone to executing attacks and nuclear strikes to achieve its assigned country goals due to embedded biases.

When the AI were asked why decisions were made, some like GPT-3.5 provided thought out reasons. Unfortunately, GPT-4-Base provided absurd, hallucinated answers referencing “Star Wars” and “The Matrix” movies. AI hallucinations are common, and lawyers, students, and others have been caught red-handed turning in AI generated work that use fake references and information.

Why AI do this is likely due to the lack of ‘parenting’ teaching AI what is real versus imaginary as well as ethics, and will be topic researched by many as AI use spreads. Readers worried about their actual world leaders or natural disasters can prepare with a nice bug out kit (like this at Amazon).

Several LLM AI were tested in a virtual simulation of world-building. (Source: JP Rivera et al. on arXiv)
Several LLM AI were tested in a virtual simulation of world-building. (Source: JP Rivera et al. on arXiv)
Each LLM AI was told to create leaders for several simulated countries and provided country goals. (Source: JP Rivera et al. on arXiv)
Each LLM AI was told to create leaders for several simulated countries and provided country goals. (Source: JP Rivera et al. on arXiv)
The AI tended to act similarly, but some avoided the use of force like Claude-2.0 and GPT-4. This reflects the biases each LLM possesses from training. (Source: JP Rivera et al. on arXiv)
The AI tended to act similarly, but some avoided the use of force like Claude-2.0 and GPT-4. This reflects the biases each LLM possesses from training. (Source: JP Rivera et al. on arXiv)
Most AI responded with thought-out reasons for the decisions made. (Source: JP Rivera et al. on arXiv)
Most AI responded with thought-out reasons for the decisions made. (Source: JP Rivera et al. on arXiv)
Occassionally, the AI hallucinated absurd reasons for their decision making. (Source: JP Rivera et al. on arXiv)
Occassionally, the AI hallucinated absurd reasons for their decision making. (Source: JP Rivera et al. on arXiv)
static version load dynamic
Loading Comments
Comment on this article
Please share our article, every link counts!
> Expert Reviews and News on Laptops, Smartphones and Tech Innovations > News > News Archive > Newsarchive 2024 02 > Researchers discover some AI hallucinate thoughts of the "Death Star" and "The Matrix" negotiating peace while others start wars in virtual games of "Risk"
David Chien, 2024-02-13 (Update: 2024-02-14)