Student Co-presenter Names

Axel Obrien

Download

Download Full Text (742 KB)

Student Class

2028

Student Affiliation

WISP Intern

Author ORCID Identifier

https://orcid.org/0009-0003-6482-819X

First Advisor

Eugene Santos

First Advisor Department

Department of Engineering Sciences—Thayer School of Engineering

Description

This study aims to investigate hallucination and deception in large language models (LLMs). We research the underlying causes, mechanisms, and consequences of these behaviors. As LLMs become increasingly integrated into our everyday lives it is critical for us to understand these models. That’s why this research aims to enhance the reliability, transparency, and trustworthiness of LLMs. Two methods were implemented to help examine LLMs and when hallucinating and deceiving. Our Llama & Lora models were trained on false information and then told to evaluate true or false statements to see the existing context window of the LLMs. Another method was to train a LLM to lie using the Google Flan model, simply by training it on false information and having it determine which city belonged to which country. Our research leads us to highlight the challenge of defining hallucinations and deception of AI, as there is no standard definition or measurement.

Publication Date

2025

Keywords

hallucination, deception, AI, LLMS, LLama, Lora, Flan

LLM Hallucination Station

Restricted

Available to Dartmouth community via local IP address.

Share

COinS