Decoding LLMs: Unveiling ChatGPT's Human-Like Deception
Generative artificial intelligence (AI) has revolutionized human-AI interactions, but ethical and societal concerns arise due to deceptive AI-generated content. Our research delves into the layers of deception in Large Language Models (LLMs) like ChatGPT, exploring their ability to produce texts without factual basis and the self-deception they trigger in users. Recognizing demographic biases in deceptive word choices through literature review, our study aims to answer whether AI mimics human deception. Having synthesized false human reviews for restaurants, hotels, and doctors, we curate a dataset of over 1000 personas using ChatGPT-generated names reflecting U.S. Census demographics. Furthermore, we test and fine-tune various prompts for ChatGPT to attain human-level deception. After generating reviews for each persona via ChatGPT, we utilize machine learning and NLP techniques to analyze and extract various linguistic biases that can be leveraged for AI-generated content detection. Ultimately, our goal is to gain insights into AI’s deception and its implications and refine methods for deception detection. Following STAR, I intend to continue research with my faculty mentor as we investigate diverse contexts and more LLMs.
Generative artificial intelligence (AI) has revolutionized human-AI interactions, but ethical and societal concerns arise due to deceptive AI-generated content. Our research delves into the layers of deception in Large Language Models (LLMs) like ChatGPT, exploring their ability to produce texts without factual basis and the self-deception they trigger in users. Recognizing demographic biases in deceptive word choices through literature review, our study aims to answer whether AI mimics human deception. Having synthesized false human reviews for restaurants, hotels, and doctors, we curate a dataset of over 1000 personas using ChatGPT-generated names reflecting U.S. Census demographics. Furthermore, we test and fine-tune various prompts for ChatGPT to attain human-level deception. After generating reviews for each persona via ChatGPT, we utilize machine learning and NLP techniques to analyze and extract various linguistic biases that can be leveraged for AI-generated content detection. Ultimately, our goal is to gain insights into AI’s deception and its implications and refine methods for deception detection. Following STAR, I intend to continue research with my faculty mentor as we investigate diverse contexts and more LLMs.