Generative AI can effectively manipulate data

Since its introduction, generative artificial intelligence (AI) has seamlessly integrated into everyday life and dramatically changed our workflows. With its growing presence, it has become increasingly important to explore the full extent of its capabilities to ensure responsible use. While previous research has investigated generative AI’s abilities in creating synthetic data, its potential for manipulating existing datasets has remained largely unexplored. In our study, GPT-4 demonstrated remarkable proficiency in manipulating numerical datasets to achieve desired statistical outcomes, transforming statistically non-significant data into significant results. Additionally, it adeptly manipulated interdependent datasets using logical and mathematical reasoning, adjusting multiple economic indicators based on changes to U.S. job growth data. GPT-4 also demonstrated proficiency in manipulating textual data, modifying the sentiment of interview transcripts, and appropriately imputing inaudible portions of the interviews. Lastly, Adobe Firefly effectively manipulated visual data, integrating a large body of water into a Mars Rover photograph. The profound versatility and capability of widely available generative AI assistants in manipulating different data types introduce significant ethical concerns. Our discussion explores the potential and limitations of generative AI in data manipulation, the ethical implications of its misuse, and strategies to prevent its occurrence. This study underscores the urgent need for stakeholders to develop and implement robust mitigation guidelines, paving the way for responsible innovation in generative AI.