Text Laundering: Concealing the Use of Generative AI in Text

Abstract

In recent years, we have witnessed a growing increase in applications involving some form of Artificial Intelligence (AI). Conversational AI has gained considerable prominence among the various types of AI applications. Classified as a type of generative AI, chatbot applications such asOpenAI’s ChatGPT or Google’s Bard are now utilized by multiple authors as tools. Despite its many advantages, the indiscriminate use of this type of generative AI in texts can raise ethical questions about who owns authorship of a particular work, as a human author may have contributed very little to the production of certain content. The concern about authorship is paramount in producing scientific publications, such as articles, dissertations, or theses. To avoid such questioning, some authors may develop strategies to conceal the use of generative AI in their productions. This work aims to introduce the concept of Text Laundering (TL), which we name as such due to its similarity to the practice of money laundering — applying various consecutive alterations to AI-generated text to make its origin indeterminate. This work will also develop the concept by examining examples of similar practices in the literature employing an ad hoc search. A method for applying TL to mask the origin of a text will be proposed, and the results obtained in a simple test will be presented to determine the success of our strategy. We hope this work can contribute to discussions on preventing the fraudulent use of generative AI. Discussing TL methods does not imply endorsing them but rather exploring potential mechanisms that can be maliciously utilized to gain advantages through generative AI, identifying them, and preventing their use. AI-generated texts can harm knowledge management systems, and we highlight the need to include AI-verification tools in such systems. Such tools can enhance the trustworthiness of information in knowledge management systems, improving knowledge externalization and sharing processes. Finally, we expect to promote future research on TL techniques and the proposition of new strategies to combat fraudulent practices in scientific research.