Man vs the machine in the struggle for effective text anonymisation in the age of large language models

The collection and use of personal data are becoming more common in today’s data-driven culture. While there are many advantages to this, including better decision-making and service delivery, it also poses significant ethical issues around confidentiality and privacy. Text anonymisation tries to prune and/or mask identifiable information from a text while keeping the remaining content intact to alleviate privacy concerns. Text anonymisation is especially important in industries like healthcare, law, as well as research, where sensitive and personal information is collected, processed, and exchanged under high legal and ethical standards. Although text anonymisation is widely adopted in practice, it continues to face considerable challenges. The most significant challenge is striking a balance between removing information to protect individuals’ privacy while maintaining the text’s usability for future purposes. The question is whether these anonymisation methods sufficiently reduce the risk of re-identification, in which an individual can be identified based on the remaining information in the text. In this work, we challenge the effectiveness of these methods and how we perceive identifiers. We assess the efficacy of these methods against the elephant in the room, the use of AI over big data. While most of the research is focused on identifying and removing personal information, there is limited discussion on whether the remaining information is sufficient to deanonymise individuals and, more precisely, who can do it. To this end, we conduct an experiment using GPT over anonymised texts of famous people to determine whether such trained networks can deanonymise them. The latter allows us to revise these methods and introduce a novel methodology that employs Large Language Models to improve the anonymity of texts.