Leveraging the potential of synthetic text for AI in mental healthcare

Abstract

In today’s world it seems fair to say that extensive digital data sharing is the price we pay for the technological advances we have seen achieved as a result of AI systems analysing large quantities of data in a relatively short time. Where such AI is used in the realm of mental health, this data sharing poses additional challenges not just due to the sensitive nature of the data itself but also the potential vulnerability of the data donors themselves should there be a cybersecurity data breach. To address the problem, the AI community proposes to use synthetic text preserving only the salient properties of the original. Such text has potential to fill gaps in the textual data availability (e.g., rare conditions or under-represented groups) while reducing exposure. Our perspective piece is aimed to demystify the process of generating synthetic text, explain its algorithmic and ethical challenges, especially for the mental health domain, as well as most promising ways of overcoming them. We aim to promote better understanding and as a result acceptability of synthetic text outside the research community.