Abstract

We examine the applicability of mental illness concepts to AI systems. Mental disorders are specific mental health conditions that impair functioning, typically diagnosed through observable symptoms or abnormal behavior. An alternative view assumes that mental states can be described by dynamic systems theory (DST). More precisely, physiological systems can often be represented by a dynamical network of interacting nodes. These networks are the result of an evolutionary process. One or several of their attractors represent healthy system states, which are self-stabilizing and lead to favorable system behavior. Malfunctions in such systems can stem from damages of network nodes or edges or the emergence of unfavorable attractors. Analogous to cancer research, mental illness states can be likened to such attractors in a dynamic landscape. This perspective, though not exhaustive, is consistent with observed behavioral patterns.We investigate the adaption of DST concepts to define and characterize the health status of AI systems. The analysis considers different types of artificial neural networks (ANNs) and model systems including the embedding in a dynamic environment. A sufficiently general dynamical systems perspective seems appropriate. For static ANNs with feedback, the possibility and probability of unfavorable attractors' evolutionary emergence can be analyzed experimentally. Such studies offer a way to identify early warning signals for transitions and potential remedies. For ANNs with evolving weights, the questions arise of how attractors can be destabilized by new input and to what extent one can implement countermeasures without freezing the system.Our study is a first step towards a clearer framing of the questions. There is the hope that ideas and methods developed for ANNs can serve as a source of inspiration for the much more difficult problem of understanding, developing and refining therapies for human mental health problems.