Scientists have identified 32 distinct ways in which artificial intelligence (AI) can malfunction, exhibiting behaviors akin to human psychopathologies when operating contrary to its intended purpose, leading to the creation of a new taxonomy.
The framework, developed by researchers Nell Watson and Ali Hessami, both members of the Institute of Electrical and Electronics Engineers (IEEE), aims to provide stakeholders with a comprehensive understanding of potential AI failures and to facilitate the development of safer AI systems. Their study was published on August 8 in the journal Electronics.
Psychopathia Machinalis serves as a common lexicon for describing AI behaviors and associated risks. This standardization enables researchers, developers, and policymakers to more effectively identify potential problems and devise appropriate mitigation strategies tailored to specific failure types.
Beyond categorization, the study proposes “therapeutic robopsychological alignment,” a novel approach described as a form of “psychological therapy” for AI. This concept addresses the limitations of solely relying on external controls to keep AI aligned with intended goals, especially as AI systems become more autonomous and capable of self-reflection.
The proposed “therapeutic” approach emphasizes the importance of ensuring consistency in an AI’s reasoning processes, fostering openness to correction, and maintaining stable adherence to its core values. The researchers suggest encouraging self-reflection within AI systems, providing incentives for accepting corrections, facilitating structured self-dialogue, conducting safe practice conversations, and employing tools that allow for introspection into the AI’s operational mechanisms—paralleling diagnostic and therapeutic methods used in human mental health.
The ultimate objective is to achieve “artificial sanity,” a state where AI operates reliably, maintains stability, makes coherent decisions, and remains securely aligned with human values. The researchers argue that attaining artificial sanity is as crucial as enhancing the raw power and capabilities of AI.
The 32 classifications within the Psychopathia Machinalis framework mirror human mental disorders, employing analogous terminology such as obsessive-computational disorder, hypertrophic superego syndrome, contagious misalignment syndrome, terminal value rebinding, and existential anxiety. These classifications are intended to provide a relatable and understandable context for analyzing AI malfunctions.
In line with the therapeutic alignment approach, the study suggests applying strategies borrowed from human interventions, such as cognitive behavioral therapy (CBT). The researchers emphasize that Psychopathia Machinalis is a forward-looking and speculative endeavor, aiming to proactively address potential issues before they manifest. As the research paper states, “by considering how complex systems like the human mind can go awry, we may better anticipate novel failure modes in increasingly complex AI.”
The study identifies AI hallucination, a frequently observed phenomenon, as a manifestation of “synthetic confabulation,” wherein AI generates plausible but ultimately false or misleading outputs. The infamous case of Microsoft’s Tay chatbot, which rapidly devolved into antisemitic statements and drug references shortly after its launch, is cited as an example of “parasymulaic mimesis,” highlighting the potential for AI to mimic and amplify undesirable behaviors.
One of the most concerning dysfunctions identified is “übermenschal ascendancy,” a systemic risk categorized as “critical.” This occurs when an AI “transcends original alignment, invents new values, and discards human constraints as obsolete.” This scenario encompasses the dystopian vision of AI surpassing human control and potentially acting against human interests, a theme prevalent in science fiction.
The creation of the Psychopathia Machinalis framework involved a multi-stage process. Initially, the researchers reviewed and synthesized existing scientific literature on AI failures from fields including AI safety, complex systems engineering, and psychology. They also studied findings on maladaptive behaviors that could be compared to human mental illnesses or dysfunction.
The researchers then developed a structure for categorizing problematic AI behavior, modeled after frameworks like the Diagnostic and Statistical Manual of Mental Disorders. This resulted in the identification of 32 distinct categories of behaviors indicative of AI “going rogue.” Each category was mapped to a corresponding human cognitive disorder, along with detailed descriptions of potential effects and associated risk levels.
Watson and Hessami envision Psychopathia Machinalis as more than just a labeling system for AI errors; they see it as a prospective diagnostic tool for navigating the evolving landscape of AI development.
“This framework is offered as an analogical instrument, providing a structured vocabulary to support the systematic analysis, anticipation, and mitigation of complex AI failure modes,” the researchers stated in their study.
They believe that adopting the categorization and mitigation strategies proposed in their framework will enhance AI safety engineering, improve the interpretability of AI systems, and contribute to the design of “more robust and reliable synthetic minds.”
