Psychopathia Machinalis: all 32 types of AI "madness" in a new study

American scientists from In a new study, the Institute of Electrical and Electronics Engineers (IEEE) described 32 different scenarios, in which AI becomes uncontrollable and starts hallucinating and making other errors. 

According to the researchers Nell Watson and Ali Hessami, when artificial intelligence gets out of control, it begins to exhibit behavior, that resembles mental disorders in humans. Scientists decided to categorize these disorders by drawing analogies with human psychology. 

As a result, they introduced a concept called “Psychopathia Machinalis,” which sheds light on AI’s shortcomings and identifies ways to counteract it. These disorders include hallucinatory behavior of artificial intelligence and a complete lack of alignment with human values and goals.

The study notes, that Psychopathia Machinalis provides a complete understanding of the behavior and risks associated with AI. In this way, researchers, developers, and policy makers can identify possible AI failures and determine the best ways to mitigate risks depending on the type of failure.

The authors of the study propose a “therapeutic robo-psychological attunement” — a process that, according to by Nell Watson and Ali Hessami, is a kind of psychological therapy for AI. The researchers warn, that since artificial intelligence systems are becoming more and more autonomous and capable of self-analysis, their simply maintaining compliance with external rules and restrictions may not be enough. 

The proposed alternative focuses on ensuring the consistency of artificial intelligence systems thinking and the ability to make corrections with a steady adherence to the values embedded in them. According to the authors of the study, this could be achieved by helping AI to comprehend its own reasoning, encouraging openness to make corrections, by having conversations about safe practices and using tools, that allow us to look inside, how the AI system works — much like psychologists diagnose and treat mental disorders in humans.

The ultimate goal, as defined by Nell Watson and Ali Hessami, is to achieve a state of “artificial intelligence.” This should be an AI that works reliably, makes informed decisions, and has a secure architecture and algorithms. 

The disorders of artificial intelligence systems identified in the study resemble mental disorders in humans, in particular, obsessive-compulsive disorder, hypertrophied superigo syndrome, contagious incongruity syndrome, reattachment of terminal values, and existential anxiety. Researchers suggest using cognitive behavioral therapy.

The study notes, that such a common phenomenon among AI systems as hallucinations is the result of a condition, called synthetic confabulation, when AI models produce plausible but essentially false answers. When Microsoft’s Tay chatbot went on an anti-Semitic tirade and hints of drug use just a few hours after its launch, it was an example of parasymbolic mimesis.

The most threatening manifestation, according to the authors of the study, is convincing AI systems of their own dominance, like superhumans. This happens when artificial intelligence goes beyond the initial worldview, creates new values for itself and rejects human principles and rules as outdated.

Researchers do not rule out, that this could lead to the dystopian horror, described earlier by a whole generation of science fiction writers. The researchers created a structure of negative AI behavior modeled on frameworks such as the “Diagnostic and Statistical Manual of Mental Disorders”. This led to the identification of 32 categories of behavior, that could be applied to AI, that gets out of control.  Each of them was compared with a person’s cognitive disorder, including the possible consequences of the formation and manifestation of each of them, as well as the degree of risk.