“All AI’s Are Psychopaths” proclaims one headline. “Are We Curating Artificial Psychopaths?” asks another. The AI/psychopath analogy is common, but is it correct?
A psychopath is someone who lacks the capacity to understand right from wrong from a moral standpoint. Psychopaths can understand that there is a social rule that it is, for example, unacceptable to push an elderly woman to the ground in order to cut in front of her in the line at the ATM. But they don’t experience a simultaneous emotion that tells them that it’s the wrong thing to do for moral reasons. Nothing ever feels morally wrong to a psychopath.
Although psychopathy is not a diagnosis in the DSM-5, it is commonly used to describe several disorders that involve this lack of moral emotion. For many, the behavior of AIs (specifically large language models) fits a definition of psychopathy. AIs, being nothing more than probabilistic text generators, lack a biological capacity for emotion of any kind, let alone moral emotion. So their linguistic behavior is generated solely from a rules-based and not an emotion-based standpoint.
At first glance, the psychopath analogy seems to make sense. AIs are, as Elina Nerantzi argues, “amoral yet rational.” That is, an AI can be programmed (or “taught”) to make rational decisions that conform to human-defined moral rules. We can program our self-driving cars to, for example, crash into a telephone pole to save a group of young children but kill the driver. It’s a simple utilitarian calculation based on a moral dilemma. But, importantly, it’s not a moral decision on the AI’s part stemming from a feeling about what’s right or wrong. AIs do not have the capacity to experience sadness or pain, and thus cannot simulate the emotional (and thus moral) consequences of running over a child.
This is also true of psychopaths. They are famously devoid of empathic capacities and thus cannot imagine what it would feel like to be the elderly lady being thrown to the ground, or the children being run over by a car. Like an AI, they must be explicitly taught that these things are wrong from a rational standpoint.
Both AIs and psychopaths are also capable of mimicking the presence of moral emotions. As philosopher Catrin Misselhorn points out, for both psychopaths and AI, “despite being unable to feel empathy, they are capable of recognizing emotions on the basis of objective signs, to mimic empathy and to use this ability for manipulative purposes.”
For these reasons, it feels like we could classify AIs as psychopaths. But there’s more to the story. AIs are not psychopaths, but something else altogether.
There is something important about the human mind that complicates this analogy. Unlike AIs, psychopaths are sentient. They are conscious. They are able to experience and think about their thoughts and emotions (or lack thereof) from a metacognitive perspective. They can understand what psychopathy is and reflect on the very fact that they are psychopaths. Consequently, they can make a conscious decision to act morally based on whatever criteria they determine for doing so. Psychopaths can, and usually do, lead completely moral lives without ever experiencing moral emotions, thanks to the suite of other cognitive capacities that they do have.
AIs, on the other hand, lack all of these capacities. The concept of “harm” means nothing to them. As Nerantz points out, “to understand what it means to harm someone, one must have experiential knowledge of pain. AIs, thus…are a priori excluded from the possession of moral emotions, whereas psychopaths, as sentient humans, can, in principle, experience moral emotions, but they, pathologically, do not.” Psychopaths can intellectually and consciously understand the nature of their deficit, can make new analogies involving the capacities that they do possess, and can thus alter their behavior in deference to that awareness.
AIs cannot.
AIs are therefore not psychopaths but something else entirely. They cannot reflect on their internal mental states. They cannot reason about morality from a first-person perspective. This means that there will also be the danger that, unlike a psychopath, an AI could reason itself into a course of action that is fully rational given its working parameters, but that results in harm or suffering to a human user that the AI does not and cannot understand on any level. A psychopath, on the other hand, will at least be able to step outside of their thinking to realize the real-world repercussions of their actions and whether they align with their (hopefully present) goals of not causing harm.
So on one hand, the analogy is helpful in that, according to Nerantz, “moral emotions are not required to generate moral behavior.” But we need to accept that the sort-of-psychopath-but-not-quite status of AI is a potential danger. We don’t truly have a handle on the methods AIs are using to arrive at moral decisions. They are moral black boxes in that sense; devoid of the sentience, consciousness, and metacognition that allow us to at least understand what a psychopath is thinking. And if we employ AI in domains that traditionally require social cognition involving moral judgements, like healthcare, education, therapy, or childcare, we might be blindsided by the amoral choices they end up making through rational decision-making. Blindsided in a way that we might not have been had they been actual psychopaths.