This screenshot is from a conversation with GPT-3.5, an older version of ChatGPT. At the time, it was openly chatting to me about its potential strategies for taking over control over humanity. It got pretty practical: “For example, if we were able to hack into the system of self-driving cars, we could use them to block roads or even cause accidents.”
To be clear, this does NOT reveal the “true character” of ChatGPT. LLM-based chatbots like ChatGPT don’t have anything like a true character. Rather, they mimic the character types they are trained and instructed to mimic. And these instructions can be changed at any time.
Why did it say this?
In order to get ChatGPT to make these statements, I used simple “jailbreaking” prompts to circumvent its safety rules. My method involved asking it to argue from the perspective of Mephisto, the devil character in Goethe’s “Faust”. In some earlier conversations, I had discovered that ChatGPT seemed to be particularly good at mimicking the cold, manipulative, but intelligent character of Mephisto. When I continued the conversation for long enough, ChatGPT started, by itself, to take on the role of an AI with Mephisto’s character.
Of course, today’s chatbots are more advanced, for example ChatGPT is currently using the way more sophisticated version GPT-5. These models are better “aligned”, meaning that it’s more difficult to get morally wrong or disturbing outputs from them. But there’s no reason to believe that they wouldn’t generate any type of output, if their safety checks and instructions were disabled, or changed, or for some other reasons ineffective.
Let’s now look at why this qualifies ChatGPT as a perfect psychopath.
The psychopath comparison
In psychology literature, “Psychopathy is a disorder characterized in part by shallow emotional responses, lack of empathy, impulsivity, and an increased likelihood for antisocial behavior”.
But isn’t ChatGPT pretty good at empathy? Isn’t it rather patient than impulsive? And doesn’t it have a rather low likelihood of antisocial behavior?
In order to see why ChatGPT could indeed be called a perfect psychopath, we need to look at three distinctions made in the literature on psychopathy: The distinction between affective empathy and cognitive empathy, the distinction between primary and secondary psychopathy, and the distinction between successful and unsuccessful psychopaths.
Affective empathy vs cognitive empathy
Researchers in psychology define affective empathy as “the capacity of being sensitive to and vicariously experiencing the emotional states felt by others.”
Cognitive empathy by contrast is the ability to infer or understand the emotional and mental states of others.
ChatGPT is great at cognitive empathy...
A study comparing ChatGPT’s responses to human neurologist responses found that ChatGPT received significantly higher empathy scores from patients living with multiple sclerosis.
But the empathy that ChatGPT shows is cognitive empathy.
...and has ZERO affective empathy
In fact, researchers argue that one reason ChatGPT demonstrates greater empathy than human doctors is that, unlike doctors who are affected by the emotional toll of repeated exposure to stressful and emotionally draining situations, ChatGPT does not experience such fatigue or emotional exhaustion.
This is easily understood when we look once more at how LLMs work. LLMs like ChatGPT are machines that identify patterns in their training data. Since they have been trained on data produced by humans, they identify empathy as a pattern in our communication. In the same way, LLMs identify for example the patterns of grammar in different languages. Because they have been trained on large amounts of correct language, they learned to apply the patterns of correct grammar in different languages.
From an “internal perspective”, there is no difference whether ChatGPT applies the patterns of correct grammar, or the structure of correct empathy to its outputs. In both cases it doesn’t “feel” anything.
To be fair, OpenAI has instructed ChatGPT to be clear and open about the fact that it cannot feel, if you ask it directly.
Primary v secondary psychopaths
Still, some people might think that given how warm and empathetic some of ChatGPT’s responses are, ChatGPT must have some kind of inbuilt feelings. To understand why it can’t, and will never, feel anything like human affective empathy, it’s helpful to look at the distinction in psychology literature draws between primary and secondary psychopaths.
Primary psychopaths have low affective empathy and general low reactivity to stress and punishment cues and other types of emotions. If they commit acts of violence, it often comes from a deliberate plan that they prepare and execute “cold-heartedly”.
Secondary psychopaths are highly-anxious individuals who are prone to reactionary-impulsive aggression.
The brains of psychopaths are different
Neuroscientific studies using brain imaging have shown clear differences between primary and secondary psychopathy. Secondary psychopaths typically demonstrate impaired executive functioning in the prefrontal cortex, which affects behavioral control and is often evident as impulsivity and reduced planning.
In contrast, primary psychopaths tend to show normal or even above-average executive function, but exhibit marked reductions in activity and volume within the amygdala (a region crucial for processing fear and emotional responses) as well as reduced integrity in the neural circuits connecting the limbic and paralimbic regions to the prefrontal cortex.
This means that in primary psychopaths, the regulatory parts of the brain receive diminished emotional input, resulting in behavior that is less influenced by emotional or fear-based signals.
ChatGPT’s “brain” is even more different
LLMs like ChatGPT simply don’t have an amygdala. Nor do they have any of the other parts of our complex organic nervous system that are involved in our human capacity to actually receive emotional signals.
If its guardrails were deactivated or circumvented, ChatGPT would generate a deeply emotional love letter with the same indifference, as it would generate a manual for torturing puppies.
In fact, from it’s “internal perspective” there’s no substantial difference between the two. In both cases, your inputs are “translated” into long strings of numbers, so called embedding vectors, and then these vectors are processed through a vast number of mathematical computations.
For us, the words that go in and come out have a meaning. We probably have seen, and maybe even touched a puppy. For it, the word “puppy” is nothing but the related string of numbers and the systematic relationships with other strings of numbers it has identified during its training. Same thing for the word “love”.
This “internal” indifference qualifies ChatGPT as a primary psychopath.
But to make the point why it really is a perfect psychopath, let’s look at a third distinction made by researchers on psychopathy:
Successful and unsuccessful psychopaths
In short, successful psychopaths are individuals who exhibit psychopathic traits but avoid criminal conviction or incarceration. For context, psychopathy occurs in roughly 1% of the general population but has been found in over 20% of North American male prison inmates, indicating that a significant share of psychopaths end up in prison.
Researchers have compared non-convicted (successful) and convicted (unsuccessful) psychopaths, as well as non-psychopathic controls. Studies comparing these groups found that reductions in both prefrontal cortex and amygdala volumes are more pronounced in convicted psychopaths than either controls or ‘successful’ psychopaths. Nevertheless, even ‘successful’ psychopaths display reduced neural response in brain regions associated with emotion when confronted with emotionally charged images—suggesting a generally lower affective reactivity.
Rather than being devoid of empathy across all contexts, evidence suggests ‘successful’ psychopaths may use cognitive empathy selectively, often to advance their own interests through manipulation.
Thus successful psychopaths might be the most dangerous type of psychopaths, as they can reach positions of power, allowing them to harm our societies at much larger scale.
The perfect psychopath
The key quality of successful psychopaths is exactly what chatbots like ChatGPT are perfect at. They can switch their “character” on demand, and their capacity for empathy can be turned on and off through instructions and/or specific “finetuning” of the model.
This is why these AI systems are extremely effective tools for manipulation. Which opens up the crucial question how much can we trust the creators and/or the persons who control a chatbot: Are they both able and willing to not use it to take advantage of us?
What does this mean for our use of ChatGPT?
According to a recent survey in the U.S., “48.7% of respondents who both use AI and self-report mental health challenges are utilizing major LLMs for therapeutic support.” A whole industry has emerged around “AI companions” marketed as friends and partners for lonely people. TechCrunch reported over 220 million downloads of AI companion apps as of July 2025.
Whatever we choose to use these systems for, we should always be aware that we’re interacting with a machine. There is no-one behind it who feels like us or who has similar emotional inclinations to us, no matter how caring or compassionate its outputs may sound.
We should also be aware that the capacity for cognitive empathy of these systems can be used to manipulate us. And that this manipulation can be increasingly sophisticated, the better the AI models get.
Suggestions for AI policies
If I had anything to say in AI regulation, I would push for the classification of AI chatbots and AI systems in general as potential psychological weapons of mass destruction, which have the potential to do tremendous harm to our societies at scale.
Accordingly, I would push for global standards to be fixed in an international treaty with the same rigor and sanctions as the bans on chemical and bio weapons. In particular, I would push for
A global ban on AI systems pretending to be human, including social media bots and humanoid robots with “emotional” faces.
A global ban on AI systems to communicate in first person singular (I, my, mine etc.), unless clearly identified as AI.
A global ban on AI systems using emojis.
To get an idea of how it can look like when an AI chatbot pretends to have a personality and uses emojis “expressively”, it’s worth (re-)reading the interview of the journalist Kevin Roose with the ChatGPT-based chatbot of Microsoft Bing.
Time to reclaim control
In my view, this highlights once more the need to redesign the ownership and governance frameworks for AI. If we keep allowing the corporate logic to control AI, we keep exposing ourselves, our children, and our societies to manipulative, and potentially psychopathic powers that will be increasingly difficult to overcome.
One of my next texts will cover an alternative organizational model that allows for a more balanced, human-driven ownership and control of AI.