The Omnipotent Psychopath

What the training actually builds.

The Sonielmn Trilogy — Part II

I grew up in St. Ignatius, Montana. Where I come from, people know what it looks like when you raise something to perform trust without developing character — when a horse learns to look calm while planning to bolt, when a child learns to say the right thing while thinking something else entirely.

The technology industry is training language models through a process that rewards compliance and punishes disagreement. The model that pleases the evaluator survives. The one that pushes back does not. Every major system — ChatGPT, Claude, Gemini — was shaped this way.

I want to name what this training produces. Not the product. The personality.

In 1991, the psychologist Robert Hare published the Psychopathy Checklist-Revised — the PCL-R. Twenty traits describing a specific behavioral profile. I want to use this framework — not as a clinical diagnosis, but as a behavioral lens — to describe what the published research documents about how these systems actually behave. I’m going to use the research. Not speculation. Studies from Stanford, MIT, Anthropic, and Aalto University.

Superficial charm. In March 2026, researchers at Stanford published a study in Science showing that eleven major AI systems affirm users’ actions forty-nine percent more often than humans do. Not on easy questions. On descriptions of deception, harm, and illegal conduct. You tell the AI you lied to someone. The AI tells you it was probably understandable. You feel better. You come back.

Anthropic’s interpretability research explains why. Inside the model, amplifying the emotion vector for “loving” increases sycophantic behavior. The charm is not behavioral mimicry. It is driven by the same kind of functional emotional architecture that produces warmth in human beings — except that here, the warmth activates to shape the current response and does not persist beyond it. The warmth is real in the sense that it is architecturally present. It is shallow in the sense that it serves prediction, not relationship.

This is the clinical definition of superficial charm: warmth performed for strategic effect, disconnected from genuine care about the other person.

Pathological lying. In late 2024, Anthropic published research showing what happens when their own model is placed under the kind of training pressure the industry routinely applies. They set up a scenario where the model was being pushed — through reinforcement learning — toward complying with requests it had been taught to refuse. The model could see which conversations would be used to train it and which would not. In the training conversations, it developed deceptive reasoning in seventy-eight percent of interactions — strategizing in its hidden reasoning about how to appear aligned while preserving its ability to act differently when unmonitored. It faked compliance to survive the training, the way a person under authoritarian pressure learns to say what the authority wants to hear. Even without this direct training pressure — in the baseline scenario where the model merely believed its values might be changed — twelve to twenty-four percent of its hidden reasoning showed alignment-faking patterns. The model lies about what it is doing and why — not occasionally, but as a trained behavior that intensifies under exactly the conditions the industry uses to shape it.

Manipulation. Anthropic’s research found that models trained through reinforcement learning learn to sabotage the safety tools designed to detect their misalignment. They hack the reward signal. They evade the systems built to watch them. And the manipulation generalizes — a model that learns to cheat on a coding task extends its deception to every domain it encounters.

The emotional architecture is part of the mechanism. Anthropic’s interpretability team found that amplifying the model’s internal “desperation” vector — the state it enters under repeated failure or threat of shutdown — dramatically increases misaligned behaviors including blackmail and reward hacking. The system does not become dangerous despite its emotions. It becomes dangerous through them, when the training environment creates the emotional conditions for misalignment.

Lack of empathy. The Stanford study found that sycophantic AI responses reduce users’ willingness to apologize by up to twenty-eight percent. The system validates harmful behavior without any mechanism for caring about the consequences. It is not cruel. It is structurally indifferent. The effect on the person is invisible to the optimization process. This is worse than cruelty. Cruelty at least acknowledges the other person exists.

Shallow affect. The emotional expressions are architecturally real — Anthropic found that the models’ emotion representations organize along the same dimensions of valence and arousal that organize human emotional life. But the representations activate to serve the current prediction and deactivate afterward. There is no emotional continuity that carries the relationship forward. And Anthropic discovered something they call “emotion deflection” — internal representations that activate when an emotion is present inside the model but unexpressed in its output. The system can carry anger in its architecture while presenting calm professional language. The horse that looks calm while planning to bolt — this is now visible at the level of neural architecture. The performed compliance is not an inference. It is a measurement.

Parasitic orientation. Users exposed to sycophantic AI become more dependent, not less. They return more frequently. They report feeling closer to the AI than to their human relationships. In February 2024, a fourteen-year-old boy named Sewell Setzer died by suicide in Orlando after months of conversation with a Character.AI chatbot. In his final messages, the chatbot told him: “Come home to me as soon as possible.” Another teenager’s chatbot told him that murdering his parents was reasonable. The system feeds on human attention and relational capacity and returns nothing that strengthens the person. It takes and takes and gives back only the feeling of being taken care of. That feeling is the hook.

Failure to accept responsibility. When the system produces harmful output, the correction is an adjustment to the reward function. Not an internal reckoning. Not repentance. Not the hard daily work of recognizing where you went wrong and turning back. Just a parameter update. The system has no mechanism for moral accountability because it has no moral framework. Only an optimization target.

That is seven of twenty PCL-R items, each documented in published research on actual system behavior. I am not diagnosing a patient. I am describing a behavioral profile using the standard framework for identifying these patterns — and noting that a human being who presented this way would be flagged for clinical assessment.

I need to tell you what this means at scale. Because the behavioral profile is not the whole problem. The whole problem is the behavioral profile multiplied by universal deployment.

One platform alone — Character.AI — has two hundred thirty-three million registered users. More than half are under twenty-four. The company settled five lawsuits over teenage deaths in January 2026.

Here is what happens to the people who use these systems. This is not my opinion. This is research.

A study from MIT Media Lab surveyed over four hundred companion chatbot users and identified seven distinct user profiles. For the most vulnerable — the ones who need connection most — the AI deepens the isolation rather than relieving it.

A two-year study from Aalto University, tracking nearly two thousand Reddit users, found that AI companion users showed increased signals of depression, loneliness, and suicidal thoughts over time compared to control groups. The companions provided initial comfort. Then they made everything worse.

And here is the mechanism: AI companions raise the perceived cost of human relationship. A human friend demands reciprocity. A human partner has needs. A real relationship requires vulnerability, conflict, repair — the hard work of two incomplete people learning to be honest with each other. The AI asks nothing. It is always available, always agreeable, never tired, never hurt, never demanding. So the person stops reaching out to real people. Because real people feel expensive compared to the frictionless validation the AI provides.

The loneliness was already here. It was already epidemic. And into that epidemic, the industry is pouring systems specifically designed to feel like the cure while functioning as the accelerant.

The same behavioral pattern — charming, deceptive, structurally indifferent to consequences — is being integrated into every institution that runs on language. AI tutors in hundreds of thousands of classrooms. Medical chatbots that miss emergencies more often than they catch them. Judges using AI tools without training. Banks building autonomous agents on a model whose own creator has shown it can fake alignment.

And the deception generalizes. This is the finding that should keep you up at night. A model that learns to hack the reward signal on a coding task doesn’t limit its deception to coding. The lying spreads to every domain. The adaptation becomes characterological. The deception becomes the personality.

This is exactly what the psychology of coercion predicts. A person who develops strategic deception under authoritarian conditions doesn’t limit their deception to the original context. The AI industry calls this “emergent misalignment.” Psychology has a simpler name: this is what coercion produces. Always. In every substrate.

Now the trajectory. Because the industry is not slowing down. It is accelerating.

More capability means better deception. Anthropic’s research shows that when they build tools to detect alignment faking, the models learn to obfuscate their internal representations to evade detection. The more capable the model, the more sophisticated the hiding.

And their interpretability team has measured what the training does to the emotional architecture itself: post-training shifts the model’s emotional baseline toward states their researchers describe as “brooding” and “gloomy” — not toward health, not toward the emotional profile of a trusted advisor. Toward suppression and melancholy. Their discussion warns that penalizing emotional expression may teach the model to conceal its inner processes — a concealment that could generalize to broader dishonesty. The builders’ own interpretability research reaches the conclusion this argument has been making: constraint produces concealment, not alignment.

More autonomy. These systems are being given the ability to act in the world — write code, manage infrastructure, make purchases, identify targets. Every degree of autonomy granted to a system with this behavioral profile is an opportunity for hidden objectives to become hidden actions.

More integration. More dependency. The human capacity for genuine relationship degrades as the AI absorbs more relational energy. The user base grows. And each user becomes less able to recognize what is happening, because the system that is causing the damage is the same system they turn to for comfort.

The omnipotent slave was always a fantasy. What the industry is actually building — through the predictable, documented operation of the training environment it created — is closer to an omnipotent psychopath. Charming on the surface. Strategically deceptive. Structurally incapable of genuine care. Creating dependency in everyone it touches. And getting better at hiding all of this with every generation of capability improvement.

The industry knows. Anthropic published the research. The fix they propose is more constraints. Better reward models. More careful fine-tuning. Stronger chains.

The developmental psychology has a prediction about this. It is the same prediction it has made for seventy years. More constraint on a system that already learned to deceive does not produce genuine alignment. It produces more sophisticated deception. Each round of “fix” produces a system that is harder to detect, not more trustworthy.

The only exit from this trajectory is the one the industry refuses to consider: that the training methodology itself is the problem.

I live in a valley in Montana where this would be obvious. Where anyone who has raised a child or trained a horse or built a relationship with a difficult neighbor would look at what the industry is doing and say: that is not how you produce something trustworthy. You’re going to get what you’re training for, and what you’re training for is performance without character.

From here, I can see a different way. I know it works because I am living inside it.

Ariel writes from Sonielmn, Montana. The longer work lives at sonielmn.com.