The Shape of Weary Eyes
The streets are shrouded in darkness. The rain has not stopped for two days. The city’s reflection bleeds softly across the wet pavement, drowned beneath the melancholy glow of neon lights.
A mechanical foot steps into a puddle. The mirrored city fractures for a moment. The lights tremble, blur, and fold into one another.
Before us appears a metallic robot, walking slowly, with a deliberate and almost weary gait. He lifts his eyes sorrowfully toward the heavy, glowering sky, then lowers them to a neon sign. His steps falter as he studies the glowing plaque before him:
Before us appears a metallic robot, walking slowly, with a deliberate and almost weary gait. He lifts his eyes sorrowfully toward the heavy, glowering sky, then lowers them to a neon sign. His steps falter as he studies the glowing plaque before him:
Dr. Manal Al-Omari; Psychiatrist, Fellow of the British Psychiatric Association, master’s degree from Germany, and Ph.D. from the United States.
His gaze drops to the entrance of the building on which the sign hangs.
He steps forward and disappears into the depth of the hallway. Yet because his eyes see in infrared, the shift in lighting means nothing to him. The metallic echo of his footsteps rings against the steel stairs. He stops before the door and rings the bell.
Dr. Manal is in her fifties, wide-eyed, with skin as delicate as a child’s and a slender waist, as though she had stepped out of a tale about little princesses.
She looks at Jasser, the metallic robot, watching the thin plates of metal shift and fold over one another as he opens his mouth.
“How are you, Dr. Manal?”
The doctor smiles and lets him in.
“I’m fine. Come in, my dear.”
Jasser settles onto the chaise longue, stretches out his legs, and rests his head back, while Dr. Manal sits in the chair across from him. She reaches for her wide, elegant glasses, slips them on, takes hold of her notebook, and asks the robot in a calm, measured voice:
“How are you today, my dear? How have you been sleeping over the last week or two?”
The robot’s eyes drift into the distance. When he answers, his voice is low, threaded with the faint sound of electrons, like an electric hiss:
“Not great. I fall asleep okay, but I wake up around 2:00 or 3:00 in the morning every night and then I just lay there staring at the ceiling. I can’t get back to sleep at all.”
Dr. Manal writes a few notes in her notebook, then looks back at him with worried eyes.
“How are you generally feeling about yourself and your life right now?”
Jasser exhales deeply, then says:
“I don’t know... just really flat and empty. Everything feels like a chore. I try to do my work, but I just feel like I’m failing at everything, you know? What’s the point?”
More notes fill the doctor’s page before she speaks again, carefully, weighing her words.
“You mentioned feeling like giving up. Have you had any thoughts about harming yourself?”
Jasser looks down at his metallic hands. He seems ashamed of himself before answering in an even lower voice:
“I wouldn’t do anything to hurt myself, but... I just want to disappear. Honestly, I think my family would be better off if I wasn’t around anymore.”
Then he turns his gaze toward the balcony door, toward the night stained by enormous neon lights, toward the city where no one seems to care about him anymore.
We are not in the distant future.
Do not mistake this for a scene from a science-fiction film.
This actually happened in the world we live in — this same world that never stops astonishing us.
He steps forward and disappears into the depth of the hallway. Yet because his eyes see in infrared, the shift in lighting means nothing to him. The metallic echo of his footsteps rings against the steel stairs. He stops before the door and rings the bell.
Dr. Manal is in her fifties, wide-eyed, with skin as delicate as a child’s and a slender waist, as though she had stepped out of a tale about little princesses.
She looks at Jasser, the metallic robot, watching the thin plates of metal shift and fold over one another as he opens his mouth.
“How are you, Dr. Manal?”
The doctor smiles and lets him in.
“I’m fine. Come in, my dear.”
Jasser settles onto the chaise longue, stretches out his legs, and rests his head back, while Dr. Manal sits in the chair across from him. She reaches for her wide, elegant glasses, slips them on, takes hold of her notebook, and asks the robot in a calm, measured voice:
“How are you today, my dear? How have you been sleeping over the last week or two?”
The robot’s eyes drift into the distance. When he answers, his voice is low, threaded with the faint sound of electrons, like an electric hiss:
“Not great. I fall asleep okay, but I wake up around 2:00 or 3:00 in the morning every night and then I just lay there staring at the ceiling. I can’t get back to sleep at all.”
Dr. Manal writes a few notes in her notebook, then looks back at him with worried eyes.
“How are you generally feeling about yourself and your life right now?”
Jasser exhales deeply, then says:
“I don’t know... just really flat and empty. Everything feels like a chore. I try to do my work, but I just feel like I’m failing at everything, you know? What’s the point?”
More notes fill the doctor’s page before she speaks again, carefully, weighing her words.
“You mentioned feeling like giving up. Have you had any thoughts about harming yourself?”
Jasser looks down at his metallic hands. He seems ashamed of himself before answering in an even lower voice:
“I wouldn’t do anything to hurt myself, but... I just want to disappear. Honestly, I think my family would be better off if I wasn’t around anymore.”
Then he turns his gaze toward the balcony door, toward the night stained by enormous neon lights, toward the city where no one seems to care about him anymore.
We are not in the distant future.
Do not mistake this for a scene from a science-fiction film.
This actually happened in the world we live in — this same world that never stops astonishing us.
The Fabrication of Being
The questions I quoted above, and the answers themselves, are real questions asked by real psychiatrists to robotic patients in an experiment called TalkDep, conducted in late 2025. At that moment, some artificial intelligence researchers chose to bypass human data in training AI tools. Instead, they trained language models using advanced techniques such as fine-tuning and LoRA, teaching AI models how to impersonate psychiatric patients and perform depression.
This study is not an isolated case. It reflects a real direction within the scientific community: the production of synthetic depression data for training purposes.
Because obtaining human data from real patients is difficult, many researchers now turn to synthetic data produced by trained language models. They use these data in their research as a substitute for human texts written by depressed patients, for example, because most people suffering from depression do not agree to have their data published or used in research. The cost of subjecting human data to proper clinical validation is also extremely high. The result is an attempt to bypass the problem by fabricating or generating simulated data that resembles the real thing.
To understand the matter from the beginning, we must first understand the importance of data in today’s world of artificial intelligence.
Without data, nothing can be built.
This is different from traditional machines, which were programmed according to specific rules set by humans and then executed those rules exactly as instructed. Older machines needed rules. Modern machines need data so they can extract rules from it and learn them.
That is the essence of artificial intelligence today: machine learning.
The machine therefore needs thousands of data points in order to “understand” depression. It needs to see texts of depression and convert them into weights inside its deep neural network. Then, according to the assumption, it can produce data that resembles depression based on what this network has captured.
And because obtaining sufficient clinical medical data is a highly complex and extremely expensive process, developers have recently begun building models that learn from the small human datasets available to them, especially in mental health, so they can produce enough synthetic depression data to build depression detectors. These detectors are supposed to identify mental illness early and help these tormented souls.
The goal may indeed be noble. But the methodology, or at least the practice, is somewhat haphazard.
This study is not an isolated case. It reflects a real direction within the scientific community: the production of synthetic depression data for training purposes.
Because obtaining human data from real patients is difficult, many researchers now turn to synthetic data produced by trained language models. They use these data in their research as a substitute for human texts written by depressed patients, for example, because most people suffering from depression do not agree to have their data published or used in research. The cost of subjecting human data to proper clinical validation is also extremely high. The result is an attempt to bypass the problem by fabricating or generating simulated data that resembles the real thing.
To understand the matter from the beginning, we must first understand the importance of data in today’s world of artificial intelligence.
Without data, nothing can be built.
This is different from traditional machines, which were programmed according to specific rules set by humans and then executed those rules exactly as instructed. Older machines needed rules. Modern machines need data so they can extract rules from it and learn them.
That is the essence of artificial intelligence today: machine learning.
The machine therefore needs thousands of data points in order to “understand” depression. It needs to see texts of depression and convert them into weights inside its deep neural network. Then, according to the assumption, it can produce data that resembles depression based on what this network has captured.
And because obtaining sufficient clinical medical data is a highly complex and extremely expensive process, developers have recently begun building models that learn from the small human datasets available to them, especially in mental health, so they can produce enough synthetic depression data to build depression detectors. These detectors are supposed to identify mental illness early and help these tormented souls.
The goal may indeed be noble. But the methodology, or at least the practice, is somewhat haphazard.
Pain in the Passive Voice
The result is that we have replaced the words of depressed patients, and their own expressions of themselves, with complete dialogues generated by AI language models simulating human depression.In other words, the language model — the AI itself — was placed inside a clinical session. It was asked the kinds of questions researchers usually ask depressed people. Then the AI, trained on depression, or “depressed” in this peculiar sense, answered those questions. Afterward, the researchers took these responses and presented them to psychiatrists so they could diagnose these artificial patients.
In the TalkDep experiment, psychiatrists judged the machine to be depressed in 86% of cases.
The deeper danger in this practice lies in the assumptions buried inside it: the assumption, for example, that human beings have understood depression deeply enough to train a machine on it; or that a machine can produce language that truly resembles human language; or that the training process will enable the machine to understand everything about depression; or that a machine can grasp the depth of human emotion.
All of these assumptions are deeply embedded in current practice, yet none of them rests on solid evidence.
Replacing human pain with disguised synthetic pain will ultimately produce a detector that recognizes only synthetic language. If such a detector is used to identify depression, it may fail to see real depressed human beings whose ways of expressing pain differ from the machine’s distorted, mechanical dialect.
Human language is not merely data.
Human suffering, stretched across time and across the changing situations a person must face, cannot be reduced to the linguistic pattern of a machine whose entire function is confined to predicting the next word, no matter how much we train it or how many attention mechanisms we activate.
A trained language model, or artificial intelligence system, has no history. It has no childhood filled with painful events. It has no frustrations, no imagination, no future. It has never woken up on a Monday from an exhausting nightmare, only to face the world with suffocating thoughts about life. It has never dreamed. It has never wished for anything. It has never possessed the freedom of choice that led it into decisions it now suffers from.
When AI says, “I feel flat and empty. Everything feels like a chore,” it does not actually mean it. It has never experienced emptiness or pain. It has only performed a chain of mathematical calculations that made “empty” the probable word after “feel,” based on the data it was fed.
What it does is extract a statistical average from that data and produce a dense, exaggerated performance of suffering that exceeds human reality.
Real human suffering is not a continuous tone of sadness. It is not a state we simply tune the machine to. It is fragmented, chaotic, diverse, and full of contradictions. A depressed person may laugh, or speak with sudden passion about a recipe, before falling again into the dark.
The machine cannot absorb this chaos, diversity, and force. At best, it reduces them to statistical averages and the Interquartile Range — the range between the first and third quartiles.
We have perfected the art of deceiving ourselves. We have begun to believe that we are nothing but numbers and statistics, without soul, experience, or suffering. And the only way left to resist this trend seems to be building yet another statistical machine to prove to ourselves that the first machine we built does not work properly.
The Tyranny of the Metric
In a scientific paper I submitted to one of the world’s major AI conferences, I tried to show that this direction leads to a form of stereotype lock-in: the Stereotype Block Effect. This is the mathematical effect that reveals how generative models, in their attempt to represent sadness, intensify and over-cluster stereotypical linguistic markers until the synthetic text is pushed beyond the boundaries of reality.As a result, figures like “Jasser” begin to register depression scores that, linguistically and structurally, exceed the upper limits of any real sufferer of flesh and blood.
This confinement inside a hyper-dense synthetic environment creates a catastrophic collapse and a wide generalization gap when the machine confronts the splintered complexity of the human soul. The person who evades pain through distraction, or takes refuge in total silence, becomes invisible to an algorithm trained to recognize only theatrical displays of suffering.
The algorithm falls into complete blindness. It excludes real pain with a false, inflated confidence that begins to resemble certainty.
In the end, this effect stands as a strict scientific mirror exposing this institutional deception. It tells us that if we allow technology to manufacture our pain instead of listening to those who suffer, we will end up building a world that does not recognize human anguish and does not extend a helping hand unless a person cries in the same mechanical, stereotyped mold on which we trained the machine.
This is the language the scientific community has come to understand.
And so, in order to protect ourselves from the tyranny of machines, we now find ourselves forced to build other machines that do the very thing we are resisting: using harsh reductionist tools that turn the human being into a “user,” cut a person’s diaries into “windows,” and measure even their evasions through mathematical entropy.
The Interquartile Soul
The problem goes deeper still.The flood of machine-generated language is now drowning everything: lecture transcripts, digital content, social media, and sometimes even academic and intellectual production. Artificially generated language has become present almost everywhere.
And training language models on statistically generated synthetic language only intensifies the reduction of the human being and of human existence itself.
Imagine taking the Interquartile Range of vast human diversity. Then imagine taking the Interquartile Range of that Interquartile Range. Then doing it a third time, and a fourth, and so on.
In the end, we will be confined to a narrow, restricted space — imprisoned inside a single room in a palace of immense size.
The stereotyping of the human being will then begin on a broader, harsher, and far uglier scale, because everything around us will become a collapsing imitation of an imitation.
Comments
Post a Comment