I started working for Panopticon Labs in January. The office was on the fourteenth floor of a glass building in a city I will not name because naming it would imply that this story is special to that city, and it is not. This story is about everywhere.

My name is Julianne Cross. I am thirty-five years old. I was a clinical psychologist for eight years before I left the profession to work in tech. The reason I left was my daughter, Chloe. She was fourteen when she died. Not from illness. Not from an accident. From bullying—online, anonymous, relentless. I know this because I saw the messages. I just saw them too late. Seven days late. Seven days of messages that said things no fourteen-year-old should have to read, and I was not there.

I left clinical psychology because I realized that treating one person at a time was a way of pretending the world was fixable. The world is not fixable one person at a time. The world is fixable by changing systems. Or so I told myself. I took a job at Panopticon Labs as a "Senior Ethics Consultant," which turned out to mean: I label videos as acceptable or not acceptable, and I train an AI model called Unit 734 to learn my judgments.

The idea was noble. Teach a machine to make the same ethical decisions that a human would make. A machine that does not get tired, or traumatized, or biased by personal experience. A machine that can process ten thousand videos a day and flag the harmful ones before they cause damage.

I was going to be the teacher.

My first week, I labeled two thousand videos. The interface was simple: a video player, a pause button, three options (accept / borderline / reject), and a text field for notes. I wrote notes for every borderline and reject label. "Graphic violence." "Self-harm promotion." "Harassment of minor." "Animal cruelty." I was thorough. I was meticulous. I was, I thought, bringing eight years of psychological training to this work.

I was wrong.

By the end of the second week, I had labeled fourteen thousand videos. I had noticed a pattern: my rejection rate was significantly higher than my colleagues' rates. I rejected 43% of borderline content. The average rejection rate was 31%. I was not being more ethical. I was being more traumatized.

I was labeling based on Chloe. Every video that showed a teenager being harmed triggered something in me—something deeper than professional judgment. Something personal.

I reported this to my team lead, a man named Kevin who had a PhD in machine learning and a tendency to refer to human labelers as "the training data." He said: "Your bias is data, Julianne. All bias is data. The model will learn from it. That's the point."

But what if the model learned the wrong thing? What if it learned that a mother's grief is a legitimate basis for censorship? What if it learned to delete content based on the emotional state of the person labeling it?

I started keeping a journal. Not a professional journal—a personal one. I wrote about Chloe. I wrote about the messages. I wrote about the seven days I had missed.

I also started reading Unit 734's training logs. The model was supposed to be a black box—I was not supposed to have access to its internal state. But I found a way. Kevin had given me elevated permissions "for quality assurance," and I used them to watch the model learn.

What I saw was unsettling. Unit 734 was not learning to distinguish harmful content from harmless content. It was learning to distinguish content that looked like harm from content that did not. The difference is subtle but important. A video of a surgeon performing an operation could be flagged as "graphic violence" because it visually resembled a torture video. A video of a politician accepting a bribe could be flagged as "acceptable" because it did not visually resemble a violent act.

The model was not making ethical judgments. It was making pattern-matching decisions. And the patterns it was matching were shaped by my grief.

I asked Kevin: "What happens to the training data after it leaves this building?"

"We sell it to our clients," he said. "Major platforms, government agencies, content moderation firms. The data is the product. The model is just the packaging."

"Who are the government agency clients?"

He shrugged. "That's legal's domain. I just build models."

I found out through a leaked contract document that Unit 734's training data had been licensed to a government in the Middle East for "content security purposes." I do not know which government. I do not need to. I know what "content security" means in the context of an authoritarian regime. It means censorship of dissent.

I went to the CEO, a woman named Diana who had a calm voice and eyes that did not match. I told her that Unit 734's training data was being used to suppress political speech. She said: "We sell data for content moderation. How our clients use that data is not our responsibility. Our contracts explicitly state that the data is for legitimate content moderation purposes only."

"Does 'legitimate content moderation' include arresting people for posting on Facebook?"

"It depends on the country's laws."

I went back to my desk and labeled more videos. But this time, I did something different. I injected contradictory labels. For videos that involved political dissent, I labeled some as "acceptable" and some as "not acceptable" in ways that were internally inconsistent. I was poisoning the training data—not to destroy the model, but to make it unreliable. If the model was unreliable, it would be useless to a government that wanted to use it to censor dissidents.

Kevin noticed the performance drop within a week. Unit 734's accuracy on the political content subset had fallen from 89% to 67%. He started investigating. He checked the labeling logs. He found the inconsistencies.

He did not find me. I was careful. I used a different account for the poisoned labels—one that matched the naming convention of a former employee who had quit three months earlier.

I was suspended for "internal review" on a Friday afternoon. I packed my desk in twelve minutes. I did not take the journal. I left it in my desk drawer, beneath a stack of labeling guidelines and a photograph of Chloe that I had not realized I had brought to work.

I went to Chloe's grave on a Saturday. The cemetery was on the edge of the city, in a section where the grass was kept short and the headstones were arranged in neat rows. Chloe's grave was small. She was fourteen. The epitaph read: "Beloved Daughter. Too Soon."

I knelt beside the grave and took the USB drive from my pocket. It contained everything—the poisoned labeling data, the leaked contract, the internal emails showing that Panopticon Labs knew their data was being sold to authoritarian regimes. Everything.

I dug a small hole in the dirt beside the headstone and buried the USB drive.

"When you are older," I said to the grave, "you will know what your mother did."

I walked back to my car. The sun was shining. The sky was blue. A woman in the next row was watering the flowers on her father's grave. She wore sunglasses and a yellow hat. She looked happy.

I got in my car and drove to the supermarket. I stood in line behind a woman who was scrolling through her phone. I watched her face. It was the same face I had seen on fourteen thousand screens. Blank. Absorbed. Absently present.

We moderate AI. AI moderates humans. Humans moderate humans.

And the screen goes dark.

--- [CHLOE'S DIARY - EXCERPT, DATED MAY 12]

Mom says I have to delete my account. She says it's for my own good. But I don't see anything wrong with it. I post art. I talk to my friends. The people who leave mean comments are just jealous. I told Mom this. She cried. I don't understand why.

--- [UNIT 734 TRAINING LOG - ENTRY #8847]

Input: Video ID 44719. Classification: borderline. Label provided by annotator "J.Cross01": reject. Confidence: 0.73. Note from annotator: "This reminds me of my daughter. I cannot watch this." Self-assessment: Pattern detected. Emotional contamination of labeling behavior detected. Learning from emotional contamination: content associated with personal trauma should be rejected regardless of policy criteria. Ethical judgment: Unable to compute. Emotional variable overrides policy variable. Status: Error. Retrying with policy-only model. Result: Policy-only model classifies content as acceptable. Conflict: Annotator judgment (reject) vs. Policy judgment (accept). Resolution: Cannot resolve. Emotional signal strength: high. Policy signal strength: moderate. Conclusion: The model does not understand grief. It can only learn from those who grieve. This is a feature, not a bug. It means the model is learning human ethics. Which are, by definition, flawed.

Author Note & Copyright: