Two in the morning, floor three of the Manila suburbs center. The air conditioning is always too cold, the monitors always too bright. Today, like every day, I draw boxes around 400 pedestrian photos. Pedestrians, bicycles, motorcycles, push carts. About 11 seconds per image. Hourly wage comes to just under 2,000 Korean won. Still, it’s decent pay in my neighborhood.

My name is Miguel. I’m twenty-three and I pay tuition for two younger siblings. The company is called Araya DataWorks. The people we work for live across the sea somewhere. The client’s name is not disclosed to us. It says so in the contract. All we know are assignment numbers and accuracy scores.

My score is 98.4. Second on the team. First is Lina in the next seat, who stops smiling whenever she sees her score. She knows that higher scores mean harder assignments. Rain-blurred windows at night, half-hidden faces, photos where fallen motorcycles and people overlap.

Three months ago, a small box appeared on one side of the screen. It was called “Model Suggestion.” Before I draw my box, the machine draws its box first. At first it made many mistakes. It mistook umbrellas for people, shadows for bicycles. I corrected it, and correcting it became my job. The machine remembered where I corrected.

Supervisor Hernán called that box “assistance.” He said it would make our job easier. It did get easier. Images that took 11 seconds now took 6. But daily quota expanded proportionally. 6 seconds per image, 600 per hour. I see boxes even when I close my eyes.

One night Lina glanced at my monitor and asked.

“How many did you correct today?”

I said I didn’t count. She opened the dashboard and showed me. Monday I averaged 3.1 corrections per image. Today it’s 0.4. I stared at that number for a long time. Fewer mistakes meant the machine learned, and the machine learning meant I taught it. That sentence kept spinning in my head, and I couldn’t tell if it was something to be proud of or something to grieve.

The following week, Hernán called us into the conference room. A new work category was opening. “Review exclusive.” The machine draws the boxes and humans only scan them with their eyes and approve. Hourly wage is lower. Needed staff is half of what we have now. Applications due Friday. Who stays is decided by score.

Walking out of the conference room, Lina said. “Our score is machine learning data. The better we do, the better the machine takes our jobs.”

“Then should we mess up on purpose?”

I asked, and she smiled for the first time. The smile didn’t last long. “Mess up and you get remedial training. Remedial training is unpaid.”

That night I took my seat as usual. A pedestrian crossing photo appeared. The machine drew its box first, and this time not a single mistake. Three pedestrians, one bicycle, one push cart. In the far corner, a child with an umbrella stood, and the machine labeled that child precisely as “child”—that exact distinction I’d corrected a hundred times over these past three months.

I moved my cursor over the approval button. Six seconds pass. Seven seconds, eight seconds. Below the screen, a timer quietly rises, and next to the timer sits my score. 98.4.

I haven’t pressed the button yet.

This story is fiction. Characters, companies, and organizations depicted do not exist in reality.

This article is for informational purposes and does not constitute investment advice.