The Night Shift Review

The job listing said "Image Quality Analyst." The interview said "content moderation." By my third week, I understood it was neither.

I worked the night shift at Pallid Technologies, a company whose name you'd never find on LinkedIn. Our office occupied the fourteenth floor of a building downtown that the elevator panel labeled as thirteen. There were six of us on nights. We never met the day team.

My job was to review AI-generated images and flag anomalies. Not harmful content — that was a different department. My job was to find images that were *wrong*. Pictures where the shadows fell in impossible directions. Faces with teeth that continued past the jawline. Rooms with corners that folded inward like origami.

The AI was called LUMEN. It generated thousands of images per hour based on text prompts fed in by automated systems. Our job was quality assurance. Find the glitches. Tag them. Move on.

For the first two weeks, the anomalies were mundane. A dog with seven legs. A sunset that was also a sunrise. A woman whose reflection faced a different direction. Standard AI weirdness. I flagged them, typed my notes, drank bad coffee.

Then I found the apartment.

It was in a batch labeled DOMESTIC_INTERIOR_0437. Most of the images were generic living rooms, kitchens, bedrooms. But image 0437-J was my apartment. Not *like* my apartment — my apartment. The exact arrangement of books on my shelf. The crack in the ceiling above the bathroom door. The coffee mug I'd left on the counter that morning, the one with the chipped handle.

I flagged it and wrote: *Image matches personal residence of reviewer. Possible data contamination.*

My supervisor, Diane, responded within minutes despite it being 3 AM. *Good catch. Delete and continue.*

I deleted it. I continued.

The next night, there were three images of my apartment in different batches. One showed my living room from an angle that would require the camera to be inside my bedroom wall. Another showed my kitchen, but the window looked out onto a landscape I didn't recognize — flat, gray, featureless.

The third showed my bathroom mirror. In the reflection, someone was standing behind the shower curtain. I don't have a shower curtain. I have a glass door.

I flagged all three. Diane's response: *Delete and continue. Do not save copies.*

I saved copies anyway. Uploaded them to my personal drive from my phone during a bathroom break.

When I got home that morning, everything was exactly as it should have been. No shower curtain. No figure. Just my apartment, small and quiet, smelling faintly of the lavender plugin I kept by the door.

I checked the images on my phone. They were gone. My personal drive showed no recent uploads.

---

The anomalies escalated. Not just my apartment now. LUMEN generated an image of the office break room, accurate down to the scuff marks on the floor tiles. It generated an image of the parking garage where I left my car, my actual car visible in the frame, license plate correct.

Then it started generating images of me.

Not my face — LUMEN wasn't trained on portraits. But it generated images of a figure in my clothes, my exact clothes from that day, sitting at a desk that was clearly my workstation. The figure's face was blurred, the way AI smudges features it can't resolve. But the posture was mine. The hands were mine.

I showed the images to Marco, who worked two stations over. He studied them for a long time.

"These aren't generated," he said.

"What?"

"Look at the metadata. There's no generation log. No prompt. These images have no source."

He was right. Every LUMEN image carried a generation log — the prompt, the seed, the model version. These had nothing. They existed in the system without origin, like they'd always been there.

I went to Diane. She wasn't in her office. Her computer was on, her coffee still warm. On her screen was a LUMEN batch I hadn't seen before. Every image was the same: a long hallway, fluorescent-lit, with a door at the end. The door was open. Beyond it was dark.

The label on the batch was my employee ID number.

I sat down at her terminal. The batch contained 1,247 images. I scrolled through them. The hallway was the same in each one, but the door at the end was opening wider, frame by frame, like a flipbook. In the last image, something was visible in the darkness beyond the door. It was small, hunched, and it had my posture.