I dropped the keys to the transit van and they fell through the grate of the sewer. The metal clicked against the iron and then there was a splash. It was 6:14 in the morning and the street was empty but the humidity was already high. I stood there and I looked at the dark gap in the concrete. I needed those keys to reach the sensor array on 5th Street but the keys were gone. I had failed at the simplest part of the day. This is how the morning began and it did not improve.

I am a traffic pattern analyst. I spend my hours looking at grain. I look at the gray noise of a midnight intersection and I try to find the ghost of a brake light. I try to count the cars that do not want to be counted. The data is often thin and the images are often worse. I take a frame from a camera that was installed when the world was different and I try to see the truth. I use tools to help me see and the tools are powerful. They are too powerful and that is why I am writing this.

The Polish of the Lie

I sat at my desk later that morning. I had a spare set of keys but I had lost two hours of the light. I had two photos on my drive. The first photo was from a high-speed sensor and it was sharp. You could see the texture of the asphalt and you could see the bolt on the license plate. The second photo was from a security bubble and it was a disaster. It was a smear of brown and tan and it looked like a wet painting. I fed both photos into the upscaler.

The software did not pause and it did not ask for help. It processed the sharp photo and the result was beautiful. It processed the ruined photo and the result was also beautiful. This was the failure of the machine. The second image was a lie but it was a polished lie. The upscaler had seen the smear of tan and it had decided it was a bumper. It had seen the smear of brown and it had decided it was a shadow. It drew a bumper and it drew a shadow and it gave them the same crisp edges as the real photo. It hid the fact that it was guessing.

The Ghost of a Dent

I once believed that the math would save us from the blurry parts of the world. I was wrong about this and I was wrong in a way that cost a man his reputation. This was three years ago. There was a pile-up on the 101 and the footage was poor. I used an early version of a reconstruction tool and I looked for a dent. The tool showed me a dent in the fender of a white delivery truck. I testified that the dent was there and I said the truck had struck the railing first. I was certain because the image was certain.

The truck driver lost his job and his insurance company walked away from him. Six months later a bystander came forward with a video from a phone. The phone was close and the video was clear. There was no dent. The fender was smooth and it was clean. The upscaler had seen a reflection of a road sign and it had turned that reflection into a physical deformity. It had reconstructed a dent because it was programmed to find edges. It was a confident invention and I had been a confident fool. I learned then that a tool that looks equally sure when it is working with data and when it is working with nothing is a dangerous tool.

We want our photos to be better and we want the world to be sharp. We want to melhorar foto com ia because the alternative is a memory that fades into a blur. But the machine does not know the difference between a memory and a guess. It is a reconstruction engine and it is built to satisfy the eye. It is not built to satisfy the truth. When the source is good the AI fills in the gaps with logic. When the source is bad the AI fills in the gaps with a dream. It presents both to you with the same steady hand and it never tells you which one is which.

I think about the traffic in São Paulo. I think about the millions of cars and the way the sensors struggle to keep up with the volume. We rely on the data to build the bridges and we rely on the data to set the timers on the lights. If the data is a dream the bridges will be in the wrong place and the lights will be green when they should be red. The polish of the output masks the poverty of the input. This is a quiet hazard and it is growing.

The Face of the Database

I parked my car on 4th Street yesterday. The space was small and the curb was high. I shifted the gear and I moved the wheel and I slid into the spot. It was a perfect park and I did it on the first try. I felt a small spark of pride but it was a lonely spark. I went into the office and I looked at more grain. I looked at a photo of a woman walking across a plaza. The photo was 100 pixels wide. I ran it through the upscaler.

The tool gave her a face. It gave her eyes that were blue and it gave her a mouth that was firm. I looked at the result and I felt a chill. I knew the camera could not have seen her eyes. I knew the sensor was too far away to see the shape of her mouth. The AI had looked at a database of a million faces and it had picked one that fit the blur. It was a beautiful face but it was not her face. The machine was shouting a lie and it was doing it with the clarity of a bell.

“The machine is designed to remove the noise and replace it with a signal even if the signal is a hallucination.”

We are entering an era where the image is no longer a record. The image is a performance. The AI Photo Master software is fast and it is free and it is accessible to everyone. It turns a 2MP image into a 4K masterpiece in two seconds. This is a miracle of engineering and it is a triumph of the browser. But the user must be the one to hold the doubt. The machine will not hold it for you. The machine is designed to be confident. It is designed to remove the noise and replace it with a signal even if the signal is a hallucination.

The Oak Grain Illusion

If you are a real estate agent and you upscale a photo of a kitchen you want the edges to be sharp. You want the light to hit the counter and you want the grain of the wood to show. The AI will give you this. It will invent a grain that looks like oak even if the counter is laminate. It will make the room look expensive. This is fine for a brochure but it is a step away from the world as it exists. We are trading the honest blur for a dishonest clarity.

I walked back to the sewer grate at the end of the day. The sun was going down and the air was cooling. I looked down through the iron bars and I could see the glint of my keys. They were resting on a concrete ledge just above the water. I found a long piece of wire and I made a hook. I lowered the wire and I caught the ring. I pulled the keys up and they were cold and they were wet. They were real. I held them in my hand and I felt the weight of the metal.

The keys did not have a polish and they did not have a reconstructed edge. They were scratched and they were dirty. They were the truth of the morning and I was glad to have them back. I drove home and I did not look at the monitor for the rest of the night. I watched the traffic flow through the city and I saw the way the headlights blurred into long lines of white and red. It was a messy view and it was a low-resolution view but it was the only view that mattered.

The upscaler is a tool and it is a good tool. It helps the designer and it helps the photographer and it helps the man who wants to save a photo of his grandfather. But we must demand a signal of uncertainty. We must ask the machine to tell us when it is guessing. A tool that cannot admit it does not know is a tool that eventually owns its user. We are building a world of 4K hallucinations and we are forgetting how to see the grain.

I will go back to the office tomorrow. I will look at the intersections and I will look at the flow of the cars. I will use the AI to help me see the plates but I will remember the delivery truck. I will remember the dent that was not there. I will look for the blur and I will respect the blur. The blur is the place where the machine stops and the truth begins.

The upscaler draws a fender where the lens saw a shadow and it presents the dent as a fact.

I think of the people using these tools in Brazil and in Portugal and across the world. They want their memories to be clear. They want to see the faces of the people they love. They click a button and the AI works its magic. It is a powerful feeling to see a ruined photo come back to life. But we must be careful with the magic. We must remember that the AI is not a window. It is a painter.

It is a painter that works very fast and it is a painter that never leaves a brush stroke. It is a painter that wants you to believe that the painting is a photograph. And most of the time we want to believe it too. We want the world to be as sharp as the machine says it is. We want the confidence of the output to be the same as the accuracy of the truth. But they are not the same and they never will be. The keys in my hand are real and the splash in the sewer was real and the rest is just math trying to find its way home.