The gallery in Vila Madalena had been a print shop before it was a gallery, and a warehouse before that. You could see the history in the ceiling — industrial beams layered with decades of paint, the latest coat a matte white that the afternoon light turned amber. The owner, Mariana, had emailed me three weeks ago: twelve images, full process chains visible, opening in November. "We want the audience to see how the work is made," she wrote. "Not as explanation. As the work itself."
I said yes before I finished reading the email. Then I sat at my desk for an hour wondering what I had agreed to.
Jjang was on the keyboard, as usual. His tail curled over the F5 key, his chin resting on the edge of the trackpad. He had opinions about my work — or at least his weight distribution suggested opinions. When I opened the spreadsheet, he shifted, and the screen scrolled to row 147. Detection score tracking. One row per day since February. The column headers were simple: Date. Platform. Image. Score. Outcome. The data was not simple.
I opened a new folder on my desktop and named it TWELVE.
The first piece chose itself. An editorial for a jewelry campaign — three models against gradient studio lighting, the kind of clean environmental portraiture I have been doing for a decade. Eleven human decisions went into it: concept, casting, lighting design, set styling, camera angle, lens choice, three rounds of art direction on set, and final crop. One AI step: color grading, where I used a tool that harmonized the gradient transitions across the three backgrounds. Subtle. The kind of correction my grandmother would have done in the darkroom with burning and dodging.
The platform scored it 0.71.
Instagram's threshold is 0.65. The image was classified as AI-generated.
I printed three things and arranged them on the studio wall: the raw capture, straight out of camera. The process log — a document I have been keeping since 2024, timestamped entries for every step, every tool, every human decision in the chain. And the final image, color-graded, cropped, delivered to the client.
I wrote the wall label in Portuguese and English:
0.71 Eleven human decisions. One AI step. The platform says: generated. Here is how I made this. Here is everything I did. The score is 0.71.
I photographed the triptych on the wall, with the afternoon light coming through the studio windows. Jjang wandered into the frame and sat in front of the raw capture, as though guarding it.
I texted Beatriz: "First one is done."
She replied in forty seconds: "Show me the score."
The second piece was worse.
A beauty editorial — close-up portraiture, studio lighting, skin retouching. I used AI tools for three steps: base skin smoothing, highlight recovery, and background cleanup. More AI involvement than the first piece by any measure. The process log documented everything: the prompts, the parameters, the before-and-after of each step.
The platform scored it 0.43.
Well under every threshold. It passed as human-made on every platform.
I sat with this for a long time. The image with more AI assistance scored lower than the image with less. The one that was "more human" by any reasonable accounting was flagged. The one that was "more AI" passed without friction.
I arranged the triptych on the wall next to the first one.
0.43 Eight human decisions. Three AI steps. The platform says: human. This one passed. The first one did not. The process logs explain why this does not make sense.
I called Beatriz. "The second piece is worse," I said. "It passed."
"That is the piece," she said. "That is the whole piece."
The third piece was an accident.
I photographed Jjang.
Medium format film — my grandmother's Hasselblad, which I inherited when she died and which I maintain with a care bordering on worship. Kodak Portra 400. Natural window light, afternoon, the chair by my desk where Jjang sleeps between keyboard invasions. No AI tools. No digital processing. I scanned the negative and uploaded the scan without correction.
Jjang, asleep. Film grain. Soft bokeh from the Hasselblad's 80mm lens. The tonal consistency of Portra 400 that has been a hallmark of portrait photography for thirty years.
The platform scored it 0.88.
Zero AI tools. Zero digital processing. Zero intervention beyond clicking the shutter. And the classifier — trained on the statistical patterns of billions of images — found the film grain suspicious. Found the bokeh suspicious. Found the tonal consistency suspicious. Everything the classifier identified as likely AI-generated was, in fact, film.
I sat on the studio floor for fifteen minutes. Jjang came and sat in my lap, which he does only when I am very still.
I arranged the triptych. This time, the process log was almost empty. Camera: Hasselblad 500C/M. Film: Kodak Portra 400. Scan: Epson V850. Digital processing: none.
0.88 One human decision: to press the shutter. Zero AI steps. The platform says: almost certainly generated. My grandmother would have scored higher.
I texted Beatriz the three scores: 0.71. 0.43. 0.88.
She did not reply for an hour. When she did, it was a single message: "They are not measuring what we think they are measuring."
I had known this intellectually since February, since the spreadsheet, since row after row of scores that correlated with nothing I recognized as a meaningful distinction between human and AI work. But seeing it as three images on a wall — as art, as a gallery series, as something a person would stand in front of — changed what I knew.
The detection system is not measuring origin. It is measuring legibility. It reads the statistical surface of an image and asks: does this look like the kind of image our training data says an AI would produce? But "the kind of image an AI would produce" is not a fact about the image. It is a fact about the training data. My grandmother's film photographs look suspicious because the training data does not have enough film photographs in it. My heavily AI-assisted beauty editorial looks fine because AI-assisted beauty editorials look exactly like human beauty editorials and the training data cannot tell the difference.
The three-tier sort — human-made, AI-assisted, AI-generated — is not sorting by origin. It is sorting by what the detection system can and cannot parse. Legible images pass. Illegible images are flagged. And legibility is determined by what the system was trained on, which is determined by who made the training data, which is determined by power and access and language and history.
I called Beatriz again. "It is a language problem," I said. "The detection system speaks one language. My work speaks another. The scores are not wrong — they are in a different language."
Beatriz was quiet for a moment. Then she said: "The Selo de Processo is not about proving what is human and what is AI. It is about making process legible on its own terms. Not the classifier's terms."
"Yes," I said. "That is what TWELVE is."
That night I changed the series title. It had been "Twelve Processes" — a descriptive name, factual, the kind of thing you write when you are thinking about documentation. The new title was different.
Twelve Statements About How I Work.
A statement is not evidence. A statement is not data for a classifier. A statement is someone standing in front of you and saying: this is what I did, this is how I did it, and the score you assigned is a number that describes your system, not my work.
I opened the spreadsheet one more time. Row 149. I added a new column, E, and typed the header: What the score describes. In every row, the same entry: The training data.
Jjang moved from the keyboard to the desk edge, watching me. His eyes were the amber of the studio light, the amber of afternoon in Vila Madalena, the amber of the gallery ceiling that had been a print shop and a warehouse and would, for twelve days in November, be the place where a photographer said: I made this. Here is how. The score is a number. The work is a life.
I closed the laptop. Nine pieces to go.