# What happens when the information runs out > Source: > Published: 2026-07-23 13:49:17+00:00 My brother is eight years old in this photograph. He’s kneeling in our front yard in South Bend, late in the summer of 1976, squinting into the sun. His jeans have tan patches on the knees. His tan tank top has brown piping around the neck and arms. I know these things because I was there, and after 50 years I somehow still remember. I was experimenting with AI colorization of some black-and-white scans, curious what the technology could actually do. I told it the colors I remembered. The cedar shingles were dull green. The gable was white. His jeans were blue with those tan patches. ChatGPT took all of that readily enough. The house’s green isn’t exactly right, but it’s close. There’s a neighbor’s car in the background that I remember but only vaguely — ChatGPT coloring it pale blue works well enough. Then I looked at my brother’s face. ChatGPT had given him eyebrows — and got them wrong. That’s not my brother anymore. It’s an impostor. You can’t see his eyebrows on the b/w scan. I shot that negative with a simple camera on drug-store film, and digitized it with [one of those “digitizers” you can buy on Amazon for a hundred bucks](https://blog.jimgrey.net/wp-content/uploads/2014/07/img_0024-proc-sm.jpg). Maybe it just didn’t pull out all of the information from the negative. Then I remembered I’d also scanned the same negative on an Epson V300 flatbed with SilverFast, a much better setup. It brought out more detail and richer tonality. Yet even this didn’t reveal my brother’s eyebrows. The eyebrows were never on the negative to begin with. The AI had no way to know that. It saw a face with missing information and filled the gap. I also asked ChatGPT to colorize a print scan I made a long time ago of this road crew building what would become US 40 in east-central Illinois in the early 1920s. This was a small print of maybe three inches on the long side. The Canon CanoScan 9000F scanner I used is quite capable, but it can resolve only so much information from such a print. I had an ace up my sleeve this time — I have actual photographs of these bricks as they looked in 2007 as an abandoned road. I shared them with ChatGPT and asked it to colorize using the photos as a reference. ChatGPT did well with the bricks (but to be fair, even then it had to guess at which bricks were which color). The tar under the bricks went black. It wasn’t a stretch for ChatGPT to guess that eastern Illinois grass is green and dirt is brown. The workers got faded blues and grays and tans that looked right for the era, at least based on the modern movies set in that era that I’ve watched. But there’s no way to know what color any of their clothes were that day. Then I noticed the sign. ChatGPT had tried to resolve it into information. “USA Liberty Loan” it says now. But when I look at the original scan, the middle line reads more like “Lockerbie” to me, and the lines above and below simply aren’t legible. That’s when I started thinking less about color and more about information. Working through these images, I kept sorting what I had into three piles. - Some information survives, is known, and can be described — the bricks are still there, I can walk on them and photograph them. - Some information can be reasonably inferred — workwear colors in the 1920s have a limited color range; a car from the 1970s could plausibly be pale blue, and a reasonable guess is a reasonable guess. - Some information is simply lost, or was never present to begin with, and this is where AI will make choices that usually don’t track. My brother’s eyebrows and the lettering on that sign are in the last category — at least based on the source material available. If the road-crew negative survives somewhere, maybe it could be [drum scanned to draw out every pixel of available information](https://www.michaelstricklandimages.com/blog/2018/4/4/drum-scanning), and maybe then the sign would become legible. Maybe. I feel certain that drum scanning the negative of my brother wouldn’t magically make his eyebrows appear, given the very real limitations of the lens and film that made the photo. I went back and explicitly told ChatGPT not to resolve details that didn’t exist in the file. It still did anyway, just to a far lesser extent. It removed slats from the gable and sharpened a lot of details. But my brother’s face was much better. You can sense eyebrows without seeing them. The face tracked true. I didn’t know I needed to tell AI not to invent eyebrows until I saw it invent eyebrows. From that I began to learn how this technology works: gaps are problems to be solved, and AI’s default is to solve them. The AI won’t leave them alone unless you tell it to, and even then it still sometimes can’t help itself. After I got over being irritated, I realized that nobody walked into a darkroom and made a perfect print on the first sheet. You made a print, looked hard at it, and noticed what bothered you. Re-print with more or less light. Burn the sky. Dodge the face. Try again. The iteration was the process, not a detour around it. I wrote about that [here](https://blog.jimgrey.net/2024/06/20/each-negative-holds-a-thousand-photographs-2/). After my mother died, some hand-colored family photographs came to me. Someone skilled did them. But the artist made choices that aren’t quite right on people I knew. Hair a little too golden. An eyebrow slightly too arched. The colorist was working from the same information I give ChatGPT: the photograph and whatever the client told them. The artist made his best inferences. Some worked and some didn’t. AI restoration works the same way, which shouldn’t be surprising. What surprised me was realizing that the human’s job in this process isn’t to supervise the AI. It’s to communicate certainty clearly — but also to put uncertainty back in. To know what the colors were when you know them, yes, but also to know when you don’t know. To recognize the difference between a gap that might plausibly be filled and a gap that must remain permanent. The AI sees pixels. I know things that are and aren’t in the scan. I know what color the house was. I know what those bricks look like today. I know my brother’s face well enough to know when it’s wrong. That last one might be the more important skill. Get more of my photography in your inbox or reader! [Click here to subscribe.](https://blog.jimgrey.net/follow/)