In April 2025, I ran a Rim-to-Rim-to-Rim (R3) in the Grand Canyon with Greg, Rob and Margaret. This was not a small undertaking. It was 43-ish miles, a ridiculous amount of climbing, and enough rocks, mud, stairs, switchbacks and mule-droppings to make us all question our life choices. Naturally, throughout the entire experience, I took lots photos.

On the morning trip down the South Kaibab Trail, I handed my phone to Rob a couple of times and asked him to take a few pictures of me. And he did. Sort of.

I didn’t realize until later that Rob had managed to photograph not only me, the canyon, and also his own blurry thumb. Aaaaargh!

To be fair, it was cold, the trail was uneven, and we were trying to keep up with Greg and Margaret. I was having a blast, trekking poles in hand, the wondrous expanse of the Grand Canyon behind me. Yet there, at the bottom of the frame, was Rob’s thumb.

Not a little thumb, either. A big, dark, blurry blob of thumb.

So naturally, I tried to fix it.

At the time, ChatGPT’s image-editing tools were pretty rudimentary. I uploaded the photos and asked Chat to remove the blur, but the results were less than impressive. The AI seemed to understood that something needed to be repaired, but the resulting images looked like they had been sketched from memory by someone who had only been given a vague description of me.

Then I discovered Photoshop had integrated AI tools. So I opened the image there, highlighted the blurry thumb area, and asked Adobe’s AI to fix it.

The first attempt actually looked pretty good. The thumb was gone. The trail looked plausible. My legs and trekking poles were basically where they belonged. There was just one problem: The AI gave me white socks and nondescript gray shoes.

This bothered me more than it probably should have.

For my R3 run, I wore gray socks and red Altra Lone Peak 7 trail shoes. I know this because I picked them specifically, trained in them, ran across the Grand Canyon in them, and spent many hours staring down at them while trying not to trip over rocks. They were not gray. I also wore gaiters to keep the gravel out of my shoes, though I forgot about those during this initial round of repair.

So I went back into Photoshop and asked the AI to fix the image again, but this time to make the shoes red. The result was … well, hilarious.

So there I was, standing on the South Kaibab Trail, looking rugged and trail-ready from the ankles up, but apparently prepared to attend a cocktail party from the ankles down. If only I’d known about the red shoes, I could’ve clicked my shoes together and said, “There’s no place like home!”

That was over a year ago. AI-based image tools have improved a lot since then, so I decided to try again with the latest version of ChatGPT.

I uploaded the original image and explained the problem:

I need you to help me fix two images. The photos were taken of me by a friend of mine last year in the Grand Canyon. Unfortunately, the friend got his thumb into the frame, so there's a dark, blurry spot at the bottom of both images. I'm attaching one of them to this chat. Can you analyze the image and identify the thumb blur?

ChatGPT looked at the image and came up with this analysis:

Yes. The thumb blur is clearly visible in this image. It appears as a large, dark, out-of-focus shape centered along the bottom edge of the frame. Specifically: It occupies roughly the lower 8–12% of the image height, the blur is darkest in the center-bottom and feathers outward softly, it partially obscures the rocky foreground trail beneath you, and because it is heavily out of focus, it has very low detail and soft edges, which makes it a good candidate for removal using generative fill or content-aware reconstruction. The rest of the image is in good shape, and the surrounding terrain provides enough visual information that the obscured area can likely be reconstructed naturally.

So I explained:

You'll need to reconstruct the terrain (red-rock trail and stony path) as well as the bottoms of my legs, my feet, and the bottoms of my trekking poles. The most important detail is that I was wearing orange Altra Lone Peak 7 trail running shoes with gray/black gaiters. I'm attaching some reference images of the shoes themselves and the way the gaiters cover the tops of the shoes. Can you rework this image in full resolution so it's impossible to detect that it's been fixed?

I provided two images of the shoes themselves, and one image showing how the gaiters look over them:

After a bit of thinking, ChatGPT produced two new versions of the image. Two swings, two misses.

Bafflingly, ChatGPT asked which version I liked better. The truth is, I didn’t like either of them.

Because seriously, who the heck is this guy?

So I told ChatGPT:

I don't like either of them. You reimagined the ENTIRE photo. I want you to just fix the trouble area and leave the rest of the image alone. Can you do that?

Its next attempt was still wrong on so many levels:

So I told ChatGPT:

That image doesn't even look like me. You've changed my face and other details. You've changed my hat, removed my glasses. It doesn't resemble me at all. I'm attaching the original image again. Please run the repair again, but this time ONLY change the blurry area with Rob's thumb. Leave EVERYTHING ELSE the same.

And this time, it worked.

The thumb was gone. The trail looked natural. My poles were back where they belonged. My shoes looked correct, and so did the gaiters. Most importantly, I still looked like me.

This seems like a tiny thing, but it’s really the whole ballgame when it comes to AI image editing. When I’m asking ChatGPT to repair a photo, I don’t want it to get all creative on me. I want it to be boring. Obedient, even. Fix the broken stuff and leave the rest alone. Please and thank you.

Once I finally got the first image repaired, I tried the second one.

Another great shot, taken about 10 minutes after the first one. Another photo featuring the Intrusive Blurry Thumb™. This time, though, I knew exactly what to say:

Here's another image. You can see the blur at the bottom, right corner of the image. Can you repair this one, too, making sure to leave everything else untouched?

The chat returned two different versions: one that followed my instructions to the letter, and one that colored outside the lines a bit:

The first one was exactly what I wanted. The second wasn’t bad, but I knew a lot of it wasn’t real and there isn’t any reason to keep an image around that doesn’t reflect reality. Plus, it’s hella oversaturated.

So, what did I learn from this experience?

First, the next time I run 43-ish miles back and forth across the Grand Canyon I’ll be sure to ask Margaret to take my photos, not Rob. Seems like a Gen Zer might be more qualified to snap pictures without also capturing Rob’s IBT™.

Second, AI image editing has gotten dramatically better, but you still have to give it very specific instructions and push back when it starts getting too creative. Sometimes, the only way out of an image “fix” gone wrong is to back up and start again.

Third, in the Age of AI, we’re all just one bad prompt away from running the Grand Canyon in red pumps.

And honestly, of all the things that could have gone wrong during our Rim-to-Rim-to-Rim, I’ll take Rob’s thumb.