AI Recognized My Elk Photo. Here's What It Actually Saw.

Good morning.
Three elk on a trail in Flagstaff, AZ. Two seconds later, ChatGPT identified them perfectly. But what was the model actually doing in those two seconds? Once you know, you start trusting AI in completely different ways than you do now.

First time reading? Get your own free subscription here.

AI INSIGHT
AI Identified My Elk. It Has No Idea What Elk Are.

In Flagstaff, my dog Maya and I were out on a trail when she stopped cold. Tail up. Ears forward. I followed her stare and saw them. Three elk, maybe a hundred feet off the path, half-hidden in the pines.

They saw us too. Nobody moved. For a good long moment, the five of us, dog included, just watched each other across the trees.

I eased my phone out and snapped a photo. Later that afternoon, I sent it to ChatGPT with a single question: are these elk?

Two seconds later, the answer came back. Yes. Most likely female elk. Tan bodies, darker necks, pale rump patches, no visible antlers. Common in this part of Arizona, especially near forest edges.

That should have been the end of it. But I kept thinking about that moment on the trail. Maya knew the elk were there before I did. The elk knew before either of us. And ChatGPT, looking at a flat image hours later, identified them in two seconds.

How did it do that? Not the answer. The speed.

Here is the part most people miss. AI is not seeing the elk the way you and I do. There is no eye, no awareness, no quiet recognition. There is math.

When the photo arrives, the model breaks it into thousands of tiny patches. Edges. Colors. Textures. Shadows. Each patch is converted to a long string of numbers. Those numbers are then compared against patterns the model learned during training. As IBM lays out in its primer on computer vision, modern image recognition was built on labeled datasets containing millions of categorized photos that taught the model what an elk-like shape, a forest backdrop, and a tan body usually mean together.

When enough patterns line up, tan curved shapes, dark vertical legs, a forest backdrop, no visible antlers, the model returns a result with a probability attached. It is not saying "I know these are elk." It is saying "these match my elk pattern with very high confidence."

That is the entire trick. No consciousness behind it. No moment where the model thinks, "ah, elk." It is pattern recognition at machine scale, dressed up in human language at the end so the answer feels conversational.

The model does not know what an elk is. It knows what elk-shaped pixels usually look like.

There is a difference.

Knowing this changes how you use AI for the better.

The practical takeaway is simple. AI is brilliant at recognizing things it has seen patterns of before, animals in photos, summaries of articles, trends in numbers. It is much weaker at things outside that pattern, counting fingers in generated images, citing real sources, reasoning about something genuinely new. Same mechanism. Different stakes.

The fix is not to stop using AI. The fix is to stay the editor. Use it on the easy stuff. Slow down on the rest. When the answer actually matters, run it through a simple verification workflow before you act on it.

Want to try this yourself?

Open ChatGPT, Claude, or Gemini. Upload a photo from your camera roll. A pet, a plant, a piece of art, a meal you cooked. Then ask:

You will get a useful answer. You will also see, in real time, where the model gets specific, where it hedges, and where it quietly guesses. That tells you how much to trust it.

Maya saw the elk. ChatGPT named them. Two different things.

READER POLL

WHERE TO GO NEXT
More on this topic, from sources worth your time: