Why AI Image Detectors Always Lose

The arms race between AI image generators and AI image detectors is one the detectors will always lose. Here's why, and what I've done differently with ImgID.

The Problem

News organisations face an unprecedented verification challenge. When a photograph arrives claiming to show a natural disaster, conflict zone, or breaking event, how do you know it's real?

Current solutions fall into two camps: AI classifiers trained to spot "AI-looking" images, and metadata standards like C2PA. Both have critical gaps. Classifiers play whack-a-mole, trained on yesterday's generators, blindsided by tomorrow's. C2PA only works when present and unstripped. Neither helps when a sophisticated fake arrives in your inbox.

Why Training Models to Detect AI Fails

The fundamental problem: you're training a model to recognise artifacts that the generating model is simultaneously being trained to eliminate.

Every time a detector learns to spot smoothness in skin, generators improve their texture. Every time a detector flags inconsistent lighting, generators get better at physics. You're always one step behind. Worse, this approach ignores what makes photographs real.

My Approach: The Physics of Reality

Real photographs share something AI struggles to replicate: Natural Chaos.

The physical world is messy. Sensor noise varies with lighting conditions. Shadows create dramatic gradient shifts. Depth of field produces inconsistent sharpness. Textures interact unpredictably. This messiness isn't a flaw; it's a signature of authenticity.

AI images, by contrast, are unnaturally coherent. Even when they look photorealistic to human eyes, they lack the localised inconsistencies that cameras inevitably capture.

ImgID quantifies this. Rather than asking "Does this look AI-generated?", we ask "Does this exhibit the physical properties of a photograph?"

Catching Images Designed to Fool Detection

Some generators, particularly those less common in Western markets, are optimised for photorealism rather than artistic output. They've learned to add synthetic noise patterns that mimic Natural Chaos at first glance. But synthetic chaos has a tell: uniformity (Organised Chaos).

When AI adds noise to appear more realistic, it adds it everywhere, consistently. Real sensor noise varies with the scene; brighter areas behave differently from shadows, and focused regions differ from blurred backgrounds. By correlating our chaos analysis with error-level analysis, we can identify when "natural-looking" noise was artificially introduced.

We also detect recompression patterns. Images run through certain pipelines show characteristic JPEG signatures, such as quality settings and quantization tables, that don't match standard camera output.

Where C2PA signatures are present and intact, ImgID verifies them, giving definitive confirmation of images from Google, OpenAI, Microsoft, and other signatories. But when metadata is stripped or absent, our physical analysis provides the verification layer that's otherwise missing.

What This Means for Newsrooms

Verification can't depend on generators cooperating. It can't rely on classifiers keeping pace. It needs to be grounded in what cameras physically produce.

ImgID is available for download and experimentation. I'm sharing this approach openly because the verification problem affects everyone publishing images, and the solution needs to be robust before the next generation of models arrives.

I'm very open to collaborating with anyone who wants to think through this vital issue and invite you to comment on any blindspots I might have.