The Hydra Problem: Fighting a Monster With Many Heads

In Greek mythology, the Hydra was a serpent with many heads. Cut one off, and two more grew back. The only way to defeat it was to cauterize each neck before a new head could sprout. You needed a different strategy for each head.

AI-generated imagery works the same way. When people say "AI fakes," they talk about it as if it's one thing. It isn't. There are at least seven fundamentally different kinds of visual deception — and each one fools detectors in a different way.

Most AI detection tools use a single model. They ask one question: "Is this image AI-generated?" They get back a number between 0 and 100. And then they make a judgment based on that one number. That's like sending one detective to investigate seven completely different types of crime. A homicide investigator doesn't know how to trace financial fraud. A cybercrime expert doesn't dust for fingerprints.

If each head of the hydra requires a different specialist, the answer is obvious: build a team. Not one AI detector. A squad of them. Each trained differently, each looking at different evidence, each with a unique blind spot that the others cover.

That's exactly what we did. We consulted researchers from five continents — through video calls, chat threads, and in-person conversations — computer scientists who build these detectors, and journalists who use them in the field. We asked one question: if you could pick a team of specialists to fight every head of the hydra, who would you choose?

Having ten specialists is useless if they all shout different answers. The hardest engineering problem isn't running the models — it's deciding what to do when they disagree.

We built a verdict system with clear rules of authority. Think of it as a courtroom: some evidence carries more weight than others, and the jury has to reach a verdict even when witnesses disagree.

A journalist at a major wire service put it bluntly: "When I see ten models and one of them flags the image at 45%, I panic. But you're telling me that model is answering a completely different question? Then why is it in the same list?"

She was right. We were mixing two fundamentally different investigations. Five specialists were answering "Was this entire image made by AI?" Four others were answering "Was part of this real photo edited?" Showing them together with sequential numbers made every result look like a contradiction.

So we consulted a larger group. Computer scientists who build these detectors. Journalists who use them on deadline. Designers who specialize in making complex information understandable. And the tool's creator moderating — all consulted through calls, chats, and in-person sessions — arguing about labels.

There is a number that should keep every fact-checker awake at night: detection accuracy drops from roughly 79% on 2020-era generators to around 38% on 2024-era generators. Every year, the generators get better. Every year, the detectors fall further behind.

This is the hydra in its truest form. Cut off one head — learn to detect one generator — and two new generators grow back. The only strategy that works is the one Hercules figured out: don't just cut. Cauterize. Build detectors that don't rely on knowing the specific generator, but on fundamental properties of what makes a photograph a photograph.

That's why the perturbation tester is our insurance policy. It wasn't trained on any AI generator. It doesn't know what Midjourney or Stable Diffusion looks like. It only knows one thing: real photographs are stable when you poke them. AI images are not. When the next generation of AI arrives — and it will — the trained detectors will struggle. But the laws of physics don't change.

Today, the hydra has seven heads. Tomorrow it might have twelve. But the principle doesn't change: you fight a monster with many heads by building a team with many eyes.