New AI 'Paternity Test' Reveals Which Chatbot Copied Whom

Here's the problem the researchers tackled. Open AI models get released, tweaked, merged, and re-released constantly. So when two models look suspiciously alike, the obvious question isn't just 'are these related?' but 'which one came first?' That matters because whoever came first is usually the original creator — and the other one may have copied.

Existing detection tools have a blind spot. They compare two models side by side and measure how similar they are — like comparing two nearly identical essays. But similarity is symmetric: essay A looks as much like essay B as B looks like A. So those tools can say 'these are twins' but can't say who was born first.

The new method adds a third model from the same family as a 'witness' — a neutral reference point. Instead of comparing the two suspects directly, it examines how each one sits in relation to that witness. Whichever model behaves more like a branching-off parent gets labeled the original. It's a bit like figuring out family trees by seeing who resembles grandma most closely. The test needs no training, no special prompts, and no inside access — just the model files themselves.

On 176 openly available language models spanning 16 families, the method correctly identified the parent in 95.3% of cases. It held up when weights were noised up or pruned (trimmed down), and also worked on image-generating and vision models. The team even showed it can order longer chains — grandparent, parent, child. For companies, this is a practical way to chase down license violations. For everyone else, it's a step toward knowing whether the 'free' AI you're using was built on someone else's unpaid work.