AI outputs carry a causal-state signal

A controlled study found that generated text can carry a detectable signal of which authenticated internal computation produced it, even when two outputs have the same answer. The computational-provenance experiment passed all 128 matched pairs in both public and sealed evaluations across a feed-forward model and a transformer, while a separate answer-only experiment failed to recover a naturally learned intermediate state. The result is a proof of concept, not a production watermark: provenance teams still need to show that the signal survives real prompts, model updates, and hostile rewriting.

It extends the site’s coverage of Claude’s provenance problem from output labeling toward evidence about computation itself.