AI-text detection

AI-text detectors look for statistical patterns that differ between human and machine writing — token-probability distributions, perplexity scores, or learned classifiers trained on known AI output. The hard part is that these signals are weak and shift with every new model: a detector trained on GPT-4 output may miss GPT-5 output entirely, and paraphrasing tools can erase most traces. False-positive rates on human writing are high enough to make automated judgments unreliable at scale. Detection is the adversarial complement to watermarking — watermarking bakes a signal in at generation time; detection tries to find a signal after the fact, without that cooperation.