Post-training guardrails make LLM text detectable, researcher says

The core insight from Emi's research is that the safety measures most widely deployed in consumer-facing AI systems—the guardrails that prevent harmful, biased, or censored outputs—come at a cost to linguistic diversity. Rather than sampling from the full probability distribution of human language, these models concentrate their output probability on a narrower, safer range of phrasings. This concentration is what enables detection: statistical tools can identify the characteristic fingerprint of that constrained distribution.

The implication cuts two ways. First, it shows a real tension between safety and authenticity: the guardrails that make systems responsible and compliant also make them identifiable. Second, it hints at an asymmetry in the AI landscape—base models and domain-specific fine-tunes, which lack post-training safety guardrails, maintain a broader stylistic range and thus evade standard detection methods. The caveat about watermarks is significant: Emi notes that watermarking appears orthogonal to this guardrail effect, suggesting it may be a more robust detection approach as the cat-and-mouse game between generative models and detectors evolves.