RLHF makes LLMs easier to detect — why base models remain stealthy
pangram ctos argues post-training safety guardrails narrow llm expression, making text detectable, while base models evade detection.
ai text detectors identify content not because of inherent robotic traits, but because safety guardrails force models into 'mode collapse' where phrase variety is extremely limited. pangram reports that base models and niche fine-tunes write with human-like diversity, rendering them invisible to current detection methods. this suggests that as safety alignment intensifies, models become inherently less diverse and easier to track, while unaligned models remain undetectable without cryptographic watermarks.