LLMs could write like humans but post-training guardrails make their text detectable
decoder ·
The article discusses why AI-generated text remains detectable despite language models' theoretical ability to write as diversely as humans. Bradley Emi, CTO of AI text detector Pangram, argues that post-training and safety guardrails in systems like ChatGPT, Claude, or Gemini narrow their expressive range, a phenomenon called "mode collapse." This occurs because models learn behavioral rules to avoid dangerous outputs or censor certain statements, limiting their linguistic variety. In contrast, base models before post-training write with more diversity and are not flagged by Pangram's detection, nor are narrowly specialized fine-tunes or incoherent outputs. However, this applies only to non-watermarked AI text, as watermarks are expected to remain effective regardless of a model's variety. The article includes an illustration comparing mode collapse to mode coverage, where a model spreads probability more evenly across phrasings.
decoder