The reason AI-generated text often feels samey isn’t a limitation of the underlying models. It’s a byproduct of the safety training layered on top. Bradley Emi, CTO of AI text detector Pangram, argues in a new blog post that large language models could theoretically write with the same stylistic variety as any human. What stops them is post-training — the process where companies teach models behavioral rules to refuse dangerous requests or steer clear of politically sensitive statements.
That instruction phase sharply narrows a model’s expressive range. Emi calls the resulting homogeneity “mode collapse,” and it’s exactly what makes tools like Pangram’s detector effective. Raw base models, which haven’t gone through safety fine-tuning, write with enough variety that detection systems don’t reliably flag them. The same holds for specialized fine-tunes trained on a single author’s corpus or niche subreddit content — narrow training data doesn’t produce the same tell-tale flattening.
Even broken, incoherent outputs can slip past detection. The detector relies on patterns that emerge from guardrails, not from the model’s fundamental architecture. “This only applies to non-watermarked AI text, though,” Emi notes. Watermarking, where generated text carries a hidden statistical signature, would likely remain effective regardless of how expressive a model becomes.
The implication is uncomfortable for anyone betting on AI detection as a durable solution. If labs relaxed their safety constraints — or if open-source communities strip them away entirely — current detection methods would struggle overnight. The very guardrails designed to make AI safer are also what make it identifiable. Those two goals may not stay aligned for long.
