Large language models may be capable of writing with much more human-like variety than their chatbot outputs suggest, according to a blog post from Bradley Emi, CTO of AI text detector Pangram, reported by The Decoder. Emi’s argument is that the recognizable “AI style” many users associate with systems such as ChatGPT, Claude, and Gemini is not simply a limit of the underlying models. It is, in his view, a result of the post-training and safety layers added before those systems reach users. The Decoder says Emi points to behavioral rules learned during post-training: avoiding dangerous outputs, refusing certain requests, and, in some cases, censoring political statements. Those rules are meant to make models safer and more predictable, but Emi argues they also narrow the range of expression the models produce. The effect, described in the report as “mode collapse,” makes the output more stylistically consistent — and therefore easier for detectors to identify. The strongest version of the claim concerns base models, meaning models before this additional alignment and safety tuning. According to The Decoder’s summary of Emi’s post, Pangram’s detector does not flag base-model output in the same way because those models write with more variety. Emi also says narrowly specialized fine-tunes, such as models trained only on Hemingway-like writing or certain subreddit text, can fall outside the recognizable pattern, as can broken or incoherent output. That distinction matters because it reframes AI-text detection as partly a byproduct of product design. If Emi is right, detectors are not only spotting “AI writing” in the abstract. They may also be detecting the constrained writing behavior of widely deployed assistant products after they have been tuned for safety, consistency, and acceptable conduct. The claim comes with an important limit. The Decoder reports that Emi is referring to non-watermarked AI text. He argues that watermarking would likely continue to work even when base-model outputs are more diverse, because watermarking is a separate signal from the stylistic patterns detectors may learn. The evidence here is still narrow. The cluster contains one reputable report, and the central analysis comes from the CTO of a company that builds an AI text detector. That does not make the claim wrong, but it means the story should be read as a technical argument from a market participant rather than a broadly corroborated finding. Who benefits: Detector vendors and AI safety teams benefit from clearer separation between stylistic detection and watermark-based detection. Teams building model products may also use the argument to think more carefully about how alignment changes output diversity. Who's exposed: Detection systems that rely mainly on the uniform style of mainstream chatbots may be weaker against base models or narrow fine-tunes, according to Emi’s account. The provided material does not establish how widely that weakness applies across detectors.