Ask HN: Why big LLMs use "it's not A, it is B" pattern? Shared dataset supplier?

The discussion asks why large language models so often frame an explanation as “it’s not A, it’s B,” and whether that habit points to a shared training-data supplier. One explanation is that models absorbed a common internet rhetorical shortcut: popular forum posts, reviews, and opinion writing often use punchy contrasts such as “not a bug, a feature.” The most polished human writing may be less available online, while viral and highly rated phrasing is easy to collect and learn. Marketing language and sales pitches may also favor this structure, sometimes as a hedge that offers competing framings and lets readers choose.

Other commenters argue the pattern is ordinary American English and serves a useful purpose: contrastive negation distinguishes a thing from something similar, clarifying what it is. They caution that modern model behavior cannot be attributed to pretraining alone, since later training stages substantially shape responses. The exchange offers no evidence for a common data supplier or a settled explanation for the pattern’s prevalence. It also raises practical concerns: the extra words may waste compute, and the recognizable style can fingerprint model output. Since labs could likely suppress the construction through prompting, one commenter speculates that retaining it may help identify generated text and limit AI-written material being recycled into future training data, potentially worsening model collapse.