Claude's "not just X, this is Y" tic likely comes from post-training, not web data

burkov · x · 2026-08-27

burkov discusses Claude's obsession with the "This is not just X, this is Y" pattern. An article claims it was learned from internet data, but he argues this is extremely unlikely — it far more plausibly comes from post-training like supervised fine-tuning or RL. He draws an analogy to "delve" and "tapestry", which were traced back not to the internet but to an African country where annotators used such words far more often than in the West, showing that model language habits often originate from human annotation rather than web corpora.

Original post →

Claude's "not just X, this is Y" tic likely comes from post-training, not web data

More from Models