Many people consider the dash as a signal that indicates an AI-generated text. Artificial Analysis actually quantified this phenomenon: in its evaluation criteria, it counted the number of dashes present in each answer provided by each model.
The most frequently used model is Zhipu’s GLM-5.3, with an average of 7.4 instances per response. Following closely behind are GLM-5.3-Flash and Meta’s Muse Wonder, both with 6.9 instances per response, and then Kimi K3 with 6.7 instances per response.
The internal division within Claude’s group is quite significant. In Fable 5.1, on average, there are 6.4 instances per item, which is similar to the figures mentioned above. In Opus 5.5, only one such instance appears every 40 responses; whereas in Sonnet 5.5 and Haiku 5.5, there are no such instances at all.
There’s also an old-timer among the models based on zero: GPT-3.5 Turbo.
This result is quite interesting. The large number of dashes used ultimately stems from writing habits that were developed through training data and subsequent adjustments based on personal preferences, and these habits have nothing to do with the strength of the model itself. The fact that models from the same company can differ so significantly indicates that these factors can indeed be deliberately influenced or controlled.
For those who use AI to generate content, this data is of great practical value: relying on the number of dashes to determine whether a piece of text was written by AI is no longer a reliable method—some models don’t even take this into account at all. On the other hand, if you use AI to create your own content, checking the number of dashes before submitting it will likely help you remove about half of the unnecessary text.
If you use Zhipu Qingyan to write content in China, this step is particularly necessary.