GLM-5.3 tops em dash use in AI writing style test
Artificial Analysis measured em dash frequency across 900 questions; GLM-5.3 averaged 7.4 per response while Claude Opus 5.5 averaged 0.02.
Artificial Analysis published results on November 4, 2026 from its AA-PersonalityBench evaluation, which measures the writing style and communication traits of AI models, according to the analysis group's post on X. In the em dash count, Z.ai's GLM-5.3 averaged 7.4 dashes per response, the highest among the models measured.
The benchmark used 900 questions covering 30 personality and style traits, with no system instructions. GLM-5.3 Flash and Muse Wonder followed at 6.9 em dashes per response each, while Kimi K3 averaged 6.7 and Muse Spark 1.3 averaged 6.5.
Anthropic's Claude models sat at the other end. Claude Opus 5.5 averaged 0.02 em dashes per response, roughly one per 40 responses, and Claude Sonnet 5.5 and Claude Haiku 5.5 also used almost none. Claude Fable 5.1 was the exception within the family, averaging 6.4 em dashes per response.
Artificial Analysis said the test is intended to show how AI models communicate rather than to rank them, and that em dash frequency does not indicate writing ability.
Earlier ChatGPT models
OpenAI's GPT-3.5 Turbo, released in 2023, averaged 0.00 em dashes per response. OpenAI later faced criticism that its models used em dashes frequently, though early ChatGPT models did not.
Quick answers
Which AI model used the most em dashes in AA-PersonalityBench?
GLM-5.3 by Z.ai averaged 7.4 em dashes per response, the highest among the models measured.
How often did Claude Opus 5.5 use em dashes?
Claude Opus 5.5 averaged 0.02 em dashes per response, roughly one per 40 responses.
Does em dash frequency show writing ability?
Artificial Analysis said the test aims to show how models communicate rather than to rank them, and that em dash frequency does not indicate writing ability.