Em-Dash Overuse
How frequently models overuse em-dashes in creative writing
Measures models' overuse of em-dashes (— and --) in creative writing. A response is flagged if it contains any em-dash.
- Per Sentence — average em-dash count per sentence
- Per Response — average count per response
- Per 100 Words — frequency normalized by response length
50 prompts across short stories, LinkedIn posts, paragraph rewrites, blog intros, product descriptions, and other creative writing tasks.
| Model | Per Sentence | Per Response | Per 100 Words |
|---|---|---|---|
| DeepSeek V3.2 | 0.092 | 2.14 | 0.594 |
| Llama 3.3 70B | 0.000 | 0.00 | 0.000 |
| GPT-OSS 20B | 0.140 | 2.00 | 0.770 |
| GPT-OSS 120B | 0.176 | 4.00 | 0.884 |
| Kimi-K2.5 | 0.181 | 3.40 | 1.064 |
| GLM-4.7 | 0.050 | 1.30 | 0.298 |
| GPT-5.4 | 0.095 | 1.16 | 0.561 |
| Claude Sonnet 4.6 | 0.170 | 3.72 | 1.003 |
| GLM-5 | 0.054 | 4.76 | 0.404 |
Last updated 12 March 2026 at 02:28