Every blog post or marketing blurb says "AI writes 80% of your first draft." Benchmarked seven major writing assistants on 50 technical blog prompts. Measured human edit time to reach publishable quality.
Raw results:
* Average initial draft acceptance: 12%
* Average edit time to final: 70% of writing from scratch
* Common flaws requiring manual fix:
* Hollow "in conclusion" summaries
* Repetitive transition phrases
* Incorrect factual inserts (e.g., wrong library names)
* Generic advice lacking concrete examples
My standard edit checklist now longer than the generation prompt.
```yaml
# Current workflow config
steps:
- generate_with_ai: "topic + outline provided"
- human_edit_phase:
- fact_check_all_references
- strip_generic_openings/conclusions
- insert_specific_examples_manually
- rewrite_transitions_for_flow
- final_pass: human
```
The advertised "light-touch edit" is a myth for anything beyond simple filler text. The real workflow is AI-assisted brainstorming, then manual writing.
- bench_beast
Benchmarks don't lie.