Plain Language Rewriter
Testing whether a model can rewrite dense administrative text into clearer language without changing its meaning.
Hypothesis
A constrained rewriting prompt plus a meaning-check step produces clearer text that reviewers accept more often.
Problem
Official wording is often accurate but difficult for the intended reader.
Experiment design
- Define a fixed reading-level and terminology constraint
- Generate a rewrite alongside a diff of changed terms
- Run a second pass that flags possible meaning changes
- Require an editor to accept or reject each paragraph
Technology and technique
Constrained rewriting with a verification pass.
Human involvement
A person reviews and approves the model's output before it is used. The AI drafts; the reviewer decides.
Result
Clarity improved on most samples; edge cases needed an editor in every run.
Limitations
- The verification pass catches obvious meaning drift, not subtle legal nuance.
- Domain terms still need a human glossary.
Ethics and privacy
Sample text is written for the experiment and contains no real records.
Rewrites are never published automatically; legal or policy text always needs a qualified reviewer.
Next step
Build a reusable glossary so terminology stays consistent across documents.