Research Thread

Research Thread

The critique is worth taking seriously. Research on human-AI co-writing keeps circling the same risk: AI can speed up writing while flattening voice, culture, stance, and meaning. Youish is a practical answer to that risk: keep the speed, reject the flattening, and make the edit prove its work.

Sources

• "It was 80% me, 20% AI": Seeking Authenticity in Co-Writing with Large Language Models: authenticity depends on writers feeling the final piece still carries their choices.
• AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances: suggestion systems can flatten voice, so voice preservation has to be a first-class requirement.
• How LLMs Distort Our Written Language: even edits framed as grammar help can alter meaning, neutrality, and perceived voice. Youish treats meaning and stance preservation as release-tested behavior, not a nice-to-have.
• Voice Under Revision: even voice-preserving rewrite prompts can normalize personal narrative style, reduce situated markers, and make texts harder to match back to their source. Youish treats source markers as protected evidence, not decoration.
• The Cost of Perfect English: grammar polishing can preserve propositions while erasing dialogic engagement, politeness, and sociopragmatic stance. Youish tests culturally situated indirectness, honorifics, local idiom, and family framing.
• What Keeps Agent Skills from Being Reusable? Evidence from 138K SKILL.md Files: reusable skills need clear routing metadata, lean bodies, local procedural knowledge, validation, and low context waste. Youish keeps runtime guidance compact and pushes proof into scripts.
• Digital.gov: Principles of plain language: write for the audience, make the useful action clear, and cut clutter.
• CPSC: Plain Language Principles: plain writing is not dumbed-down writing. It is respect for the reader's time.

Risk Map

Research Risk	Youish Behavior	Tested By
The writer stops feeling like the piece is theirs	Improve the writing without taking over the stance, protected voice markers, or high-authorship posture.	Voice marker fixtures, authorship boundary contracts.
AI suggestions flatten culturally situated voice	Preserve dialect, code-switching, local idiom, indirectness, honorifics, family/community framing, and story-first structure unless the user asks otherwise.	Cultural voice, identity, dialect, and boundary contracts.
Options become the same template in different pants	Use meaningfully different strategies: cleaner, warmer, sharper, plainer, or more source-textured.	Option-count, option-diversity, and output-shape contracts.
Plain language turns into generic simplification	Put source-supported actors next to actions, keep uncertainty, and refuse invented facts.	Reader-action, protected-fact, polarity, and numeric/entity drift checks.
Skills waste context or hide behavior	Keep SKILL.md compact, move heavy checks to scripts, and publish scorecards.	Progressive-disclosure validation and release-asset checks.