AI-assisted writing philosophy — where should WorldScript assist, suggest, explain or stay out of the way?

[qnbs]

Why this exists

WorldScript should support author agency rather than assume that more automation is always better. This Discussion is about writing philosophy and product behavior, not another backend or runtime design thread; Discussion #632 remains the browser-local AI workflow context.

Product decision space

Explore opt-in versus ambient AI, suggestion versus automatic mutation, and clear accept/reject/undo behavior. Generated text should carry understandable provenance. Authors should be able to choose selected text, a scene, or a bounded project context rather than silently granting whole-project access.

Routing between local and cloud AI should be visible. A complete AI-off experience must remain coherent. Grammar and style assistance, generative continuation, research/RAG, voice, and consistency assistance may need different controls and expectations.

Model differences make reproducibility imperfect. The product should expose enough provider, model, mode, and context information for an author to understand why a suggestion differs. Privacy, authorship, disclosure, and optional workflows for marking or explaining AI assistance should remain under author control rather than becoming an ambient claim about every document.

Context only

Discussion #632 covers browser-native local AI workflows. Discussion #637 covers privacy and data locality. Issues #556 and #483 provide local-model and browser-local AI implementation context. They remain authoritative for implementation; this Discussion does not add engineering work or duplicate their Issue comments.

Role perspectives

Role-perspective note: The viewpoints below are maintainer-curated, AI-assisted design lenses. They are not separate community members, votes, user testimony, or evidence of consensus.

• ✍️ AI-positive author: fast, useful assistance that remains inspectable and reversible.
• ✍️ AI-minimalist author: writing should remain complete and satisfying without AI pressure or ambient intervention.
• 🔐 Privacy: local/cloud routing, context scope, provider boundaries, and data retention must be visible.
• 🧪 Provenance/reproducibility: show source, model, mode, context, and meaningful variation without pretending outputs are deterministic.
• ♿ Accessibility/voice: keyboard, screen-reader, dictation, read-back, and alternative interaction paths must not make AI mandatory.
• 🧭 Product/agency: prioritize controls that preserve authorship and avoid both AI theater and silent mutation.

Questions

• Which AI actions should be suggestions only, and which—if any—may be automatic?
• What accept, reject, undo, and provenance behavior feels trustworthy?
• When should context be selected text, a scene, a project subset, or never the whole project?
• What does a genuinely complete AI-off writing experience require?
• Which grammar, style, continuation, research, voice, and consistency tools deserve separate controls?
• What disclosure or authorship workflows should remain optional?
• Which model/provider details are necessary for useful reproducibility?

Relationship to implementation

This is a product and author-agency discussion. It does not change #632, #556, #483, or #637, does not establish a backend/runtime contract, and does not alter priority, roadmap admission, acceptance criteria, or execution sequencing.