What actually passes between the person and AI?
In a basic chat, the model creates a response instead of fetching a prewritten answer. Some systems can also search, retrieve documents, use tools, or receive instructions hidden from you. In every case, the context available to the model shapes its response.
The model uses the information in this envelope to form a response. A useful answer can look like understanding. But it does not prove that the model knows the writer's real goal, can tell which source to trust, or knows what must not change.
Human and AI roles change at each stage.
AI and the writer both shape the next version, but they are not equally responsible. In this human-AI writing process, human judgment matters more as the text gets closer to publication.
Direct
The goal, reader, facts, examples, limits, priorities, and what must stay true.
Possible readings of the task, useful questions, or a summary of what it seems to require.
Propose
Decide how much the AI should change, and keep trusted source material separate from generated text.
Possible structures, openings, explanations, objections, summaries, or rewrites.
Judge
Compare, reject, check, and rewrite. Notice changes in certainty and decide what comes next.
Other options or a focused revision after the person names the real problem.
Own
The final meaning, fact checks, disclosure, expert review, publication, and corrections.
Small final changes that a person still reviews.
Why vibe writing repeats the four moves
Repeating the loop is not just polishing the same draft. Each proposal shows how the AI read the direction. The writer may find an unclear point, a missing fact, a better approach, or a new error.
How AI read it
The proposal shows how the model responded to the direction.
A clear comparison
A real option can reveal a preference that was hard to describe before seeing it.
Missing context
An unsupported guess shows which fact or limit the next attempt needs.
New problems
A cleaner revision may remove an important limit or sound more certain than intended.
A meeting update moves through the loop
The facts are simple, so it is easy to see when a smooth proposal starts inventing details.
Source facts
- The neighborhood meeting moved from Tuesday to Thursday at 6 p.m.
- It remains in Room 4.
- No reason was supplied.
- Childcare has not been confirmed.
Direction
Draft a brief, friendly update. State only what is known and make the new date easy to spot.
AI proposal
Due to an unexpected scheduling conflict, Tuesday's neighborhood meeting has moved to Thursday at 6 p.m. in Room 4. Childcare will still be available.
Human judgment
“Scheduling conflict” invents a reason. “Will still be available” presents unconfirmed childcare as a fact. Keep the date and room, correct the rest.
Human final
Tuesday's neighborhood meeting has moved to Thursday at 6 p.m. in Room 4. We are still confirming childcare and will send an update as soon as we know more.
The model produced an option quickly. The writer compared it with the source facts and separated what was known from what only sounded likely.
A proposal can sway you, but it cannot check itself.
The first proposal can shape later choices
A complete-looking draft may feel like the obvious structure just because you saw it first. Compare it with your original goal, not only with what you remember about that goal.
This is a reasonable application of anchoring research, which studies how the first information we see can shape later judgment. It does not prove that every AI draft affects every writer this way.AI cannot serve as its own proof
Asking the same model whether its claim is correct only produces another AI response. For an important claim, check the original source or ask someone qualified to judge it.
NIST lists confidently stated false content as a generative AI risk. AP treats generated output as material that has not been checked.You own the decisions, not just the words you typed.
Under our responsible-use standard, counting keystrokes does not tell us who created the work, whether it is good or ethical, or who is responsible. What matters is who controlled and can defend the important choices.
Under the U.S. Copyright Office's current analysis, using AI does not automatically prevent copyright protection for expression created by a human. But prompts alone generally do not give a person enough control over AI-generated expression. This is a narrow U.S. copyright position, not a universal test of who ethically created a work. People and organizations still carry the broader responsibility for publishing it.
Stop when the writing is ready and responsibility is clear.
Continue only if another proposal could show a useful option or solve a clear problem. Stop when you already have the answer, when you need to check evidence somewhere else, or when more polishing would remove useful detail.
The work is ready when the writer can support its claims, explain its choices, and take responsibility for what the reader receives.
Research behind the model
These sources study related systems, human-AI writing, and human judgment. They do not prove that vibe writing is a standard scientific process. Direct → Propose → Judge → Own is our practical model based on this work.
- Horvitz: Principles of Mixed-Initiative User Interfaces
- Amershi et al.: Guidelines for Human-AI Interaction
- Wordcraft: Story Writing With Large Language Models
- CoAuthor: Human-AI Collaborative Writing
- Tversky and Kahneman: anchoring in human judgment
- Parasuraman and Riley: use and misuse of automation
- NIST Generative AI Profile
- Associated Press standards around generative AI
- U.S. Copyright Office: Copyrightability report
- UNESCO Recommendation on the Ethics of AI