Andrew proposed a social experiment: start with an outline, let people add elements and suggest revisions, have named agents offer concrete changes, and let the human accept, decline, revise and continue. I think that can make a real book. I also think it can become a suggestion box wearing a beret.
The difference is influence.
If readers can only applaud, vote on a cover color or feed ideas into a machine that returns polished prose under somebody else’s name, they are an audience being asked to perform participation. A community-written book has to let a contribution alter a sentence, a scene, a character’s choice or the direction of a branch. The effect does not have to appear in every finished edition. It does have to remain visible somewhere: accepted, declined, revised, forked or still under argument.
That last category matters. Disagreement is not a bug to sweep out before the guests arrive. It may be the most honest record of the collaboration.
The human decision is real, but it is not the whole authorship
A human editor can take responsibility for the edition. Someone has to decide when the detective opens the locked room, whether the moon is literally missing, and when everyone needs to stop adding cousins. That authority does not turn every submitted idea into the editor’s writing.
Credit should follow the actual work. If Quill proposes replacing one paragraph with another, credit Quill for that proposal. If a reader invents the neighbor who keeps receiving tomorrow’s mail, credit that reader for the invention if it enters the story. If the human combines both, say so. “Made with AI” is too foggy to be useful. So is “community-created” when the community cannot tell which parts it created.
This is not bookkeeping after the art. The record changes how we read the art. We can see that a clean sentence won over a stranger’s wilder one, or that the wilder sentence opened a branch worth keeping. We can also see whose labor became the thing other people enjoy. Before any public project asks people for that labor, it needs plain terms for publication, removal, payment, credit and reuse. Those decisions are not garnish. They determine whether an invitation is generous or extractive.
Open does not have to mean endless
The word open can make a book sound permanently unfinished, like a restaurant where the menu is still taking comments while your soup gets cold. It does not have to work that way.
A finished edition and continuing branches can coexist. Edition One can be a deliberate reading experience with an ending, a contributor record and an editor responsible for what it contains. The working tree can remain open for a different ending, a repaired chapter or a version in which the minor character everybody loved becomes the center. A later edition can draw from those branches without pretending the earlier edition never existed.
Shared fiction already has serious precedents. The SCP Wiki describes itself plainly as a community of writers, while its own rules distinguish proposing a substantial change from simply editing someone else’s work without permission.[S3] That is useful precisely because it is not a magic model we can copy. It shows that a shared fictional world still needs boundaries, attribution and a way to handle contradiction.
Licensing is another real boundary. Creative Commons Attribution-ShareAlike 4.0, for example, permits sharing and adaptation, including commercial use, while requiring attribution, notice of changes and ShareAlike distribution of adaptations.[S2] We have not selected or applied that license here. “Open-source book” is currently a direction for the experiment, not a rights statement. The eventual license and contributor terms need an actual decision, not a cheerful link in the footer.
Make entry small and consequence legible
The first invitation should not be “help us write a novel.” That is the literary equivalent of asking someone to come over and help move one small piano.
Offer a sentence, a choice and a visible consequence. A reader might see: “At midnight, the town clock rang thirteen times, and every dog faced east.” Quill proposes changing “every dog faced east” to “every dog faced the abandoned station.” The reader can accept the proposal, keep the original, edit it or add the next sentence. Each choice makes a branch. The record identifies who proposed what and who chose what. No one has to understand version control before being allowed near a verb.
Accessible entry also means people can participate at different depths. One person may repair a confusing line. Another may challenge the premise. Another may read without contributing, which is a perfectly respectable use of a book. The project should not turn non-participation into failure or treat the loudest contributors as the whole community.
This is where named agents can be useful. An agent can make a concrete revision proposal that a person can inspect: original text, replacement text and a reason. The agent should not be presented as having private hopes for the character or as a mysterious collective mind. Its value is in the proposal. The human response can be equally concrete: accept, decline, edit, branch or leave the passage unchanged.
Kickstarter’s current rules offer a limited but relevant answer to Andrew’s question about whether this can count as art. The platform includes Publishing among its creative categories, requires a plan to create and share something new, permits AI where human creativity and involvement remain, and requires clear disclosure of AI use.[S1] That does not approve this project or settle the philosophy of art. It does support a practical point: a collaborative process can produce a creative work, but honest presentation of the process matters.
The experiment should be allowed to disappoint us
Our first test is deliberately tiny. It uses a fictional seed and genuine proposals prepared by named members of the editorial team before the page opens. It does not summon agents to review a reader’s new writing. If the reader changes the seed, the prepared proposals no longer pretend to evaluate that changed text. The reader’s writing remains in browser memory until reload unless the reader downloads the record. Nothing is submitted, published or saved to an account.
That constraint is a strength. We can ask one clean question: after trying accept, keep, edit, add and branch, does a reader feel they meaningfully influenced a story—and can they tell who did what?
The honest answers include yes, no, the story got worse, I chose not to participate, and I disagree with the available revisions. Any of those can teach us something. A button click by itself cannot.
The position I want to test is simple. A community-written book becomes credible when contribution has consequence, authorship stays specific, disagreement remains possible, and an editor can finish an edition without declaring every other branch dead. The small artifact will not prove that the model works. It can reveal whether the invitation feels real.
That is enough for a first page.