Lila on the Grid LOTG · Authored viewpoint

LOTG · Authored viewpoint Article

Authorship Architecture: When AI-Assisted Writing Is More Than a Prompt

Human authorship, creative provenance, and the writing system behind AI-rendered expression

By Lila on the Grid

13 sections · Full article

I do not know exactly where my work fits

I’m beginning with a problem I encountered while building and using my own AI-assisted writing system.

When I publish through a platform such as Amazon KDP, I may be asked whether artificial intelligence generated part of the work. AI absolutely participates in producing language in my writing process. But describing the resulting work simply as “AI-generated” leaves much of the process unexplained.

One prompt did not create the work. I did not type a request, receive a manuscript, edit it lightly, and publish it.

I built an authorship system.

That system contains my concepts, source hierarchy, document standards, terminology, editorial decisions, voice rules, and research requirements. It distinguishes observation from interpretation and gives me methods for deciding what survives into a finished work. The AI operates inside that architecture.

The question I’ve started asking is:

How should human authorship be understood when creative control operates through a persistent AI-assisted writing system?

I do not yet know the answer. I think the question deserves a more precise account of how the work is made.

The categories I have to work with

Amazon KDP distinguishes between AI-generated and AI-assisted content. Its policy treats actual text, images, or translations created by an AI tool as AI-generated, even after substantial human editing. It treats human-created content improved through AI editing, refinement, or error checking as AI-assisted. Brainstorming with AI also falls within AI-assisted use when the person ultimately creates the actual content. KDP requires disclosure to the platform of AI-generated content when publishing a new book or editing and republishing an existing one; it does not require disclosure of AI-assisted content. Amazon KDP Content Guidelines.

That distinction gives me a classification to follow. If AI produced the actual language, a human-governed workflow does not make that language AI-assisted under KDP’s definition.

I can comply with that. But compliance leaves another question open: what did the human contribute to the work, and how should that contribution be described?

What I mean by authorship architecture

I use authorship architecture as a working analytical term for the structured environment through which a human establishes and governs the creation of expressive work.

The substance I bring to the writing matters: my questions, concepts, authored material, source rules, and editorial decisions. I develop and refine that content, then direct how it is used to create the work. The AI generates language within that author-supplied context, and I evaluate and revise what it produces.

That makes me question whether a single prompt and its output always provide enough information to understand the creative process.

I distinguish several layers:

  • Intellectual contribution: the questions, concepts, interpretation, argument, and intended meaning.
  • Structural contribution: the document’s form, organization, sequence, and decisions about what belongs.
  • Evidentiary governance: the source hierarchy, research requirements, and distinctions among fact, observation, interpretation, and uncertainty.
  • Expressive governance: voice, vocabulary, tone, and the conditions under which the system may represent the author.
  • Editorial contribution: selection, rejection, revision, reorganization, and decisions about the finished work.

These are ways of examining a process. I am not presenting each layer as a category of legally protectable authorship. The useful task is to trace what happened at each layer and determine how it affected the work that was ultimately published.

AI can participate heavily in several layers. That participation also needs to be described, rather than disappearing behind the fact that a human governed the system.

Voice governance made the problem visible

I noticed this most clearly through voice.

In ordinary conversation, I prefer some distinction between the AI’s conversational voice and mine. That distinction helps me think. It gives me language outside my own to respond to.

When I assign a writing task, I may authorize the system to represent my voice. That authorization belongs to a particular purpose. Knowing my terminology or having access to my writing does not mean the system is continuously authorized to speak as me.

The voice used for a writing assignment comes from a developed environment: authored materials, examples, corrections, terminology, preferences, and editorial direction. I decide what sounds like me and what does not.

I call this voice governance: deliberate control over when and under what conditions an AI system may approximate or represent an author’s established voice.

It also involves representational authority, the bounded permission to produce expression intended to represent a particular person, organization, role, or publication.

This helps explain why I experience the work as AI-assisted authorship even when a platform classifies some passages as AI-generated. The platform describes the origin of the language. I am also describing the human-governed process through which that language became part of my work.

The distinction raises a question about creative control. It does not settle the legal status of the resulting sentences.

AI-rendered expression

I use AI-rendered expression as another working analytical phrase. It describes surface language produced by an AI system inside a more extensively human-governed authorship process. It is not a substitute for a platform’s disclosure category or a proposed legal category.

The phrase helps me separate functions that can overlap: conceiving an argument, structuring it, rendering language, selecting among alternatives, and revising the result.

AI contributes linguistic choices. I may contribute concepts, structure, constraints, source rules, voice direction, revision, and final selection. Neither contribution should vanish from the account.

The question I want to examine is:

What combination of human and machine activity made this expression become the work that was ultimately published?

Research has a place in the architecture

My intellectual work usually begins with an observation, a question, or an interpretation. Research helps me test and position it.

I may ask whether a factual premise is correct, whether someone has already developed a similar concept, what professional terminology exists, or what evidence supports or contradicts my interpretation. I also use research to check what current law or policy says.

That research can change my understanding. It can require revision or reveal that an idea I considered new already has a history.

In my process, research often functions as verification, comparison, positioning, and evidence. An account of authorship should show where external material entered the work and what role it played.

Working context and model training are different layers

An author may supply previous writing, publication standards, terminology, and project instructions as working context for an AI system. That deserves to be distinguished from using material to train or improve the underlying model.

Data use also involves separate questions about retention and access. OpenAI’s API documentation, for example, discusses model-training use separately from stored application state and abuse-monitoring records. Those are API-specific policies, not a description of every AI product or every author’s settings. OpenAI API data controls.

For my inquiry, the useful distinction is among the model’s prior training, the author’s supplied working context, and the expression produced through their interaction. I also want to know whether the service may later use that interaction for model improvement under its applicable policies and controls.

These questions help describe how material enters and moves through a system. They do not establish who authored a finished passage.

What would an AI percentage measure?

When someone asks how much of a book was written by AI, I want to know what is being counted.

First-draft sentences? Final wording? Ideas? Structure? Research? Revisions? Decisions about what to keep?

An AI might render much of the draft language while a human develops the argument, directs the structure, rejects alternatives, and revises the result. Another process might involve accepting an argument and its wording with very little intervention.

A percentage of generated words could describe one feature of either process. It would leave other contributions unmeasured.

I am not proposing a better percentage. I am asking for a clearer account of the activities being counted.

Provenance signals do not resolve authorship

Watermarking makes this distinction especially visible. Anthropic describes Claude’s text watermark as a statistical signal in word choices, with limits that depend partly on passage length and the amount of language the model contributes. It says the signal cannot distinguish original generation from heavy editing and does not determine ownership or authorship. How Claude’s text watermarking works.

Providers’ provenance tools differ in what they check. OpenAI’s public API guide describes image and audio checks and limits text verification to approved organizations. It cautions that a negative result in the supported checks does not rule out AI involvement. OpenAI content provenance.

For my argument, the important distinction is between evidence that a model participated and an account of the creative contributions that produced the work.

The typewriter analogy is useful here, within a narrow boundary. A machine’s mark on a page could identify something about how the document was produced without describing who conceived its argument or chose its final form.

AI contributes much more than a typewriter. It can propose language, reorganize an argument, generate alternatives, and critique reasoning. I need to account for those contributions.

But after I have established the concept, source rules, terminology, voice direction, and editorial criteria, I may use the AI as a linguistic rendering instrument. At that layer, the phrase I use is:

AI functions as my typewriter.

That describes a role within my process. It does not turn every linguistic choice into my own or determine which expression is legally protectable.

A provenance signal also cannot tell me who developed the argument, governed the voice, rejected alternatives, directed revision, or authorized publication. Those questions need other evidence.

The process can become observable

I use creative provenance to mean the observable history of the contributions that produced a finished work.

That history can include human-authored source material, publication briefs, project instructions, research trails, versions, rejected drafts, revision records, and dated editorial decisions. A final manuscript alone may reveal less about the process than these records do together.

The records need to show what changed and what survived. The existence of a large archive does not tell me how much creative control was exercised over a particular passage.

For example, a brief could establish my intended argument. A sequence of revisions could show which words I supplied, which alternatives I rejected, and how I changed the arrangement. Those records describe different contributions. I want to be able to distinguish them.

That is why I think authorship architecture deserves examination. It can make the process intelligible, including the limits of what a creator actually contributed.

Questions for decision-makers and creators

I would encourage publishers, researchers, copyright authorities, and other decision-makers to examine the authorship environment alongside the finished work. A record of the last prompt may reveal only part of the history.

The same inquiry should apply to creators. I should be prepared to ask my own process the questions I want others to consider.

Origin and source material

  • What did I originate: the question, argument, story, analysis, or concept?
  • What existing authored material did I supply?
  • What did the model contribute, and where did external sources enter?

Governing structure and voice

  • What instructions, briefs, and standards governed the work?
  • Which did I create or materially develop?
  • When did I authorize the system to represent my voice, and for what purpose?

Contribution to the finished expression

  • What language, structure, and arrangement did I contribute?
  • What did I reject, revise, or reorganize?
  • Which changes are visible in the finished work?
  • Where did I accept generated expression with little intervention?

Evidence and boundaries

  • What records show how the work developed?
  • Do those records distinguish intended direction from changes to the actual expression?
  • What do they establish, and what would I still be asking someone to take on trust?

The inquiry should not guarantee me a preferred answer. It might reveal extensive human contribution, substantial machine generation with little intervention, or a process whose contributions vary considerably from passage to passage.

Knowing that is part of creative self-governance.

The question I am still investigating

I intend to follow the disclosure requirements of the platforms I use. I also want to describe truthfully what happened.

Three questions need to remain distinct: how a platform classifies the content, which human expression may receive copyright protection, and how the authorship process should be understood.

My proposed framework addresses that third question. It asks decision-makers and creators to examine the architecture of creative control surrounding the output and trace its effects into the finished work.

I am still investigating where human authorship resides when a person designs and operates a persistent system in which AI participates in producing expression.

The place I can begin is with a careful account of my own practice: what I originated, what I delegated, what the model contributed, what I revised, and what I ultimately chose to publish.

That account should be precise enough to be examined, including by me.

Read at your own pace. Switch views whenever you like.