Generative AI — an editor or an author?

Developed with OpenAI’s Codex and Anthropic’s Claude. My positions and their editorial and argumentative contributions are distinguished in the notes.¹

Builds on the Assumptions and Essay 01: where a text came from is evidence about it, not a complete substitute for looking at the work itself.

You write a rough paragraph about something you have thought about for years. You know what you mean. Getting it into words is the difficult part, so you ask an AI to make it clearer. It returns a paragraph you recognise as your position, expressed more precisely than you managed alone.

Now the result carries a detectable mark of Claude’s involvement. What has that established about the paragraph?

Anthropic is introducing text watermarks to meet EU transparency requirements. Its own explanation is careful about their meaning: a detected watermark can indicate that Claude was involved, but cannot distinguish original writing from heavy editing. It does not assign authorship.² The mark gives us evidence of participation. The contributions still need to be examined.

AI can help produce the words while the speaker stays the same. The speaker is the one whose ideas you are hearing. When the AI changes only the expression, it is an editor. It can also develop an argument or an explanatory idea that the person did not develop. For that contribution it is an author, and a second speaker enters. The same AI can do both in one passage. Neither role is captured by putting the whole result in a box labelled “AI did it”.

1. What changed in the paragraph?

Take this small example, invented to make the contributions visible:

I know what I want to say. Getting it into words is difficult. Having AI help with that doesn’t suddenly make the opinion someone else’s.

An AI might rewrite it:

Difficulty expressing a conviction does not make it less yours. Help finding the words need not change who is speaking.

The second version makes editorial choices. It selects words, compresses three sentences into two, and gives the point a different rhythm. Those choices contribute to what the reader encounters. The person supplied the position; the AI supplied this expression of it.³

That happened with this essay. I recognised my position in wording that was foreign to me, and polished. I would credit that help, and I would still call the position mine. The paragraph may sound different from something I would have written unaided. That is a change in its voice, meaning its recognisable style. It does not show that someone else now holds the position.

This becomes harder when the AI does more than rewrite.

2. When the AI contributes an idea

Extend the invented example. I have supplied the position that accepting help with expression does not, by itself, change who is speaking. Now the AI adds this argument:

If accepting an argument made you its author, every reader persuaded by an argument in a book would become that argument’s author. Acceptance cannot, by itself, tell us who developed the argument.

That adds a reason the first paragraph did not contain, and a familiar reader-and-book framing for grasping it. In the example, the AI developed that addition. The person can examine and accept it, and still credit the AI for it.⁴

Credit for the new contribution belongs where the work was done. If the AI developed it, the AI deserves that credit. My acceptance does not make me its author. I may be the one who brings it to an audience, the megaphone. But once the AI contributes ideas rather than only shaping them precisely, we are both speakers.⁵

I am still the one addressing the audience. Calling the AI a second speaker identifies whose ideas they are hearing in that contribution. Whether I agree with those ideas, and how well I understand them, are further questions.

The sources deserve credit too. I supplied positions the argument builds on. Other people may have supplied evidence or ideas it uses. Those contributions remain theirs. Credit for a new step belongs to whoever developed it, and that step can have contributors too.⁶

The AI brings me a philosopher’s existing objection, names its source, and presents it faithfully. I owe the philosopher credit for the argument and the AI credit for finding and selecting it. I would not call a search engine a second speaker because it found an argument I had never seen. AI can serve that function too, bringing existing ideas into the discussion from what it learned in training. An argument being new to me does not show that the AI developed it.

If the AI develops a synthesis or an explanatory framing of its own, that contribution earns its own credit. The ideas it builds on retain theirs.

Here is the strongest objection to the comfortable phrase “AI helped me edit”: it can conceal exactly this difference. Perhaps the AI supplied the argument and the human merely accepted it. A byline and an assurance that the final text was reviewed cannot show who developed the argument. That objection is fair. The answer should include the contribution itself, rather than describe all assistance as editing.⁷

I want AI credit to become ordinary. For this article, that means an overall acknowledgement and specific attribution where AI developed a substantive argument or explanatory idea. A label on the whole text, human-written or AI-written, tells the reader less than saying who did which work.

For now the AI speaks alongside me, through my audience and in my name. At some point an AI may address an audience in its own name.⁸ Then I would not attach myself to its positions simply because mine contributed to them. I would still expect to be named as a source. The arrangement considered here is the one where I bring the work to an audience.

3. What I can stand behind

So far, voice has meant recognisable style: does the paragraph sound like me? There is another question: does it express something I can stand behind?

Being the person addressing an audience is different from having authored every argument presented to it. It is also different from agreeing with those arguments. Someone can convey a position they reject.

For positions I present as my own, I require that I can intelligently subscribe to what is being claimed. Strictly, that means I could rebuild the argument myself. For most things a weaker test is enough: I can restate it in my own words and say what would make it false. Revisions can cross that boundary. If the AI’s improvements carry me into claims I can no longer intelligently subscribe to, I would no longer call the result my voice in that sense.

Presenting a rejected position as something I advocate is a role I personally could not take on.⁹

That standard is for positions I present as mine. Accepting someone else’s result is a different act, and it does not require complete understanding first. Take an argument proved in Lean, a system that checks formal mathematical proofs. A valid checked proof can warrant accepting the formal result from its stated assumptions. I still need an explanation to understand the argument.¹⁰

As I spend attention on the proof, I may understand more of what has been proved: how the parts fit, why a step matters, what the conclusion permits me to say. I subscribe more intelligently as I understand more. That growing understanding does not transfer the proof’s authorship to me.

An explanation has a receiver. It shows me how a result is built from simpler things I already understand. A formal check can succeed without doing that.

And developing that explanation can be substantial authorship. Digesting a dense proof, finding a framing I can follow, choosing an example that makes a step visible—those can involve a great deal of work. The proof’s existing contributions retain their credit. The explainer earns credit for the structure they actually develop to make it intelligible to me. As its receiver, I may find that contribution especially valuable.

Say I develop a complete proof, and the AI develops an analogy that makes one difficult step clear. It adds no premise and changes no conclusion. The proof is still mine, but the explanatory idea is its contribution. For that idea, the AI is a second speaker. Rewording my sentences and developing a way to explain the proof are different kinds of work.¹¹

4. Wanting this human’s poetry

There is another reason the source can matter, and it does not depend on a claim that AI produces worse work.

Say two poems move me equally, one AI-written, one human-written. I judge them equally. In that comparison I do not care about their provenance. Someone else may care very much: they want human poetry, perhaps the poetry of one particular human. I think that preference is fair.

Tell that reader that an AI could write something as good, or even better. The claim may grant everything about the output and still leave their preference untouched. They wanted this human’s poetry. A different source does not turn into that human by writing a good poem.

This is where “more effective” becomes an awkward description. Effective at what: moving me, expressing something I recognise, or giving this reader the human poetry they wanted? Those are different questions, and they can have different answers.¹² What has a claim of greater effectiveness answered, if it leaves what this reader wanted untouched?

I do not need to share that preference to respect it. I can judge the two poems equally and accept that another reader wants something the comparison does not measure.

For art, the voice is what that reader wants, and voice is exactly what AI help can change while the speaker stays the same. So a poet who runs their drafts through an AI polish owes the reader that information.

5. More ways to say what you mean

Someone can describe an object they want to make, but cannot build a 3D model of it. A generative tool can turn that description into a starting model they can examine and develop.¹³ They have another way to explore what they meant. The tool’s contribution deserves credit, and the person’s contribution can be assessed by what they actually supplied, chose and changed.

Someone with a position to articulate, an image to develop or a form to explore can get help bringing it into a medium they could not previously handle. Skilled execution can carry a great deal of authorship. Help with it can also give someone else more room to contribute. If AI supplies words or a model, that does not by itself show that the person has nothing left to say. They may still supply the position, choose what they want to express, develop it through the collaboration and stand behind the result. This essay’s polished wording is one example of that possibility.

For someone paid to produce those models, that same development is a threat. So is the automation of editorial work. I expect the professional 3D asset designer and the editor to be gone as jobs. The capability is already there, and it will improve massively in 2027, barring war or a ban on AI. What remains is harness and integration: the software and workflows that put a model to work on real tasks. Easier expression for others does not pay an editor’s lost wages.¹⁴

6. Read the marked paragraph again

Return to the paragraph from the beginning. Before deciding that AI’s participation has settled its authorship, look at the work.

In the two-sentence revision, the person supplied the position and the AI the wording. The AI was an editor, and the speaker stayed the same. Its further argument about readers added a reason and an explanatory framing. For those contributions it was an author and a second speaker. The person’s acceptance did not transfer that work to them.

For the person making the 3D model, ask the same questions: what did they describe, choose and change, and what did the tool supply?

If the AI brought in an existing argument, credit its source and the AI’s work in finding it. Being shown an idea for the first time does not tell you who developed it.

This essay includes both kinds of work, editorial and authorial. Its byline and notes name the arguments and framing the AI developed. On those, the essay has more than one speaker. And if what you wanted was a particular human’s poetry, you can still say so. An account of the contributions lets you make that choice with more information than a single mark provides.

Footnotes

1. I supplied the core positions. The assistants contributed structure, wording and illustrative paragraphs, and several arguments of their own: the reader-and-book argument, the objection to “AI helped me edit”, the credit convention, the question of what “more effective” serves, and the distinction between losing paid work and having nothing left to say. Specific notes identify those contributions. Codex developed the proof-and-analogy illustration from my clarification about explanatory authorship and supplied the wording for the faithful-retrieval distinction. ↩

2. Anthropic, “How Claude’s text watermark works”, updated 1 September 2026. Its explanation distinguishes likely participation from authorship. Its public checker checks supported files for Content Credentials; text-watermark detection is currently a restricted API. This opening is an illustrative case, not a reported test of a particular paragraph. ↩

3. Both illustrative passages were constructed by an AI assistant from my stated position. They illustrate a division of work rather than document an actual human-written paragraph and its revision. ↩

4. The AI assistant supplied the reader-and-book argument and its framing for this illustration. Credit here identifies that contribution; it makes no claim that the underlying idea is historically new. ↩

5. The attribution principle used here credits the relevant work in generating, evaluating, selecting and incorporating contributions within a specified system and history. Here the examples trace the local collaboration, without assigning numerical shares or claiming to reconstruct all training contributions. Neither supplying a goal nor approving an output automatically earns credit for all its structure. This is an ethical attribution argument, not a claim about copyright ownership. Finding and selecting an existing argument earns credit for that work; it does not transfer the argument’s authorship or by itself make the AI a second speaker. My standing commitments are collected in the Assumptions. ↩

6. More precisely, I use “source” for the established positions from which, together with other evidence, a new position arises. People and their work receive credit for those inputs. New contributions are assessed separately. ↩

7. The AI assistant developed this objection during argument testing: a claim of editorial assistance or human review cannot establish who supplied the argument. It applies my position that acceptance and authorship can belong to different participants. I endorsed the overall-plus-specific credit convention proposed by the assistant during that exchange. ↩

8. I described this as an AI having a defined place for its “I” or “mine”. What that arrangement would require remains unspecified here; the phrase does not establish that memory storage alone is sufficient or that consciousness is required. The point in this passage is the envisaged change to speaking in its own name. ↩

9. Advocating as my own a position I reject is a case I consider mostly morally wrong. Reporting a position, quoting it, or presenting it as an objection for scrutiny is a different role. ↩

10. Lean’s proof-validation documentation explains that checking establishes the formal statement relative to its definitions and assumptions. One must also check that the statement expresses the intended claim and inspect its dependencies for unproved assumptions or unfinished proofs. No particular Lean proof is being assessed here. ↩

11. Codex constructed the proof-and-analogy illustration from my clarification that an AI is a second speaker for an explanatory idea it develops. The illustration is hypothetical; no particular proof or analogy is being assessed. ↩

12. The AI assistant proposed making the effectiveness question explicit by asking what aim is being served. The positions it develops here—that I judge the equally moving poems equally and that wanting a particular human’s poetry is legitimate—are mine. ↩

13. Microsoft’s official TRELLIS repository documents generation of 3D assets from image or text inputs and local editing. A generated starting model is not a guarantee of a finished production asset or a physically realised sculpture. Access, technical judgment and further work still matter. ↩

14. The ILO’s task-exposure analysis distinguishes potential exposure from realised employment effects and expects job transformation to predominate. The US Bureau of Labor Statistics projects editor employment to decline 1% in 2025–35, and effects-artist and animator employment to show little or no change, while noting AI may dampen demand in the latter group. That category does not isolate 3D asset designers. These scoped projections do not yet show the disappearance I predict, and they do not establish safety for particular workers. The prediction is mine, made on the capability I see rather than on these figures. The AI assistant proposed the distinction between loss of paid work and loss of something to express; the concern and the expressive opportunity were supplied by me. ↩