Who Is Writing When We Write with AI?

Every act of writing involves assistance.

Language precedes the writer. So do the books that taught that language what a sentence might contain. Dictionaries settle small disputes. Editors expose weaknesses the writer has learned not to see. Research supplies facts the imagination could not safely invent. A conversation overheard years ago returns in altered form. A photograph suggests a room. A piece of music changes the rhythm of a paragraph.

None of these things is equivalent to generative artificial intelligence. A dictionary does not propose a scene. A photograph does not answer when questioned. An editor may offer a sentence, but does so from a human life whose judgement can itself be questioned.

The distinction matters.

What generative AI changes is not merely the amount of help available. It changes the intimacy of the intervention. A writer can now work with a tool capable of producing language that did not previously exist: a scene, a continuation, an alternative gesture, another ending to the same conversation. It can alter the temperature of a paragraph, move an image from one character’s perspective to another, or produce several plausible versions of a moment in the time it might once have taken to write one.

That capacity should be described without either embarrassment or exaggeration.

I used generative AI materially in writing Mirrors. It proposed language that entered drafts. It also helped test, vary, compress and reconsider language already there. Some of its proposals survived in altered form. Many disappeared. The process through which the novel was written included generation, rejection, redirection, rewriting and repeated comparison with the book it was gradually becoming.

The translations now being developed have followed a related process: generation, close comparison with the English text, correction, and a return to the particular meaning of a passage whenever fluency threatened to become more persuasive than fidelity.

To call all of this merely assistance would make the role of the technology smaller than it was.

To say that the technology simply wrote the novel would make everything else disappear.

The truth lies in the more difficult territory between those statements.

A sentence is not yet a decision

A generative system can produce an extraordinary number of sentences.

This is both its power and its limitation.

A novel does not require merely the existence of possible sentences. It requires decisions about which possibilities belong to the work and which must be refused, however persuasive they appear.

A character can answer a question tenderly, defensively, cruelly or not at all. A scene can disclose its meaning immediately or carry it unresolved for two hundred pages. An object can become evidence, memory, accusation or nothing more than something a person has been unable to throw away.

Generation multiplies these possibilities. It does not relieve the writer of choosing among them.

In some respects, it makes the choice more severe.

When producing language was slow, scarcity performed part of the selection. A paragraph that had taken an hour to write acquired a certain authority simply because it existed. Effort could disguise itself as value. The writer had already invested something and therefore became reluctant to lose it.

When ten plausible alternatives can appear rapidly, existence becomes much less persuasive. Fluency is plentiful. Completion is easy to imitate. The fact that a paragraph has arrived is no reason for it to remain.

The question changes from Can I write this? to Does this belong?

There were passages in the making of Mirrors that were coherent, elegant and wrong. They understood a character too completely. They brought tenderness into a relationship before tenderness had been earned. They allowed someone to say what that person had spent years learning not to say.

Nothing was visibly defective in the sentences.

The defect was that they belonged to another novel.

The work begins before the prose

Mirrors did not begin with a request for a novel.

Before its scenes acquired their final language, the book needed a form: seven Mirrors, forty-two interlocking sections, recurring characters encountered at different ages, relationships that had to remain coherent as their meanings changed, and questions that could deepen without becoming answers.

At its centre was a person who could exert pressure on the entire novel without being granted a definitive voice.

That decision excluded possibilities before a sentence was written.

The absent person could not appear directly. No final letter, recovered account or privileged memory could settle who they had really been. Contradictory recollections had to remain capable of being simultaneously sincere. Characters could revise their understanding without the novel declaring their earlier selves mistaken. Objects could return with altered meanings, but could not become clues arranged around a solution.

These constraints were not imposed afterwards to give the book an appearance of coherence. They determined what the prose was permitted to do.

The same was true inside individual scenes.

Who was standing? Who was seated? Which character knew something at that moment, and which could not yet know it? Had a relationship already acquired a formality it did not possess twenty years earlier? Could this person plausibly use this word, or had another scene established that they would avoid it? Was a silence carrying several meanings, or was it merely vague? Did an image deepen the scene, or explain what should remain unresolved?

Together, these decisions became the private law of the book.

Generated language had to live within that law.

When it did not, the language was wrong no matter how accomplished it sounded.

Fluency can conceal what is wrong

This distinction became increasingly important because generative systems are very good at producing coherence.

Coherence can be dangerous.

A sentence may sound inevitable without belonging to the character who speaks it. An explanation may be psychologically convincing while damaging an ambiguity the novel needs much later. A beautiful image may repeat another image too closely. A reconciliation may feel emotionally satisfying while violating everything two people have spent the book becoming.

Fluency conceals these errors because fluency resembles authority.

I learned to distrust the passage that arrived too finished.

Sometimes the work was subtractive. A scene appeared with too much understanding and had to lose some of it. A character articulated an emotion too precisely and needed to become less eloquent. A connection became visible before the people involved could have recognised it. An ending achieved such perfect balance that it closed a question the novel had no right to close.

The improvement was not always another sentence.

Sometimes it was the removal of the sentence that explained why the others mattered.

At other times, the passage had to be refused completely. This could be more difficult than rejecting something obviously poor. Bad language creates little attachment. A fine sentence asks to be protected. It encourages the writer to adjust the scene around it, then the character, then perhaps the book, until the sentence has acquired a place it never deserved.

The finished page records what survived. It does not preserve the branches that were abandoned, the explanations removed, the scenes redirected or the apparently good sentences discarded because they carried the work towards a different emotional truth.

Yet those absences belong to the authorship of the novel.

A book is made from what the writer permits to remain.

It is also made from what the writer is willing to lose.

Selection is not enough

It would be convenient to define authorship simply as choosing among alternatives.

That definition is too generous.

A person can select from generated possibilities while surrendering almost every meaningful decision to the system. If the instruction amounts to give me something good, followed by acceptance whenever the result appears competent, selection becomes little more than approval.

Authorship requires an independent conception of the work.

The writer must be able to recognise not only that a passage is wrong, but why it is wrong—and what preserving it would cost elsewhere.

In Mirrors, the important question was rarely whether a scene worked in isolation. It was whether the scene altered the pressure of other scenes without reducing them. A line in the thirty-fourth section could change the meaning of something said in the fourth. A gesture had to remain consistent with a relationship shown at another age. A remembered phrase could recur, but each recurrence needed to belong to the person remembering it rather than to the architecture that placed it there.

This meant carrying the whole while attending to the part.

No individual passage could assume that responsibility. It could not know that a tender sentence had arrived too early unless the entire history of the relationship remained present. It could not decide that a minor contradiction was human while another would break the chronology. It could not feel that a silence had become repetition rather than accumulation.

A system could produce the language of the scene.

The responsibility for recognising the scene belonged elsewhere.

The tool changed the room

It would also be dishonest to claim that the technology affected only speed.

A new tool changes the kinds of thought available around the work.

Generation made comparison unusually powerful. A scene could be approached at different emotional temperatures. An exchange could be rewritten with more or less explanation. A moment could be viewed from another character’s centre of gravity. Possibilities that might once have remained hypothetical could be made concrete enough to judge.

Most of them did not survive.

But seeing them changed the act of choosing.

The process sometimes felt like working in a room filled with possible versions of the same object. One could turn them, place them beside one another, notice where the weight had shifted, and discover that the most immediately attractive version was also the least capable of carrying time.

Occasionally, a discarded version contained one small thing that revealed what the scene had been missing: not a complete solution, but a gesture, a hesitation, the recognition that one character should be looking at the wrong person.

At other times, the alternatives merely confirmed that the first instinct had been right.

Both outcomes mattered.

The tool did not stand outside the writing process. It altered the room in which decisions were made. It increased the number of visible possibilities and therefore changed the conditions under which intention had to be defended.

I see little value in pretending otherwise.

The relevant question is not whether the tool changed the writing. It did.

The question is whether the writer remained capable of recognising what the change was doing to the work.

Some questions cannot be delegated

There were questions no generative system could answer for Mirrors because they were not questions of probability, elegance or narrative efficiency.

How much certainty should the reader finally be denied?

When does ambiguity become evasion?

How long can a father remain understandable before understanding begins to resemble forgiveness?

Should an apology repair a relationship, or merely reveal that repair is no longer what either person needs?

At what point does caring for someone become a way of controlling them?

Which version of a person deserves to survive when every available version is partial?

These questions do not have optimal answers.

They require commitments.

A novel acquires its identity by making such commitments repeatedly and accepting their consequences. One choice closes possibilities another would have preserved. A character is permitted to remain silent, and the book must live without the knowledge that speech would have provided. A relationship is denied reconciliation, and the ending must find another form of peace. An absent person is refused the right to explain themselves, and everyone else must continue carrying incomplete versions of them.

The book gradually becomes unable to be another book.

Generative technology can propose language within that narrowing field. It can illuminate consequences the writer has not yet considered. It can expose a choice by offering the road not taken.

It cannot decide why the field should narrow in one direction rather than another.

The answer does not exist inside the generated sentence.

Responsibility remains with the writer

For me, this is the most important question surrounding authorship and artificial intelligence.

Not who pressed each key, but who accepts responsibility for the work.

A writer cannot attribute a weak character to the tool, or a false emotional note, or a contradiction that should have been caught, while claiming authorship of everything that succeeds.

The same process that permits selection requires accountability.

If a generated sentence remains in the novel, it remains because it was accepted. If a structure fails, the failure belongs to the person who preserved the structure. If a character becomes simpler than a human being should be, it is not enough to say that a system produced the language.

Someone decided that the language was finished.

This responsibility is not made smaller by the fact that writing has always involved influence, conversation, editing and tools. Generative AI is different from those forms of assistance. It can participate directly in the production of prose, and honesty requires that participation to be named.

But naming it does not settle the question of authorship.

Authorship, as I understand it, is not a claim that every word originated in an unaided act of composition. It is the acceptance of responsibility for the form that finally remains: for its choices, omissions, patterns, failures and consequences.

That responsibility cannot be delegated to the tool without delegating the work itself.

Transparency is not the work

There is a temptation around new creative technologies to make the method more important than the thing made.

I would rather not do that.

Mirrors should stand or fail as a novel.

A reader does not owe the book admiration because its process was elaborate. Nor should the method excuse a sentence that fails, a scene that does not move or a character who does not feel alive.

The process matters because honesty about process matters.

It does not replace the encounter with the work.

This is why I prefer to describe the method plainly rather than turn it into either a secret or a manifesto. Generative AI was materially involved in producing and revising the language of Mirrors. Human direction established and preserved the architecture, constraints, characters, continuity and final form through which that language became this particular novel.

The distinction is not perfectly clean. Creative processes rarely are. Some decisions emerged through exchange rather than arriving fully formed in advance. Some possibilities became visible only because the tool proposed them. Direction itself was occasionally changed by what the process revealed.

That does not make responsibility impossible to locate.

Influence can be shared.

Responsibility cannot.

The choices remain

Every technology used in writing eventually becomes less visible.

The pen disappears behind handwriting. The typewriter disappears behind the page. Word processing, spell-checking, digital research and electronic editing are usually absent from our experience of a finished novel, even though each changed the conditions under which novels could be made.

Generative systems may never disappear in quite the same way. Their ability to produce language raises questions the earlier tools did not raise, and those questions deserve serious answers.

But the deeper problem of authorship remains strangely familiar.

A writer encounters possibilities.

Some are kept. Others are refused. A form gradually emerges from those decisions. What began as an open field becomes a path, then a structure, and finally a work capable of resisting further change.

Eventually there comes a point at which the writer can no longer say that the book is merely becoming something.

It has become something.

At that point, no individual phrase stands outside the pattern of judgement, revision and responsibility that has brought the work into its final form.

The tool can generate another sentence.

It can always generate another sentence.

The writer’s task is to know when the novel does not need one.