The Originality Question: Who Is the Author When Humans and AI Write Together?
Maurício Pinheiro
AI authorship became an unavoidable question in January 2023, only weeks after ChatGPT became publicly available. At the time, a deceptively simple question occurred to me: Could we determine whether a text had been written by a human or generated by an artificial intelligence?
That question immediately opened several others. Could AI-generated text constitute plagiarism? Who should receive authorship credit? Could an invisible watermark reveal that a machine had participated in producing a text? More broadly, could we reliably distinguish human writing from machine-generated language simply by examining the final result?
What happened next was almost symbolic.
On January 31, 2023, OpenAI released a classifier designed to distinguish AI-written from human-written text. Less than six months later, the company withdrew it because of its low accuracy.
That short-lived experiment came to represent something larger than the failure of a particular detection system.
It suggested that perhaps the problem itself had been framed incorrectly.
In 2023, the central question seemed to be:
Can we determine whether a machine wrote this text?
Three years later, that question feels increasingly insufficient.
Generative AI is no longer merely producing texts alongside human writers. It is becoming embedded within the act of writing itself: suggesting ideas, reorganizing arguments, searching for alternatives, rewriting passages, translating, criticizing, editing, and sometimes generating substantial portions of a work.
The boundary we once hoped to detect has therefore become much harder to define.
And this leads to a more difficult question:
When human thought and machine generation become intertwined in the same creative process, what does it even mean to say who wrote the text?
The Detector Was Asking the Wrong Question
The early debate assumed a relatively clean division.
A text was either written by a human or generated by an artificial intelligence.
If that distinction existed, then perhaps some sufficiently sophisticated algorithm could inspect the finished document and determine which side of the boundary it belonged to.
But the boundary itself has become increasingly artificial.
Consider a researcher who spends months developing an argument, asks an AI system to improve the English, rewrites several passages, verifies every reference and publishes the final version.
Was the article written by AI?
Now imagine someone who writes three sentences describing a subject, asks an AI system to generate two thousand words, makes a few superficial edits and publishes the result under their own name.
Was that article written by a human?
Between those extremes lies almost every possible combination.
A person may formulate the ideas while AI generates sentences. AI may suggest ideas while the person rejects most of them. A human may write the first draft and use AI as an editor. AI may produce the first draft and a human may transform it beyond recognition. One model may help search the literature, another may organize an argument, and another may polish the language.
There is no longer a single binary variable called AI-generated that adequately describes all of these processes.
And this is why text detectors face a deeper problem than imperfect accuracy.
Detectors inspect the artifact. Authorship belongs to the chain of decisions that produced it.
A detector can map the statistical topography of language, but it cannot reconstruct the chain of intellectual choices that built the argument.
It cannot know why one interpretation was selected and another discarded. It cannot know whether a paragraph originated in months of research or in a five-second prompt. It cannot know whether the person publishing a statement understands it, agrees with it or merely accepted what a model proposed.
Trying to infer authorship solely from the finished text therefore confuses linguistic origin with intellectual origin.
The consequences are not merely philosophical. Research has already shown that AI-text detectors can systematically misclassify some human writing, including writing by non-native English speakers.
The failure is therefore not simply technical.
It is conceptual.
We tried to infer authorship from linguistic fingerprints at precisely the moment when authorship was ceasing to leave a unique fingerprint.
AI Generation Is Not the Same as Plagiarism
Another distinction that was much less clear in the early debate now needs to be explicit.
AI use, plagiarism, copyright infringement and lack of originality are not the same thing.
A human can plagiarize without using AI.
An AI-assisted text can contain no plagiarism at all.
A work can be legally permissible yet intellectually derivative.
A work can contain genuinely original analysis even though AI participated extensively in its editing and composition.
The categories must therefore be separated.
Plagiarism is a theft of attribution; copyright infringement is a violation of legal protection; lack of originality is a failure of intellectual novelty; AI use is, by itself, merely a choice of instrument.
The distinctions are not perfect, and the concepts sometimes overlap, but collapsing them produces more confusion than clarity.
This has become particularly important because the copyright debate has moved far beyond the relatively simple question of whether a generated sentence resembles an existing sentence.
The use of copyrighted works during model training is one problem. The copyrightability of AI-assisted outputs is another. Whether a particular passage reproduces protected expression is another. Whether an author has properly acknowledged intellectual sources is yet another.
These questions interact, but they cannot be reduced to a single test for “AI-generated text.”
Nor can originality be measured simply by textual similarity.
A completely new sentence can express an entirely conventional idea.
A familiar sentence structure can convey a profoundly original one.
The originality question was never really about vocabulary.
It was about intellectual contribution.
The Return of the Human Author
Recent legal thinking has increasingly returned to one particularly important concept: human creative contribution.
This offers a more useful way of approaching authorship.
The important variable may not be the percentage of words generated by a machine.
It may instead be the location of creative control.
Who chose the problem?
Who established the central claim?
Who determined which evidence mattered?
Who rejected incorrect or irrelevant suggestions?
Who shaped the structure of the argument?
Who made the interpretive decisions?
And who is prepared to take responsibility for the finished work?
These questions tell us considerably more about authorship than asking whether an algorithm generated sentence number 37.
The distinction becomes especially important when people argue that prompting itself constitutes authorship.
There are certainly prompts that require knowledge, imagination and skill. A sophisticated prompt can constrain style, define objectives, establish a conceptual direction and guide a system toward a particular result.
But instruction is not necessarily authorship.
A film director does not become the cinematographer merely by telling the cinematographer what image is needed. A person commissioning a portrait does not become the painter because they specify the pose, clothing and background. Giving detailed instructions can constitute creative direction without constituting execution.
Prompting occupies a similar territory.
At one extreme, a prompt may contain nearly the entire intellectual substance of the resulting work, with the model doing little more than reformulating it.
At the other, a prompt may contain little more than:
Write an article about consciousness.
It would be difficult to argue that those two acts represent equivalent creative contributions.
Prompting can therefore be part of authorship.
It is not, by itself, proof of authorship.
The decisive question remains what intellectual work the person actually contributed.
The act of typing is no longer proof of thinking.
The Strange Case of the AI Editor
The problem becomes clearer when we compare generative AI with technologies and practices we already accept.
Nobody normally argues that a spellchecker becomes co-author because it changes several words.
We do not transfer authorship to Microsoft Word because it corrects grammar.
Professional writers also work with human editors who suggest cuts, reorganize sentences, challenge weak reasoning and sometimes rewrite substantial passages.
Yet authorship generally remains with the person responsible for the intellectual work.
Generative AI initially appears to belong somewhere on this continuum.
But the analogy eventually breaks down.
A human editor is another intentional agent.
An editor can disagree with the author, defend an objection, refuse an instruction, explain why an argument fails or accept professional responsibility for a decision. The relationship contains a reciprocal structure of accountability.
An AI system does not participate in that relationship in the same way.
It produces probabilistic continuations.
It can simulate disagreement, criticism or editorial judgment, sometimes extraordinarily well, but it does not possess an independent professional obligation to defend the integrity of the work.
If it proposes a bad argument, the responsibility for accepting that argument remains elsewhere.
This makes human supervision more important, not less.
There is also a meaningful difference between asking:
Improve the clarity of this paragraph without changing my argument.
and asking:
Develop the argument, find supporting reasons and write the article.
The first delegates primarily expression.
The second may delegate substantial parts of reasoning.
The distinction begins as one of degree, but at scale, it becomes a difference of kind.
Between those extremes lies the territory in which modern authorship has become genuinely difficult to describe.
Watermarking Was Not a Bad Idea
One of the earliest proposed solutions to the problem of identifying AI-generated text was watermarking.
The idea was straightforward: if machine-generated language could contain a hidden statistical signature, then perhaps its origin could later be detected without relying solely on stylistic clues or imperfect AI classifiers.
That idea did not disappear.
It became considerably more sophisticated.
Modern watermarking techniques can alter the statistical process through which tokens are selected so that machine-generated material contains patterns that are difficult for readers to notice but potentially detectable computationally.
Such approaches may become useful components of a broader provenance infrastructure.
But watermarking does not solve the authorship problem.
A watermark may provide evidence that a particular generation system participated in producing some content.
It cannot tell us who conceived the argument.
It cannot reveal how extensively the material was subsequently rewritten.
It cannot establish whether the person publishing the text understood it.
It cannot determine who deserves intellectual credit.
And it certainly cannot tell us who should accept responsibility for an error.
Watermarking can help establish provenance.
It cannot establish originality.
That distinction may prove fundamental.
From Detection to Provenance
The debate is consequently moving from forensic detection toward provenance.
With the recent applicability of the European Union’s AI Act transparency obligations, this movement has acquired a regulatory dimension as well. The emerging logic is increasingly concerned not simply with guessing whether artificial intelligence touched a piece of content, but with establishing how synthetic material can be identified, documented and responsibly presented.
This is a more promising direction.
Instead of examining a finished article and asking software to guess how it was produced, provenance attempts to preserve information about the process itself.
Technical initiatives such as the Coalition for Content Provenance and Authenticity (C2PA) illustrate this approach by providing mechanisms through which information about the origin and transformation of digital content can accompany the content itself.
For writing and research, provenance may also require something less technological and more procedural: disclosure.
Academic publishing already offers a useful conceptual precedent in contributor frameworks such as the CRediT taxonomy, which separates different forms of contribution instead of treating authorship as a single indivisible act.
Something similar could eventually emerge for AI-assisted intellectual work.
Was AI used for language editing?
Literature discovery?
Translation?
Data analysis?
Argument generation?
Drafting?
Critical review?
Those distinctions are considerably more informative than a label saying simply AI-generated.
A serious publishing culture may increasingly document sources, drafts, human review, editorial responsibility and the role played by computational systems.
The conceptual change is profound.
Detection asks:
Can I discover whether AI was involved?
Provenance asks:
Can I understand how this work came into existence?
The second question is far more important.
Originality After Generative AI
Generative AI also forces us to reconsider what originality itself means.
We often confuse originality with wording.
But the deepest forms of originality rarely reside in individual sentences.
They reside in choosing a question nobody considered important, connecting ideas that previously appeared unrelated, recognizing a pattern hidden in existing evidence, constructing a new hypothesis, rejecting an accepted explanation or seeing an old problem from a different perspective.
A machine may help express such an idea.
Increasingly, it may also help discover the connection.
That makes the attribution problem harder, but not meaningless.
Imagine a scientist who spends twenty years developing a theory and then asks an AI system to rewrite the final paper in clearer English.
Calling the resulting work “AI-generated” would tell us almost nothing about its intellectual origin.
Now reverse the situation.
Someone asks an AI system to propose an unexplored hypothesis, construct the argument, identify supporting literature, generate examples and write the article. The person reads the result, changes several sentences and publishes it.
Calling that simply “human-written” would be equally misleading.
The relevant dimension is no longer human versus machine.
It is the distribution of intellectual agency between them.
And this distribution may vary within a single document.
The hypothesis may be human.
The literature search may be machine-assisted.
The argument may emerge through interaction.
The prose may be substantially generated.
The verification may return to the human.
What, then, is the percentage of AI authorship?
Perhaps the question itself has become meaningless.
Authorship Is a Chain of Decisions
Perhaps authorship has always been less about writing words than we assumed.
An author chooses.
The author decides what problem deserves attention, what information matters, which objections are serious, which analogies illuminate rather than distort, what should remain and what should disappear.
An author converts possibilities into commitments.
Generative AI can now participate in almost every stage of this sequence.
It can propose hypotheses, organize arguments, identify weaknesses, generate counterexamples and produce alternative formulations.
But generation is not selection.
Possibility is not commitment.
And suggestion is not responsibility.
This may be where the distinction between human and machine remains most important.
The model can produce ten arguments.
Someone must decide which one is worth defending.
It can generate a confident paragraph containing an error.
Someone must decide whether to publish it.
It can suggest a conclusion.
Someone must be willing to put their name beneath it.
Authorship therefore need not disappear merely because machines participate in writing.
But we may need to understand it less as an act of textual production and more as a structure of judgment, selection and responsibility.
That change will not always be comfortable.
It may reveal that some texts bearing human names contain remarkably little human intellectual contribution.
It may also reveal the opposite: that some heavily AI-assisted texts remain fundamentally expressions of human thought.
A detector cannot resolve that distinction.
Only an account of the creative process can.
So Who Wrote This?
This essay itself provides a useful demonstration of the problem.
Artificial intelligence participated in its production.
It was used to test formulations, challenge sections of the argument, suggest structural improvements, identify repetition and help refine the language.
But every one of those interventions passed through a sequence of judgment gates.
Should this argument remain?
Is this distinction valid?
Does this source support the claim?
Is this paragraph saying what I intend to say?
Has the machine introduced an idea I reject?
Has it expressed one of my ideas better than I did?
Should I accept the formulation, modify it or discard it completely?
The final document emerges from that interaction.
So who wrote it?
Counting keystrokes would not answer the question.
Counting AI-generated tokens would not answer it either.
The more relevant questions are whether I determined what I wanted to say, whether I evaluated what the system proposed, whether I rejected what I considered wrong, whether I verified the evidence, whether the final argument represents my judgment and whether I accept responsibility for publishing it.
That is not an attempt to make AI participation disappear.
Quite the opposite.
It is an attempt to describe that participation more accurately.
We are witnessing the emergence of a new form of intellectual production, while much of our vocabulary was developed for a world in which minds produced ideas and machines merely recorded them.
That world is disappearing.
The Originality Question, Revisited
In 2023, the fear was that AI-generated text would become indistinguishable from human writing.
That prediction was simultaneously correct and beside the point.
The deeper transformation is that human and machine writing are becoming difficult to separate because they are increasingly elements of the same creative process.
The future of authorship therefore cannot depend on proving that no machine participated in creating a work.
Such a standard would become both unrealistic and intellectually uninteresting.
Nor should we replace it with the opposite extreme and pretend that using AI automatically makes every resulting work equally legitimate, original or human-authored.
What matters is deeper.
Where did the central ideas come from?
Where was judgment exercised?
Who converted information into argument?
Who decided what was worth keeping?
Who recognized what was wrong?
And who takes responsibility for what remains?
Those questions cannot be answered by an AI detector.
They cannot be settled by a watermark.
And they cannot be reduced to the sophistication of a prompt.
Perhaps this is the central lesson of the last three years.
The originality question was never really about detecting machines.
It was about locating intellectual agency.
Generative AI has not destroyed originality.
It has exposed how poorly we had defined it.
The author of the future may not be the person who wrote every sentence.
It may be the person who can explain why every sentence is there—and is willing to answer for it.
Selected References
OpenAI. New AI Classifier for Indicating AI-Written Text (2023; classifier withdrawn July 2023 because of low accuracy).
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. & Zou, J. GPT Detectors Are Biased Against Non-Native English Writers. Patterns 4, 100779 (2023).
U.S. Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability (2025).
U.S. Copyright Office. Copyright and Artificial Intelligence, Part 3: Generative AI Training (2025).
Dathathri, S. et al. Scalable Watermarking for Identifying Large Language Model Outputs. Nature (2024).
European Commission. Guidelines on Transparency Obligations under Article 50 of the AI Act (2026).
Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.
Copyright 2026 AI-Talks.org
