# Who Owns a Story Nobody Wrote? Authorship, Copyright and the Problem of Machine Creativity ## 1. The Author Who Could Not Prove She Wrote Her Book The novel had taken six months to produce. Its author, Elena Ward, had developed the original idea, designed the characters and constructed a complicated plot involving a missing mathematician and a manuscript that appeared to predict the future. She had written some chapters herself. For others, she had used an artificial intelligence system to generate preliminary drafts. She had rewritten dialogue, replaced entire scenes, altered character motivations and rejected hundreds of suggested passages. The final manuscript contained approximately eighty thousand words. When her publisher asked whether she owned the copyright, Elena replied confidently. "Of course. It's my novel." The publisher's legal adviser was less certain. "Did you write every sentence?" "No." "Did you choose every sentence?" "Not individually." "Did the machine generate passages that remain substantially unchanged?" "Some." "Did you record which passages?" Elena hesitated. She had kept the final manuscript, but not a complete history of the drafting process. The adviser continued. "Then we need to distinguish what you created from what the machine generated." Elena was irritated. She had designed the story, directed the process and spent months revising the work. Without her, the novel would not exist. Surely that made her its author. Yet the adviser was asking a different question. Not whether Elena had contributed creatively. Not whether she deserved recognition. But which parts of the finished text were protected by copyright, who owned those rights and what evidence supported the claim. This scene is fictional. Elena, her novel and the conversation with the publisher have been invented for this chapter. But the problem is genuine. Artificial intelligence allows people to create literary works through processes that do not fit neatly into familiar descriptions of writing. A person may compose a prompt, generate thousands of words, select promising passages, revise them and arrange the results into a finished narrative. Another person may use AI only to check punctuation. A third may create a program that produces hundreds of poems without selecting any particular output. All three have used computers. But their creative contributions differ substantially. Copyright law must distinguish these situations without assuming that every use of software has the same consequences. The question becomes even harder when the generative system has been trained using material written by other authors. Those writers may never have agreed to participate. Their works may have influenced the system's capabilities. And the generated text may sometimes resemble protected material. Who, then, can claim the resulting story? To investigate the problem, we must separate several questions that are often confused. Who created the work? Who owns the legal rights? Who contributed creatively? And who is responsible for its publication? They may not all have the same answer. Review dated 31 October 2026. This comparative chapter describes selected statutes and official reports; it does not decide the legal status of a particular manuscript. The US reports are administrative analyses. Direct retrieval of the live UK statute and EUR-Lex texts was unavailable during this review; the source ledger distinguishes the official corroboration used from provisions not fully retrieved. ![Authorship, ownership, permission and responsibility are separate questions about the same manuscript.](images/contribution-questions.svg) ![A manuscript page from Mary Shelley’s Frankenstein, 1816. A historical example of writing and revision.](images/frankenstein-draft.jpg) A manuscript page from Mary Shelley’s Frankenstein, 1816. A historical example of writing and revision. Credit: Mary Shelley (1797-1851). [Source](https://commons.wikimedia.org/wiki/File:FrankensteinDraft.jpg). [Public domain](https://creativecommons.org/publicdomain/mark/1.0/). Resized; no crop. ## 2. The Difference Between an Author and an Owner We often use the words author and owner as though they mean the same thing. A novelist writes a book. The novelist owns its copyright. The novelist sells copies. Everything seems straightforward. But authorship and copyright ownership are distinct legal concepts. The author is generally the person who creates the relevant expression, subject to the definitions and rules of the jurisdiction. The copyright owner is the person or organisation entitled to exercise the relevant economic rights. Initially, these may be the same person. They do not necessarily remain so. A writer may assign copyright to another party. An employer may own certain works created by employees in the course of their employment. A publisher may receive an exclusive licence to publish a work without becoming its copyright owner. The details depend on the law and contracts involved. ### What does copyright protect? Copyright generally protects particular original expressions, rather than ideas in the abstract. Consider the premise: A detective discovers that a machine has been predicting murders. This is an idea for a story. Many writers could develop it differently. One might produce a psychological thriller. Another might write a philosophical satire. A third might construct an interactive mystery. The general premise is not equivalent to the detailed expression of a finished novel. Specific prose, dialogue, characterisation and creative arrangements may qualify for protection when the applicable legal requirements are met. This is why two novels can share a broad premise without necessarily infringing each other's copyright. But copying substantial protected expression may create legal problems even when the copied material appears inside a different story. The boundary is not always obvious. ### Ownership is not the same as permission Suppose Elena uses a writing assistant to help draft her novel. She must consider at least three separate matters. First, whether copyright exists in the resulting material. Second, who owns any copyright that exists. Third, whether she has the contractual and legal permission required to use and publish it. A service's terms may grant users broad rights concerning generated outputs. That does not itself establish that every output receives copyright protection under national law. Nor does a contractual permission guarantee that the output cannot infringe someone else's rights. Contracts, copyright eligibility and infringement are different questions. This distinction is essential when analysing AI-generated literature. ## 3. The American Position: Human Authorship In January 2025, the United States Copyright Office published Part Two of its report Copyright and Artificial Intelligence, addressing the copyrightability of generative-AI outputs. Its conclusion was that existing US copyright principles can accommodate AI-assisted creativity, but protection requires sufficient human authorship. The Office distinguished between using AI as a tool and allowing a system to determine expressive elements without sufficient human creative control. It concluded that human-created expression, creative modification and sufficiently original selection or arrangement may be protected, while providing prompts alone does not ordinarily establish authorship of the resulting AI-generated expression. This does not mean every work involving AI is unprotected. It means the human contribution matters. ### Three different manuscripts Imagine three versions of Elena's novel. Version A: Elena writes the entire manuscript herself and uses software to check spelling and punctuation. Version B: Elena writes the plot, selects passages generated by AI, rewrites many scenes and creatively arranges the finished material. Version C: Elena enters a short prompt, receives an entire novel and publishes it unchanged. These examples illustrate different degrees of human control. Under the Copyright Office's analysis, Version A presents no unusual obstacle merely because software was used. Version B may contain protectable human-authored contributions, but the scope of protection depends upon the actual creative work performed. Version C presents substantial difficulty if the expressive content was determined by the generative system rather than a human author. The important question is not how many hours Elena spent operating the software. It is what original expression she contributed. ### A documented example: Zarya of the Dawn In February 2023, the US Copyright Office reconsidered the registration of Zarya of the Dawn, a comic book created by Kris Kashtanova using human-written text and AI-generated images. The Office recognised protectable human contributions, including the text and the creative selection, coordination and arrangement of text and images. It excluded the AI-generated artwork itself from the revised registration. The decision illustrates how a larger work can contain both protected human-authored material and material that is not independently protected through the same claim. The example concerns visual material, not an AI-written novel. Nevertheless, the underlying distinction is relevant to computational literature. A work may possess copyright protection in its human-created structure or expression without every component receiving identical protection. ### Creativity cannot be counted by percentage Suppose Elena rewrites 40% of an AI-generated chapter. Does she own 40% of the copyright? There is no general legal formula that operates this way. A small but highly original creative contribution may be significant. A large quantity of mechanical changes may contribute little protectable authorship. The question concerns the nature and originality of human expression, not merely the percentage of words changed. This is why records of drafting and revision can be useful. They help demonstrate what the human author actually contributed. ## 4. The British Exception: Computer-Generated Works The United Kingdom has an unusual legal provision addressing computer-generated works. The government’s March 2026 report discusses this existing protection and unresolved questions about its application; policy discussion should not be mistaken for enacted reform. Section 9(3) of the Copyright, Designs and Patents Act 1988 states that, for qualifying computer-generated literary, dramatic, musical or artistic works, authorship is attributed to the person who undertakes the arrangements necessary for creation. Section 178 defines the relevant category as work generated by computer in circumstances where there is no human author. Section 12(7) provides a special copyright term of fifty years from the end of the calendar year in which such a work was made. This is a different statutory approach from the US position. But it would be a mistake to conclude that every AI-generated passage automatically belongs to the person who typed the prompt. ### Who made the necessary arrangements? Imagine that a publisher commissions a programmer to build a poetry generator. The programmer designs its selection procedures. A researcher supplies training material. A publisher operates the program. A reader enters a prompt. The system produces a poem. Who undertook the arrangements necessary for its creation? The question may not have an obvious answer. The statute provides a principle, but identifying the relevant person in modern generative-AI workflows can be difficult. There are also questions about how originality requirements apply and whether particular AI-assisted works fall within the statutory definition. These questions should not be treated as conclusively resolved merely because the legislation contains a computer-generated-works provision. ### A historical mismatch The 1988 legislation predates modern large language models. Lawmakers were addressing a technological environment very different from today's generative systems. Early computer-generated works might have involved explicitly programmed procedures and relatively clear relationships between programmer, operator and output. Modern AI services can involve model developers, training-data providers, application developers, prompt writers and editors. The creative process is distributed across several participants. Applying an older statutory category to these arrangements can be difficult. This illustrates a recurring problem in technology law. A legal rule may remain in force even when the technology it originally addressed has changed substantially. Its wording matters. But so do interpretation, legal precedent and the facts of the particular case. ## 5. Does a Machine Have to Be an Author? Our legal investigation reveals an important philosophical distinction. A machine might contribute causally to the production of a work without being recognised as its legal author. But does the absence of legal authorship establish that the machine is incapable of creativity? No. Legal categories do not automatically resolve philosophical questions. Copyright law serves particular public and economic purposes. It defines rights, obligations, ownership and periods of protection. Philosophy asks different questions about creativity, intention and agency. A computer might produce an output judged original and artistically valuable. Whether that makes the computer creative depends upon the definition of creativity being used. Whether the computer possesses conscious creative intentions is another question. Whether the computer can legally own copyright is another. ### Four claims that must remain separate Consider the following statements. 1. The machine generated the words. 2. The words possess literary value. 3. The machine understood what it generated. 4. The machine owns copyright in the words. The first concerns technical causation. The second concerns artistic evaluation. The third concerns cognition and understanding. The fourth concerns law. One statement does not automatically establish the others. A poem can be meaningful to readers even if the system producing it lacks demonstrated subjective experience. A technically sophisticated system can create an output that nobody considers artistically worthwhile. A human may contribute important creative decisions without writing every word. And legal ownership may depend on statutes and contracts rather than philosophical theories of consciousness. The title of this chapter therefore contains a deliberate provocation. Who owns a story nobody wrote? In practice, many supposedly machine-written stories involve substantial human creative contributions. The harder question is how to identify those contributions accurately. ## 6. The Library Inside the Machine We now arrive at the most contentious part of the problem. A generative model does not begin with an independent understanding of literary language. Its capabilities are developed through training. Training material may include books, webpages, articles, computer code and other forms of expression, depending on the particular model. Some material may be licensed. Some may be in the public domain. Some may have been made available under permissions that allow particular uses. Other material may be subject to copyright and disputed uses. This creates two different questions. Was the material lawfully used to train the model? Does a particular output infringe copyright? They are related, but they are not the same question. ### Learning from a book is not identical to republishing it A human novelist may read hundreds of detective stories and learn about the genre. They may acquire a sense of pacing, dialogue and narrative structure. They may then write an original mystery. Reading and learning do not automatically constitute copyright infringement. But machine learning introduces technical operations that complicate the comparison. Training may involve making copies of source material, processing those copies and incorporating statistical relationships into model parameters. The legal significance depends upon the jurisdiction, applicable exceptions, permissions and particular circumstances. A broad analogy to human reading cannot settle the question. Nor can the assertion that a model learns statistical patterns establish that every use of copyrighted training material is lawful. ### The US debate In May 2025, the US Copyright Office released a pre-publication version of Part Three of its AI report, examining generative-AI training. It analysed the application of copyright law, including fair use, to different kinds of training activity. The report recognised that uses of copyrighted works for AI training are not all equivalent and that the legal analysis depends upon their circumstances. Fair use is a doctrine of US copyright law permitting certain unauthorised uses after consideration of statutory factors. It is not a blanket exemption for new technology. Nor is it a general international rule. The legal treatment of AI training remains subject to judicial decisions and evolving interpretation. ### The European approach The European Union's 2019 Copyright in the Digital Single Market Directive contains provisions addressing text and data mining. Article 4 establishes an exception or limitation for certain reproductions and extractions of lawfully accessible works, subject to conditions including appropriate reservations of rights by copyright holders. In addition, Article 53 of the EU AI Act establishes obligations for providers of general-purpose AI models, including a policy for compliance with EU copyright law and a public summary of training content. These provisions address different aspects of the problem. The copyright directive concerns the legal framework for certain uses of protected material. The AI Act includes obligations concerning compliance policy and transparency. Neither should be described as giving unrestricted permission to train on every copyrighted work. The detailed legal position depends upon the applicable rules and facts. ### Why writers care An author may believe that their books were used to develop a commercial system without appropriate permission. A model developer may argue that its training practices fall within an applicable exception or licence. A publisher may worry that generated text could reproduce protected expression. Readers may want to know whether a work involved AI assistance. These are different interests, and they cannot all be resolved by asking who pressed the Generate button. The dispute concerns the economic and cultural relationship between existing literature and systems capable of producing new language. ## 7. Can a Writing Style Be Owned? Suppose Elena asks a model to write a detective story using short sentences, restrained dialogue and a melancholy atmosphere. Many writers use such techniques. The resulting text might resemble a broad literary tradition without copying any particular work. Now imagine a different request. Elena asks the model to imitate a specific contemporary novelist. The output reproduces familiar patterns of imagery, rhythm and narration. Is that copyright infringement? The answer is not automatically yes. But it is not automatically no. ### Style and expression Copyright generally does not grant exclusive ownership over an abstract style. A novelist cannot ordinarily monopolise the use of short sentences, unreliable narration or atmospheric descriptions of rain. Other writers may employ similar techniques. But particular protected expression can be infringed. If generated text reproduces sufficiently substantial protected material, the fact that it was produced by AI does not automatically remove the legal concern. The analysis may involve similarities, access, originality, applicable exceptions and the substantiality of what was taken. Jurisdiction matters. ### The difference between imitation and copying Consider a deliberately broad instruction: Write a Gothic scene involving an isolated house and an unreliable narrator. This describes genre conventions. Now consider a request for a passage closely reproducing the distinctive wording of a living novelist. The second presents a different kind of risk. Even when direct copyright infringement is not established, deliberate imitation can raise ethical or commercial concerns. Readers may be misled into believing that a writer endorsed or participated in the work. A publisher might exploit another author's reputation. A generated passage might also reproduce elements that go beyond general style. The law's boundaries and the ethics of literary imitation do not coincide perfectly. Something may be legally permissible while remaining artistically derivative or ethically questionable. ### Literary influence has always existed Writers have always learned from earlier writers. A novelist might be influenced by Dickens, Austen, Woolf or Kafka. Literary history is full of transformations, parodies, allusions and adaptations. Generative AI does not invent literary influence. It changes its scale, speed and mechanisms. A writer can request many imitations almost instantly. A publisher can generate large quantities of material resembling established genres. An individual author's style can become the subject of automated reproduction. These capabilities make questions about attribution and consent more pressing, even when copyright law does not provide exclusive rights over style itself. ## 8. The Story That Remembers Too Much There is another problem. A language model may produce text that resembles material encountered during training. Such resemblance can arise for different reasons. Some phrases are common. Some literary conventions are widely shared. Some passages may be memorised sufficiently closely that generated outputs reproduce substantial portions. The distinctions matter. A familiar sentence does not automatically indicate unlawful copying. But a lengthy passage closely reproducing a protected work may require investigation. ### Memorisation and reproduction Researchers have demonstrated that some language models can reproduce portions of training material under particular conditions. This does not mean that all models store their training data as a directly searchable library of complete books. Nor does it mean that every output is an exact collage of existing passages. Modern models learn complex numerical relationships, but memorisation can still occur. The practical concern is whether particular outputs contain protected expression and whether their use is authorised. ### The publisher's problem Imagine Elena has generated a passage describing a coastal town. She likes the result and includes it in her manuscript. Months later, a reader recognises several sentences from an existing novel. Elena did not knowingly copy the original. Nevertheless, the similarity may create a problem. Her lack of deliberate copying is relevant to the circumstances but does not, by itself, guarantee that publication is lawful. The publisher may need to examine the material and obtain professional advice. A careful workflow should therefore treat generated text as material requiring editorial review, rather than assuming that machine generation guarantees originality. That principle applies to factual accuracy as well. The computer may produce a convincing historical quotation that never existed. It may invent a source or misattribute a statement. Literary fluency does not establish provenance. ## 9. The Difference Between Credit and Responsibility Suppose Elena's finished novel contains a factual allegation about a real person. The allegation is false and damaging. The paragraph was generated by AI. Who is responsible? Elena cannot necessarily escape responsibility by explaining that the machine wrote it. The publisher cannot assume that software-generated material is exempt from ordinary editorial standards. Other legal and ethical considerations may apply, including defamation, privacy, contractual duties and consumer protection, depending on the facts and jurisdiction. Copyright ownership is only one part of the publishing problem. ### Creative credit A writer might disclose that AI was used in brainstorming, drafting or editing. Such disclosure can help readers and publishers understand the creative process. But a credit statement is not a substitute for determining copyright eligibility. Nor does it establish that the model itself possesses legal authorship. ### Editorial responsibility Someone must decide whether a manuscript is ready for publication. That includes checking whether its factual claims are supported, its fictional continuity is coherent and its use of existing material is appropriate. These responsibilities remain important regardless of whether AI contributed. Consider a manuscript containing an invented historical document. Within a novel, the document may be a legitimate fictional device. But if the publisher markets it as a genuine discovered manuscript, the claim changes. The issue is no longer simply literary invention. It concerns representation to readers. ### A useful publishing record For complex AI-assisted projects, an author may find it useful to maintain a record containing: - The original creative brief and outline. - Human-written drafts and revisions. - Generated passages incorporated into the manuscript. - Material substantially rewritten or rearranged. - Sources used to verify factual claims. - Relevant permissions and licensing information. This record does not create copyright by itself. It can, however, provide evidence of the creative process and help resolve later questions about provenance. The distinction between having rights and being able to demonstrate the basis of a claim is particularly important when many participants or tools contribute to one work. ## 10. The Literary Work as a Chain of Contributions We can now describe Elena's novel more accurately. It was not produced by one isolated event. It emerged through several stages. Elena developed an idea. She constructed a plot. She wrote prompts. A language model produced passages. She selected and revised some of them. She integrated the results into a manuscript. An editor proposed further changes. The publisher prepared the book for readers. Each stage contributed to the finished artefact. But not every contribution has the same legal or creative status. ### Causation is not authorship The software developer helped make generation possible. The model provider supplied a service. The prompt writer directed particular operations. The editor influenced the final manuscript. But causal participation alone does not establish legal authorship of all resulting expression. A manufacturer who builds a camera does not ordinarily become the author of every photograph taken with it. A word-processing company does not own every novel written using its software. Generative AI complicates these familiar analogies because it may determine expressive details that were not individually supplied by the user. Nevertheless, the lesson remains: we must identify the nature of each contribution. ### The importance of human selection Imagine that Elena generates one hundred possible openings. She chooses one. She rewrites it. She changes the point of view and integrates it with the novel's later revelations. The final passage reflects her creative decisions. But suppose she instead accepts the first generated paragraph without modification. The human contribution to that passage is different. The legal consequences depend on applicable law and facts, not a universal formula. The artistic consequences may also differ. A writer's selection can be meaningful, but not every selection is equally original. Creative control is a matter of substance, not simply the existence of a human somewhere in the process. ### Authorship becomes more visible Generative AI may therefore encourage a more explicit understanding of what authors do. They do not merely produce words. They establish intentions. They choose structures. They create relationships. They revise for meaning. They decide which possibilities belong in the work. But those activities should be demonstrated rather than assumed. A compelling account of authorship identifies the decisions that actually shaped the finished expression. ## 11. A New Literary Economy The questions surrounding AI-generated writing are not confined to individual manuscripts. They concern the wider literary economy. Books are written, edited, published, distributed and read within institutions. Authors depend on different combinations of sales, advances, royalties, commissions and other income. Publishers invest in production and marketing. Libraries preserve works and make them accessible. Readers develop expectations about originality, authenticity and attribution. Generative AI can affect each of these relationships. ### The abundance problem A system capable of producing large quantities of text can dramatically reduce the cost of generating preliminary material. But producing text is not the same as producing books readers value. Editing, research, design, distribution and reputation remain important. In a market containing abundant generated writing, readers may place greater value on trusted editorial selection and demonstrable human creative vision. Alternatively, some forms of routine writing may become increasingly automated. The effects will vary by genre, market and institution. They cannot be inferred solely from the technical ability to generate fluent paragraphs. ### Consent and compensation Writers may seek greater control over the use of their works in training. Model developers may seek licensing arrangements that provide access to suitable material. Publishers may negotiate permissions concerning their catalogues. Different economic arrangements are possible. But these are matters of policy, bargaining and law as well as technology. A computer cannot decide what constitutes fair compensation simply by calculating the most efficient arrangement. The values involved are contested. This brings us back to the distinction explored in Chapter Four. Technical capability does not automatically confer legitimate authority. The same principle applies to the literary marketplace. A system's ability to process books does not itself determine which uses are lawful, fair or socially desirable. ## 12. The Name on the Cover Elena finally completed her novel. She reviewed the chapters, reconstructed the development of important scenes and documented how the manuscript had been created. Some passages were entirely her own. Others had originated as generated suggestions but were extensively transformed. A few remained largely as the model produced them. She decided that the most accurate description of the work would acknowledge those distinctions. But one question still troubled her. Was she less of a writer because she had used a machine? There is no reason to assume that using a computational tool automatically removes artistic achievement. Writers have always employed tools, research materials, collaborators and editorial assistance. What matters is how the tools contribute and what creative decisions the writer makes. But neither should we assume that every machine-produced passage becomes human-authored merely because somebody requested it. The distinction requires attention to the actual process. ### The central lesson We began this chapter by asking who owns a story nobody wrote. The question turns out to contain several different problems. A text may have been generated by a machine. Its literary value may depend upon human interpretation. Its copyright status may depend upon the creative contributions of particular people and the law of a particular country. Its use may also depend upon contracts and permissions. And responsibility for publishing it may remain with humans and organisations even when substantial portions were generated automatically. These questions are connected, but they cannot be resolved through a single definition of authorship. The United States has emphasised sufficient human creative expression. The United Kingdom retains a special statutory provision for certain computer-generated works. The European Union has introduced important requirements concerning general-purpose AI providers and copyright compliance, while copyright protection remains governed by relevant EU and national rules. The legal landscape is evolving. The philosophical questions are older and deeper. A machine may produce language that readers value. That does not automatically establish conscious creative intention. A person may direct and shape computational output. That does not automatically establish exclusive rights in every resulting sentence. And a publisher may have permission to distribute material while still carrying responsibility for its accuracy and legal compliance. Authorship, ownership, creativity and responsibility are different relationships with a work. The arrival of generative AI makes those distinctions impossible to ignore. The name on a book's cover remains important. But increasingly, understanding how the book was made may be just as important as knowing whose name appears there. ## Questions to Take Away 1. What is the difference between being the author of a story and owning its copyright? 2. Why does copyright generally protect particular expression rather than a broad fictional premise? 3. How does the US Copyright Office distinguish AI-assisted human creativity from expression determined by a generative system? 4. Why does the UK's computer-generated-works provision create different questions from the American human-authorship approach? 5. Can a machine produce an artistically valuable text without possessing conscious creative intentions? 6. What is the difference between the legality of AI training and the copyright status of a particular generated output? 7. Why is imitating a literary style not necessarily equivalent to reproducing protected expression? 8. If an AI-generated passage contains a false allegation or an unattributed quotation, who should investigate and correct it before publication? 9. Why might a record of drafts, revisions and human creative decisions be useful in an AI-assisted publishing workflow? 10. Could an abundance of automatically generated writing make editorial judgement more valuable rather than less? ## A Small Reader Experiment: Who Created the Paragraph? This experiment examines creative contribution rather than attempting to provide a legal ruling. Consider three versions of a fictional scene. ### Step 1: The generated draft Imagine that an AI system produces this original teaching passage: The detective entered the old house. It was dark and quiet. She found a letter on the table and realised that it contained an important secret. The passage is grammatical, but its literary details are limited. ### Step 2: A human revision A writer changes it to: The house smelled of damp plaster and extinguished candles. On the kitchen table lay an envelope addressed in the detective's own handwriting. She had never seen it before. The second passage introduces specific sensory detail and a mystery involving the protagonist's identity. The writer has changed the language, imagery and dramatic mechanism. ### Step 3: A different contribution Another writer keeps the original paragraph unchanged but places it at the end of a chapter in which the detective has repeatedly denied that she can remember the house. The words have not changed. But their position within the larger story gives them new significance. :::lab contribution ### Step 4: Compare the contributions | Version | Human contribution | Question to consider | | ------------------- | ----------------------------------- | ----------------------------------------------------------- | | Generated draft | Task direction | Who determined the particular wording? | | Rewritten paragraph | New expressive language and imagery | What original expression did the writer add? | | Rearranged passage | Selection and narrative placement | Does the arrangement involve sufficiently creative choices? | Under US copyright principles, creative human revision and original arrangement may be eligible for protection, while merely supplying a prompt does not ordinarily establish authorship of the generated expression. But the specific legal outcome cannot be decided from this simplified table. ### Step 5: Make your own revision Rewrite the scene so that the letter is not important because of its contents. Instead, make its physical existence the mystery. Perhaps it is addressed to someone who has not yet been born. Perhaps the envelope is older than the house. Perhaps the handwriting belongs to a person who cannot write. Choose one possibility and write a paragraph. Then identify the decisions that made it different from the starting text. Consider which elements involve your own creative expression. ### The deeper question Now imagine publishing the revised paragraph. Would you describe the process as writing, editing, collaboration or computational assistance? There may be more than one reasonable description. But your explanation should identify what actually happened. What this experiment demonstrates: Creative contribution cannot be measured simply by counting generated and human-written words. It involves choices about language, structure, arrangement and meaning, while legal protection must be assessed under the applicable law. ## Further Reading The following are primary legal sources and official institutional publications. They describe legal rules and administrative interpretations, rather than universally applicable conclusions about every AI-assisted work. ### United States US Copyright Office — Copyright and Artificial Intelligence, Part 2: Copyrightability (2025) The Office's analysis of human authorship and copyright protection in works involving generative AI. [Official report — PDF](https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf) US Copyright Office — Zarya of the Dawn registration decision (2023) A documented decision distinguishing human-written text and creative arrangement from AI-generated imagery. [Official decision letter — PDF](https://www.copyright.gov/docs/zarya-of-the-dawn.pdf) US Copyright Office — Copyright and Artificial Intelligence, Part 3: Generative AI Training (2025) Analysis of copyright questions surrounding the use of protected material in AI training. The Office released a pre-publication version in May 2025. [Official AI copyright study and reports](https://www.copyright.gov/ai/) ### United Kingdom Copyright, Designs and Patents Act 1988 The principal UK statutory source relevant to computer-generated works. [Section 9 — Authorship](https://www.legislation.gov.uk/ukpga/1988/48/section/9) [Section 12 — Duration of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/12) [Section 178 — Definition of computer-generated works](https://www.legislation.gov.uk/ukpga/1988/48/section/178) ### European Union Directive (EU) 2019/790 — Copyright in the Digital Single Market Includes Article 4 concerning text and data mining and the reservation of rights. [Official directive — EUR-Lex](https://eur-lex.europa.eu/eli/dir/2019/790/oj) Regulation (EU) 2024/1689 — Artificial Intelligence Act Article 53 addresses obligations for general-purpose AI model providers, including copyright-compliance policies and summaries of training content. [Official AI Act — EUR-Lex](https://eur-lex.europa.eu/eli/reg/2024/1689/oj) [European Commission — General-purpose AI obligations](https://digital-strategy.ec.europa.eu/en/factpages/general-purpose-ai-obligations-under-ai-act) ### Suggested Reading Order Begin with the US Copyright Office's 2025 copyrightability report. It provides a clear framework for understanding the distinction between AI-generated expression and human creative contribution. Then compare the UK statutory provisions, observing how their treatment of computer-generated works differs. Finally, examine the EU text-and-data-mining rules and AI Act obligations, keeping the legality of training separate from the ownership of generated outputs. The most useful approach is comparative: identify which question each source addresses before attempting to draw conclusions. ## Up Next — The Infinite Library Information, Randomness and the Search for Meaning Suppose every possible book already exists somewhere. Every novel that has been written. Every novel that will ever be written. Every alternative ending. Every false history. Every accurate prediction. Every masterpiece. And every meaningless arrangement of symbols. In Jorge Luis Borges's The Library of Babel, the universe takes the form of an immense library containing books assembled from combinations of characters. Some volumes contain recognisable language. Others contain apparent nonsense. Many appear to contain fragments of meaningful information buried in oceans of meaningless text. The problem is no longer how to produce a book. It is how to discover which books matter. Chapter Fifteen will return to Borges and connect his imagined library with combinatorics, probability, Claude Shannon's information theory and the limits of automated generation. We will investigate why the number of possible texts becomes astronomically large, why information is not identical to meaning and why finding one useful sentence in an enormous collection can be harder than generating the entire collection. The central question moves beyond authorship. If every possible story can be generated, what makes one story worth finding, reading and remembering? [UK Government — March 2026 report on copyright and AI](https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence/report-on-copyright-and-artificial-intelligence) [Carlini and collaborators — Original 2021 training-data extraction research](https://www.usenix.org/system/files/sec21-carlini-extracting.pdf)