When Machines Become Creative

A writer enters a sentence and receives a page. A designer enters a description and receives an image. A musician hums a melody and receives an arrangement. A filmmaker can now generate storyboards, voices, backgrounds, effects, and moving scenes without assembling the traditional chain of specialists who once produced each part. The result can be beautiful. It can also be commercially useful, derivative, forgettable, startling, or legally uncertain. The machine can generate the file in seconds. That does not answer who authored the work, who owns it, whose material helped produce it, or who should be paid.

Chapter Fourteen argued that the objective is the hidden decision. Creative AI makes that principle visible. A generative system is not simply a neutral fountain of possibilities. It has been trained on some material and not other material, optimized toward some outputs and away from others, constrained by provider rules, and placed inside a market with incentives about what should be easy to make.

A machine can make the file. It cannot settle who owns the work.

Generation Is Not Authorship

The first question sounds philosophical but is already legal: what counts as authorship when a machine can produce expressive material?

The U.S. Copyright Office has taken a relatively clear position. Copyright remains available when AI assists human creativity. A human-authored work does not lose protection merely because AI was used somewhere in the process. Human selection, arrangement, modification, and other creative contributions can also be protected. But purely AI-generated material is not protected by copyright, and prompts alone generally do not give the user enough control over expressive elements to establish authorship.

That distinction is more practical than it first appears. A photographer can use computational tools without surrendering authorship of the photograph. A writer can use software to reorganize notes, propose alternatives, or edit prose while remaining the author of the human expression. A designer can incorporate generated material into a larger work and own the protectable human arrangement or modifications. The analysis turns on what the person actually contributed, not whether the word AI appears somewhere in the workflow.

The difficult case is the one-button masterpiece. If a user asks for a striking image, receives it, and publishes it without meaningful creative control over the expressive result, the image may have commercial value without receiving the same copyright protection as a human-authored work.

The model does not become the author. The user does not become the author merely by being the person who pressed Generate.

The Work Is More Than the Prompt

A prompt is easy to see because it is typed into a box. Creative control is often distributed across a much longer process. A person may choose the source material, generate dozens of alternatives, reject most of them, combine fragments from different outputs, rewrite passages, change composition, alter pacing, replace details, and decide what finally reaches the public. The visible prompt can therefore be the smallest part of the human contribution.

The first draft does not tell the whole story of creative work. A person can make important choices before the AI generates material, while trying alternatives, and while revising the result. An editor may reshape a manuscript, a producer may combine generated music with recorded performances, and a filmmaker may arrange generated and filmed scenes. To understand the human contribution, examine those choices across the process.

The opposite is also true. Repetition alone does not necessarily create authorship. A user can generate hundreds of images until one looks appealing without determining the expressive details that made that image appear. The Copyright Office’s case-by-case approach therefore asks a useful practical question: where did the human actually determine expression rather than merely request or select a result?

For creators and publishers, the practical habit is simple: keep a record of the creative work when ownership matters. Save the human drafts, source files, edits, arrangements, selections, and decisions that show how the final work was made. Provenance means where the material came from and how the work was made. It can help show both the origin of training material and what the human creator contributed.

The Creative Chain Is a Rights Chain

Creative work has always involved chains of contribution. A film may involve a screenwriter, director, actors, cinematographer, editor, composer, studio, and distributor. A song may involve writers, performers, producers, publishers, labels, and platforms. Contracts decide who owns which rights because creation and ownership are rarely the same question.

AI adds more contributors to that chain. Someone created the source material. Someone assembled or licensed the training data and built the model. A user supplied instructions or material, and someone selected and revised the output. A business may have commissioned the work, and a platform may distribute it. Several of these roles may belong to the same person or company.

That makes the creative chain a rights chain. Different actors may own different things, owe different permissions, and bear different risks.

The accountability mistake is to look only at the final file. An AI-generated image can lack copyright protection of its own and still raise an infringement claim if it reproduces someone else’s protected material. A contract can give a user permission to use an output without giving the user copyright in every part of it.

A provider may also offer indemnification: a contractual promise to cover specified losses or legal costs. That promise has limits. It may exclude particular uses, changes made by the user, material the user supplies, or models made by another company. The contract determines what is covered.

Ownership therefore has to be traced backward through the process rather than inferred from possession of the final file.

Training Is One Question. Output Is Another.

The most confused arguments about creative AI often mix two separate issues. The first is whether copyrighted works may be copied to train a model. The second is whether a particular output infringes a particular work. Those questions can produce different answers.

The Copyright Office’s 2025 report on generative AI training rejected a universal rule. It concluded that fair use remains fact-specific. Some training uses may be transformative and fair, particularly where access is lawful and the use does not substitute for the original market. Other uses may weigh against fair use when copyrighted works are obtained unlawfully, when models produce material that competes in the same market, or when workable licensing markets exist and are bypassed.

The Anthropic book litigation demonstrates how sharply the distinction can matter. In 2025 Judge William Alsup ruled that using lawfully acquired books to train large language models could be fair use, while finding no fair-use justification for pirated copies retained in a central library. The litigation later produced a $1.5 billion class settlement over the pirated-book claims. Judge Araceli Martínez-Olguín granted final approval in July 2026. The settlement covered 482,460 works, required destruction of the pirated copies covered by the agreement, and expressly excluded claims based on AI outputs.

That history should make both camps uncomfortable. It rejects the simple claim that every act of AI training is infringement. It also rejects the idea that transformative training creates a permission slip for acquiring the underlying material however a company chooses. Judge Alsup’s formulation was unusually direct: Anthropic had no entitlement to use pirated copies for a central library.

Training is one rights question. Acquisition is another. Output is another. Collapsing them into one slogan makes the debate easier and the law less intelligible.

Market Harm Is Becoming the Battleground

Copyright was never designed only to reward copying that looks identical. Fair use also asks what a new use does to the market for the original and for reasonable derivative markets.

That question becomes difficult when a system does not ordinarily reproduce a book, painting, or song verbatim but can generate enormous quantities of competing material. The harm claimed by creators may come less from one copied output than from a market flooded with cheaper substitutes trained on creative works.

The Copyright Office’s training report treats that concern seriously. It notes that licensing markets for AI training are developing and that an available licensing market can matter to fair-use analysis. It also warns that market dilution from stylistically similar or substitutive outputs can be relevant even when no single output is substantially similar enough to infringe.

Courts are still drawing the boundary. The consolidated New York litigation involving OpenAI, Microsoft, The New York Times, and authors including John Grisham and George R.R. Martin reached a major fair-use stage in September 2026, with both sides seeking summary judgment. The plaintiffs argue that unlicensed training creates competitive substitutes and harms the market for journalism and books. OpenAI and Microsoft argue that the training is transformative and learns statistical relationships rather than redistributing the works themselves. The Times has also placed output behavior inside the training case by pointing to examples in which its articles, including paywalled material, were reproduced in substantial part. OpenAI attributes those examples to adversarial prompting, and Microsoft disputes that the record shows outputs substituting for the original works. The court has not yet resolved the core fair-use question.

The federal government has now entered that rights chain. On September 1, 2026, the Justice Department filed a statement of interest supporting OpenAI’s fair-use position and arguing that generative-AI training can be highly transformative and that overly restrictive rules could harm scientific progress, U.S. competitiveness, and national security. Reuters described the filing as the first time the federal government had taken a position in this wave of AI-training litigation. The intervention does not decide the case. It does make the state another actor in the creative chain, arguing that a private use should be permitted partly for public reasons.

This Living Edition uses OpenAI systems in its production workflow. That relationship does not determine the merits of the litigation, but it does require the disclosure. Evidence should not become friendlier because the vendor is useful.

Style Is Not the Same as Identity

Creative people often describe the harm of generative AI in the language of style: a song sounds like a particular performer, an image feels like a particular illustrator, or prose appears to mimic a particular writer.

Copyright law does not generally grant ownership of an artistic style as such. It protects original expression, not the broad idea of painting, writing, or composing in a recognizable manner. That limitation matters because overly broad control over style could allow one creator to own techniques that later creators need in order to make new work.

Identity is a different issue. A generated song using a recognizable replica of a singer’s voice can implicate publicity, digital-replica, consumer-protection, contract, or other rights even if the melody and lyrics are new. The Copyright Office concluded in 2024 that existing protections were too fragmented and recommended a federal law covering unauthorized digital replicas of individuals’ voices and likenesses.

The distinction is useful. Influence is unavoidable in culture. Deception and appropriation of identity are not the same thing as influence.

A system should not get a free pass to impersonate a person merely because copyright does not own style.

From Lawsuit to License

The music industry is beginning to show what governance looks like when litigation becomes contracting.

Major record companies sued AI music companies Suno and Udio in 2024, accusing them of using copyrighted recordings without authorization. The industry did not move in one direction after that. Warner Music settled with Suno in November 2025, and BMG reached its own arrangement in August 2026. Believe and TuneCore joined the licensing side in September. Universal Music Group and Sony Music, however, remained plaintiffs against Suno. The shift from lawsuit to license is therefore real but incomplete.

On September 9, 2026, Suno released new music models in partnership with Warner Music Group and BMG. Reuters reported that the arrangements contemplate products in which participating artists can opt in to uses of their names, images, voices, likenesses, and compositions and share in revenue. Suno also said its newer models block prompts seeking the likeness of a specific known artist. That is Principle 3 implemented at the interface: the permission boundary is not only contractual language behind the product; it changes what the product will do.

Licensing is not a universal solution. A creator may refuse. Transaction costs can be high. Different rights may be controlled by different parties. Small creators can have less bargaining power than major labels. But licensing does something accountability systems need: it identifies the owner, defines the boundary, and attaches economic consequences.

Permission becomes architecture when the product is built so the creator can say yes, no, and on what terms.

The Objective Is the Hidden Editor

Chapter Fourteen argued that personalization systems reveal their values through what they optimize. Creative systems do the same.

A model tuned for photorealism will make different creative choices from one tuned for safety. A music service optimized for instant familiarity may reward resemblance. A writing system optimized for agreeable completion may flatten disagreement or surprise. A video generator optimized for commercial usability may avoid outputs that are difficult to license.

The objective is the hidden editor. The goals chosen for the system help determine which kinds of work it produces easily, which requests it refuses, and how often it returns to familiar patterns. Those choices also reflect what risks the provider is willing to accept. Creative judgment begins before the user types a prompt.

Adobe offers one useful counter-model because it made training provenance part of the product proposition. Adobe says its Firefly models are trained on licensed content such as Adobe Stock and on public-domain material, not on Creative Cloud customer content, and that it compensates Stock contributors for training use. It also offers intellectual-property indemnification for certain eligible enterprise Firefly features.

Those choices do not establish that Firefly is risk-free or legally superior in every situation. They show that a provider can make provenance, permission, compensation, and liability part of product architecture instead of treating them as arguments to have after deployment.

The hidden editor can be governed when its objectives become visible.

This Book Has to Answer Its Own Question

A book about delegated intelligence should not pretend that its own production process sits outside the argument.

This Living Edition uses AI systems to research, compare, challenge, draft, format, and prepare update and distribution assets under defined rules. The author retains final editorial responsibility for the argument, selection of evidence, revisions, approvals, and publication. AI-generated material is reviewed rather than treated as self-authenticating.

That description does not settle copyrightability by itself. Copyright law still asks what human-authored expression and creative control exist in the final work. It does, however, answer the accountability question. There is a named human decision owner. The machine assists the process; it does not acquire editorial authority merely because it can produce sentences.

The same discipline should apply elsewhere. A publisher, studio, agency, or musician using generative tools should be able to describe what the system did, what the human creator did, what sources or licenses govern the tool, who reviewed the result, and who stands behind publication.

Creative transparency should describe the division of labor, not perform a ritual confession that software was present.

Current Evidence: Ownership Is Moving Into Product Design

Reviewed September 11, 2026.

The U.S. Copyright Office’s current position on copyrightability remains centered on human creativity. AI assistance does not disqualify a work, but purely machine-generated expression is not protected, and prompts alone generally do not establish sufficient human control. That makes process evidence increasingly important for creators who use generative tools.

The training question remains less settled. The Copyright Office’s Part 3 report says fair use depends on the circumstances and gives substantial weight to lawful access, transformation, market substitution, and developing licensing markets. Its final edition is still pending, but the Office says it does not expect substantive changes from the pre-publication analysis.

In September 2026, the Third Circuit rejected Ross Intelligence’s fair-use defense for using Westlaw headnotes to train a competing legal-research system. The court emphasized the similar commercial purpose and expressly distinguished Ross’s nongenerative system from generative-AI disputes. The decision does not settle every AI-training case.

The training material passed through outside contractors. For this appeal, Ross did not contest that their copying was attributable to it. The practical lesson is to follow the material through the supplier chain: who obtained it, what permission supported its use, and who authorized it for training? Hiring another company to prepare the data does not make those questions disappear.

The Anthropic settlement put a dollar figure on acquisition discipline. In July 2026 Judge Araceli Martínez-Olguín approved the $1.5 billion settlement covering 482,460 works and resolving class claims involving pirated books. The agreement required destruction of the pirated copies and excluded claims based on AI outputs. The earlier ruling by Judge William Alsup had treated model training on lawfully acquired books as fair use while finding no entitlement to maintain the pirated central library. The settlement does not convert that district-court reasoning into nationwide law, but it makes one practical boundary hard to miss: how training material is obtained matters.

The OpenAI and New York Times litigation is now approaching a major merits decision. In September both sides asked a federal judge to resolve core fair-use questions on summary judgment. The Times has placed alleged regurgitation at the center of the record by identifying outputs that reproduce substantial portions of articles, including paywalled material; OpenAI attributes those examples to adversarial prompting, while Microsoft disputes that the outputs substitute for the original works. On September 1 the Justice Department filed a statement of interest supporting OpenAI’s fair-use defense and arguing that restrictive training rules could harm scientific progress, competitiveness, and national security. The court has not resolved the issue, and appeals are likely whatever the ruling.

Music is moving through contracts faster than courts can produce final doctrine. Suno’s September 2026 launch with Warner Music and BMG creates licensed and opt-in pathways for participating artists, while Believe and TuneCore have also moved onto the licensing side. Universal and Sony remain plaintiffs against Suno, so the market is splitting rather than settling. Suno’s block on prompts seeking a named artist’s likeness also shows permission being implemented at the interface. The economic question becomes not only whether a model may learn from music, but whether artists can choose participation and share in the value created from their identities and works.

Visual media remains adversarial. Disney and Universal sued Midjourney in 2025, alleging that its image generator copied and reproduced copyrighted characters without authorization. Midjourney has disputed copyright claims in the broader wave of AI litigation. The case remains another test of the output side of the problem, where courts can examine generated images directly rather than infer harm only from training.

The evidence is producing a more complicated market than either side originally imagined. Some training is being litigated. Some is being licensed. Some outputs receive human-authored copyright protection and some do not. Providers are differentiating themselves through provenance and indemnity. Ownership is moving upstream into dataset design and downstream into contracts, editing, disclosure, and distribution.

Accountability Analysis

Principles 1, 3, 5, and 6 carry most of this chapter. Principle 1 begins with authorship and publication: the model is not the accountable creator. A person or institution still chooses whether to use the output, claim rights in it, sell it, publish it, or place another person’s identity inside it.

Principle 3 asks for the rights boundary. What material may be used for training? What licenses govern it? What may a user request? What forms of imitation or digital replication are prohibited? Which outputs may be used commercially? The boundary belongs in data governance, model behavior, contracts, and product controls rather than in a disclaimer nobody reads.

Principle 5 reaches the hidden editor. Training-data choices, optimization targets, refusal policies, licensing rules, and interface design shape creative output before the user sees it. A provider that controls those choices participates in the creative architecture even when the user supplies the final prompt.

Principle 6 follows the money and the risk. If creators supply the cultural material while platforms capture most of the economic value, the allocation becomes part of the accountability question. Licensing, compensation, attribution, indemnification, and opt-in mechanisms are ways to reconnect benefit and consequence.

The creative chain is therefore not a contest to identify one owner of everything. It is a map of who contributed what, who had permission, who exercised creative control, who receives value, and who answers if the resulting work crosses a legal or ethical boundary.

Describe the creative contribution

Choose one work made with AI assistance. Record what material you supplied, what the system generated, and what you selected, arranged, rewrote, or otherwise changed. Identify who reviewed the result and decided to publish it. Keep the drafts that show those steps. This record explains the division of labor; it does not by itself settle ownership or permission.

Accountability Verdict

Generative AI should be used aggressively as a creative tool. It can lower the cost of experimentation, give individuals capabilities once reserved for studios, make revision faster, and allow people to create across media they never formally learned.

The price of that expansion should not be pretending that authorship, training rights, identity, and output ownership are one question. They are separate links in the creative chain.

The practical rule is to follow the work backward. Who supplied the source material? Who had permission to use it? What did the model contribute? What did the human choose, arrange, revise, or transform? Who controls publication? Who receives the value? Who bears the claim if the work infringes or impersonates?

Creativity can be delegated. Authorship still has to be earned.

Chapter Sixteen moves from created works to created relationships. When an artificial system can remember, flatter, comfort, advise, and simulate intimacy, the question is no longer who owns the output. It is what happens when the product is designed to become someone’s confidant.

Sources and Notes

1. U.S. Copyright Office, “Copyright and Artificial Intelligence, Part 2: Copyrightability,” January 29, 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf. Accessed September 11, 2026. The Office concludes that AI assistance does not bar copyright protection for human-authored expression; purely AI-generated material is not protected; prompts alone generally do not provide sufficient control over expressive elements; and human selection, coordination, arrangement, and creative modification may be protected. The report states current Copyright Office analysis and is not a judicial decision. Source record: DA-CH15-S1. Archive reference: pending.

2. U.S. Copyright Office, “Copyright and Artificial Intelligence, Part 3: Generative AI Training,” pre-publication version, May 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf. Accessed September 11, 2026. The Office concludes that fair use analysis for generative-AI training is fact-specific and discusses lawful access, transformation, market substitution, licensing markets, and market dilution. The Copyright Office states that a final version will be issued without substantive changes expected in the analysis or conclusions. Source record: DA-CH15-S2. Archive reference: pending.

3. Reuters, “US judge approves Anthropic’s $1.5 billion settlement of copyright lawsuit,” July 20, 2026; and related reporting and court records in Bartz v. Anthropic. https://www.reuters.com/world/us-judge-approves-anthropics-15-billion-settlement-copyright-lawsuit-2026-07-20/. Accessed September 11, 2026. Judge Araceli Martínez-Olguín granted final approval to a $1.5 billion settlement covering 482,460 works and resolving class claims involving pirated books retained in Anthropic’s training library. The settlement required destruction of the pirated copies covered by the agreement and expressly excluded claims concerning AI outputs. The underlying June 2025 ruling by Judge William Alsup treated training on lawfully acquired books as fair use while finding that Anthropic had no entitlement to use pirated copies for a central library. The settlement does not create nationwide precedent on AI training. Anthropic systems are also used in this Living Edition’s external critique workflow; see the transparency statement. Source record: DA-CH15-S3. Archive reference: pending.

4. Reuters, “OpenAI, New York Times case tees up key test of AI training under copyright law,” September 8, 2026, and related reporting on the Justice Department’s September 1 statement of interest. https://www.reuters.com/legal/litigation/openai-new-york-times-case-tees-up-key-test-ai-training-under-copyright-law-2026-09-08/. Accessed September 11, 2026. The consolidated litigation involves The New York Times, other publishers, authors, OpenAI, Microsoft, and related parties. Cross-motions for summary judgment were filed September 4, 2026. The plaintiffs argue that unlicensed training, alleged regurgitation of article passages including paywalled material, and competing outputs infringe copyright and harm markets; OpenAI argues that training is transformative fair use and attributes regurgitation examples to adversarial prompting; Microsoft disputes that the output record shows substitution. On September 1 the Justice Department filed a statement of interest supporting OpenAI’s fair-use position and arguing that restrictive rules could harm scientific progress, U.S. competitiveness, and national security. Reuters described it as the first federal-government position in this wave of AI-training litigation. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH15-S4. Archive reference: pending.

5. Reuters, “Suno releases new AI music models in partnership with Warner Music, BMG,” September 9, 2026; and Reuters, “Warner Music Group settles copyright case with Suno for licensed AI music,” November 25, 2025. https://www.reuters.com/legal/litigation/suno-releases-new-ai-music-models-partnership-with-warner-music-bmg-2026-09-09/ and https://www.reuters.com/legal/litigation/warner-music-group-settles-copyright-case-with-suno-licensed-ai-music-2025-11-25/. Accessed September 11, 2026. Reuters reports that Warner settled with Suno in November 2025, BMG reached an agreement in August 2026, and Believe and TuneCore joined the licensing side in September, while Universal Music Group and Sony Music remained plaintiffs against Suno. Suno’s newer licensing relationships contemplate opt-in participation and revenue opportunities for artists and rightsholders, and the company says its newer models block prompts seeking the likeness of a specific known artist. The chapter uses these developments as evidence that permissions and compensation are becoming product architecture, not as proof that licensing resolves every copyright dispute. Source record: DA-CH15-S5. Archive reference: pending.

6. Reuters, “Disney, Universal sue image creator Midjourney for copyright infringement,” June 11, 2025. https://www.reuters.com/business/media-telecom/disney-universal-sue-image-creator-midjourney-copyright-infringement-2025-06-11/. Accessed September 11, 2026. Disney and Universal alleged that Midjourney used and generated unauthorized copies of copyrighted characters and sought damages and injunctive relief. The allegations remain contested and are used as an example of output-focused infringement litigation rather than as established liability. Source record: DA-CH15-S6. Archive reference: pending.

7. Adobe, “Approach to generative AI with Adobe Firefly,” current edition; Adobe Firefly FAQ; and Adobe Firefly Product Description, published August 27, 2026. https://www.adobe.com/ai/overview/firefly/gen-ai-approach.html, https://helpx.adobe.com/firefly/web/get-started/learn-the-basics/adobe-firefly-faq.html, and https://helpx.adobe.com/legal/product-descriptions/adobe-firefly.html. Accessed September 11, 2026. Adobe states that Firefly models are trained on licensed content such as Adobe Stock and public-domain material, not on Creative Cloud customer content, and that Stock contributors are compensated for training use. Adobe also offers contractual IP indemnification for specified eligible Firefly features and customers, subject to terms and exclusions. These are Adobe’s own representations about its product and contracts, not an independent finding that Firefly eliminates infringement risk. Source record: DA-CH15-S7. Archive reference: pending.

8. U.S. Copyright Office, “Copyright and Artificial Intelligence, Part 1: Digital Replicas,” July 31, 2024. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-1-Digital-Replicas-Report.pdf. Accessed September 11, 2026. The Office recommends a federal right protecting individuals during their lifetimes against knowing distribution of unauthorized digital replicas of their voices or likenesses, with licensing, remedies, safe-harbor, and First Amendment provisions. The recommendation reflects gaps the Office identified in the existing mix of state and federal protections; Congress had not enacted the proposed comprehensive federal right as described in the report at the time of access. Source record: DA-CH15-S8. Archive reference: pending.

9. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153, U.S. Court of Appeals for the Third Circuit, opinion filed September 29, 2026 and unsealed September 30, 2026. https://www2.ca3.uscourts.gov/opinarch/252153p.pdf. Accessed October 1, 2026. The court rejected fair use for Westlaw headnotes used to develop a competing legal-research system, emphasizing similar commercial purpose and expressly distinguishing nongenerative Ross technology from generative-AI disputes. For the appeal, Ross did not contest attribution of its contractors’ copying. The decision does not resolve every AI-training case; supplier-chain questions in the text are author analysis. Source record: DA-CH15-S9. Archive reference: pending.