AI-Proof: What Machines Can Never Be — Volume II: Application, Part IV: Conscious Creation, Chapter 20: Intellectual Property in the Age of Consciousness

Part IV looks at what we make. After products in Chapter 17, organisations in Chapter 18 and roles in Chapter 19, this last chapter of Part IV asks what is still yours when anyone can make what you make.

← Previous: Chapter 19: Rethinking Roles and Structure · Series hub · Next: Part V: Conscious Strategy (overview) →

Arjun deleted 312 files because the work spoke for itself. Then a box from Shenzhen looked like his, and certainty was not proof. The law does not ask who was aware. It asks what you chose, and someone must answer.

In Brief: For readers who make things with machines and wonder what is still theirs. Chapter 20 asks what is still yours when a machine can make what you make, and may have learned from what you made. A designer in Mumbai cannot show how his packaging was made; a photographer in Berlin finds her way of seeing on sale for almost nothing; an annotator in Noida teaches a model that will never carry her name; a law firm in Kobe finds its lawyers’ judgement inside a machine; a founder in Daejeon learns what her patents protect. Jefferson and Locke on owning ideas, Ekalavya’s thumb, Learned Hand on Keats, the Bṛhadāraṇyaka’s line of teachers, Benjamin on the original and Foucault on the author help separate what the law lets you own, which changes by country and year, from what you can answer for, which does not.

What the law lets you own changes with country and year; what stays yours is the work you can answer for, and that is a claim of authorship, not a promise of income.

Opening: The 312 Files in Parel

There were 312 files in the folder, and Arjun Mehta selected all of them.

It was June 2025, in his studio in Parel, above the printing shop. A brand and packaging designer, he had just delivered a brand identity and packaging for an audio company in Mumbai. The folder held the work underneath the work: explorations made with a machine, colour studies, versions of the type, dead ends. His friend Imran, a photographer, was sitting on the other chair. In 2024 a client had used one of his photographs far beyond its licence, and his dated original files had settled the argument in a week. Since then he kept everything.

“Keep those,” Imran said. “Every version. You will need them one day.”

Arjun listened politely, with the private certainty that the advice did not apply to him. Months earlier he had turned off automatic backup to the cloud, because his clients, the audio company among them, asked that unreleased work stay off other companies’ servers. The finished design was on the audio company’s server. The rest, to him, was scaffolding.

He deleted the files with the specific satisfaction of a craftsman cleaning his bench.

“You will regret that,” Imran said.

“The work speaks for itself.”

In October 2025 the audio company forwarded him a link. It was a listing for headphones made by a company in Shenzhen. The box had the same curved space around the product, a similar contrast of heavy and light type, the same underlying geometry. Placed side by side, the two boxes looked like the work of one studio. Arjun had never seen the Shenzhen box. He was certain of this. But certainty, he was about to discover, is not the same as proof.

The audio company’s lawyers worked on it for three weeks, and they asked two questions. The first was quickly settled. The Shenzhen design had been registered in China in 2024 and on sale there since late that year, months before Arjun began his sketches in April 2025. Under China’s patent law¹, nobody may make or sell a product with a registered design without the owner’s permission. The test is permission, not copying. Whether Arjun had ever seen the Shenzhen box did not enter the question. For the Chinese market, his box would have to change.

The second question was the one that hurt. Outside China, the danger was a claim that he had copied. There the law asks something different, and a designer who did not copy has a complete defence, if he can show it. Could he show how the design had grown, step by step? He opened the project folder. It was empty. The backup was off. The tool he had used kept his history for thirty days. His first email to the client showed only the finished design. His phone still held photographs of old Bakelite radios at Chor Bazaar, the market where Mumbai sells its old things, dated by the phone. They were the start of the curve, he said. Everything after them existed in his memory and nowhere else.

Then the lawyers ran an image search across registered designs and old catalogues. It found the curved space around a speaker in the grille of a radio from the 1950s, and in a speaker box older than either design. Both boxes had drawn on a shared past. The search was more honest about the origin of his design than his memory had been. He had remembered the radios at Chor Bazaar as his discovery. They were also everyone’s.

One more thought stayed with him, and it was harder. He had never seen the Shenzhen box. Could he say the same of his tool? The box had been public since 2024. A machine trained on public images after that might have seen it. He could not know, and nobody could tell him.

Nobody sued. In December 2025 the audio company changed the box for the Chinese market. Arjun did the revisions without charge and lost three weeks he could not bill. When he told Imran, Imran did not say that he had warned him. He asked, “So now what?”

Imran’s advice had been half right. Saving everything would not have changed the outcome in China, where only the date of registration counted. It would have been his evidence on the second question. And the decision that cost Arjun his records had not been only his. The audio company had asked for secrecy while the work was made and for proof of process when it was disputed. A contract that demands both, with no safe place for the records, asks for the two sides of a contradiction. The cost falls on the one who obeyed.

So this chapter asks: When a machine can make what you make, and may have learned from what you made, what is still yours, and who decides?

The print edition gave this chapter its title, and the title needs an argument before anything else. It called this the age of consciousness. It said that what a person can still own is “the presence of a specific person, paying specific attention, for specific reasons”. It said that a designer’s records were “evidence of the kind of engagement that only awareness produces”. This edition cannot say that. No law in this chapter asks whether anyone was aware. As 20.1 shows, the laws ask for a human whose choices show in the work, or for the person who made the arrangements. Chinese courts have refused images whose makers kept no record. Even the law that seems to speak of experience does not ask about the machine. Since an amendment of 2018, Japan’s copyright law² has let works be used for purposes such as data analysis. The condition is that “it is not a person’s purpose to personally enjoy or cause another person to enjoy the thoughts or sentiments expressed in that work”. The test is a person’s purpose, not a machine’s inner life. So read the title as the print edition’s question. In an age when people argue about whether machines are conscious, ownership turns on something more checkable: human choices that show, and a person who answers.

The word “yours” hides three different things, and the print edition let them run together. One is what the law lets you stop other people doing: copying your book, selling your design, using your invention. Call that ownership. It is made by each country, for its own reasons, and it changes every year. The second is where a work came from: whose choices made it. Call that origin. The third is who must explain the work, stand behind it and put it right if it harms someone. Call that answering. Run together, they let taste be called “the most valuable IP asset”, though no law protects taste. This chapter keeps the three apart.

20.1 What Survives When Generation Becomes Cheap

Three things are often said to survive when machines can generate anything: data, the workflow around the tools, and taste. Whether any of them gives a business an advantage is a question about markets. This section asks a narrower one: what does the law protect, and what does it not?

Begin with an old court judgment about a poem. In 1936 the American judge Learned Hand³ was deciding a case about a play and a film. He imagined that “by some magic a man who had never known it were to compose anew Keats’s Ode on a Grecian Urn”. Then, Hand wrote, “he would be an ‘author,’ and, if he copyrighted it, others might not copy that poem, though they might of course copy Keats’s”. Originality, in copyright, means that a work came from you. It does not mean that it is new to the world. Two people who never met can each own the same poem, if each truly made it. Novelty, the test of patents and registered designs, is the opposite. It asks what was public before a date, and it does not care who copied whom. Arjun’s two questions were these two tests.

Hand’s magic used to be a thought experiment. Machines have made a version of it ordinary. When many people use the same tools, their work drifts towards the same places. A study of 2024⁴ found that stories written with machine help were rated more creative and were more like one another. A wedding photographer in Jaipur heard a bride’s mother say that her pictures looked just like those of a niece’s wedding, shot by someone else with the same tools. Similar boxes in Mumbai and Shenzhen need not mean that anyone copied. That is one problem: convergence, similar work from shared tools. A second is different. A model can be shaped by one creator’s work without copying any of her pictures, as the photographer in 20.2 fears. Copyright’s answer to copying does not settle what is owed for training.

But Hand’s test rested on something the machine has made hard to check. His author “had never known” Keats’s poem. A person can usually say what he has seen. A machine that learned from millions of pictures may have seen the one that matters, and in 2025 its user usually could not know. Independent creation needs a witness, and the person who would have been the witness no longer sees everything that shaped the work.

What, then, does the law protect? Copyright protects particular expressions. An international treaty⁵ says it plainly: “Copyright protection shall extend to expressions and not to ideas, procedures, methods of operation or mathematical concepts as such”. A way of seeing, a method of working and a taste are not property under that rule. Information kept secret can be protected as a trade secret, while it stays secret. The rest is free to learn. That does not make taste worthless. It makes it not property.

For work made with machines, the law adds a condition, and countries state it differently. On 29 January 2025 the United States Copyright Office⁶ reported on such work. It “can be protected by copyright only where a human author has determined sufficient expressive elements”, “but not the mere provision of prompts”. Protection may also reach work where “a human makes creative arrangements or modifications of the output”. On 2 March 2026 the Supreme Court⁷ declined to hear a challenge to the rule that an author must be human, so a lower court’s ruling of 2025 stands. In China, the Beijing Internet Court⁸ held in November 2023 that an image made with an image generator was protected, because its maker’s prompts and settings showed his choices. In March 2025 a court in Changshu⁹ protected an image made with an image generator and then refined by hand, because the prompt and the changes showed its maker’s choices. The same month a court in Zhangjiagang¹⁰ refused to protect three images of a chair shaped like a butterfly. Their maker could not document what she had chosen and changed. In September 2025 the Beijing court¹¹ rejected recreations of a process made for a lawsuit, because the original records were gone. Britain’s law of 1988¹², which the government proposed in March 2026 to remove, has its own rule.¹³ For computer-generated works, it says “the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”. India’s law¹⁴ says the author is “the person who causes the work to be created”.

These laws are not converging. They ask different questions, and in China and the United States the rulings point in different directions. Arjun works in India, so the Chinese rulings are not his rules. Where records decide, his deleted folder would have been his strongest witness.

Notice what the law has done to the word “author”. For ownership, it asks for a person whose choices show. Where it finds none, a machine’s output can belong to nobody and be free for anyone to copy. For blame, it rarely lets the output stay without someone for long. Chapter 8 put it this way: “in practice the law does not let a text stay without an author for long”. Ownership can be absent for good. Answering is absent only until someone is found. The law has split the author in two, and the two halves are moving apart. The Prologue’s second never is about the half that stays: “a machine’s makers and users will never be free of what it does”. As the Prologue puts it, “the people who make it answer for what it is, and the people who use it answer for what they use it to do”. That is a commitment, not a prediction, and it is the only part of this chapter that does not carry a date.

What helped produce a work, whose choices shaped it, whom the law names as author, and who must put it right need not be the same party. Annotators like Sapna, later in this chapter, are paid to teach machines what a good reply looks like; whose work is inside the tool Arjun used, he cannot know. Arjun’s choices shaped the box. India’s copyright law names him its author, and his contract gave the copyright to the client.¹⁵ If the box infringes someone’s rights, the client who sells it, the company that sold the tool and Arjun should each have something different to put right. Each share follows what each controlled: the client what it sells, the toolmaker what its tool learned from, and Arjun what he chose.

So what survives cheap generation, in the law’s sense, is less than is often promised: particular expressions a person made, secrets kept secret, registered designs and marks. A work is yours, in the sense this book cares about, when you can say where it came from, what you chose, and what you will put right.

20.2 Data Ownership in a World of Synthetic Content

Jana Voss has photographed Berlin since 2017. She makes long exposures at the blue hour, the short time between day and night when artificial light and the fading sky have a colour that belongs to neither. Traffic dissolves into rivers of light while the buildings stay sharp. She learned which bridges, at which hours and in which seasons, give the light she wants, by standing on them hundreds of times. She charges about €400 to license an image.

In November 2024 a client sent her a link to a library of stock images made by a machine. “Have you seen these?” the email asked. “They look like your work.” They did. They were not copies. They had her grey-lavender palette and her rivers of light. They were less precise, and they cost almost nothing. She could not prove that any machine had trained on her pictures. Some of them had been on platforms whose terms, which she had accepted years before, allowed it. “The machine has eaten my eye,” she told the client, “and now it is selling the view.” She meant it, and she could not prove it.

The buyers who chose the cheaper images were not wrong. Most of them needed a banner for a website, and the machine’s picture did that job at a tiny fraction of her price. This is replacement by substitution: work she was paid for, now done by a machine. Her licence sales fell by about a third in 2025, and they have not come back. She still photographs at the blue hour. What she lost was income, not her eye.

She did what the law allows. In 2025 she added a reservation of rights to her website, in a form that machines can read. An EU directive of 2019¹⁶ lets machines mine works they can lawfully reach. The owner can refuse, but only “in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. In December 2025 a German appeal court¹⁷ ruled against a photographer, Robert Kneschke, who had sued the makers of a training dataset. It held that a reservation written in plain words on a picture agency’s website in 2021 was not enough; it had to be one that “can be interpreted by machines”. It also held that the copying was scientific research, which no reservation can stop. Germany’s highest civil court heard the case on 3 September 2026 and said it might ask the EU’s Court of Justice. It is due to give its decision on 17 December 2026. Since August 2025 the EU’s AI Act¹⁸ has also required the makers of general-purpose models to “identify and comply with” such reservations. Her reservation protects her future pictures. It does nothing for the ones already learned. A photographer was asked to speak to machines in their own language, while the companies that built the machines wrote the terms of the platforms she had used.

Has the machine taken something from her? Two old voices disagree. John Locke¹⁹, writing in 1690, built his account of property on the claim that “every man has a property in his own person”. From there he argued that a person who mixes labour with what is common makes it his. On that view, Jana’s eye, built by years of her labour, is hers, and a machine that learned it took it.

Locke’s argument meets its plainest test in facts gathered by labour. Chapter 6 told of Tomas Villanueva, who runs a grocery price tracker in Cebu City, in the Philippines. His team walks Carbon Market every Saturday to check the prices of rice, cooking oil and tinned sardines, and he publishes them. In 2025 shoppers began to ask the assistants on their phones instead, and sometimes the answers used his prices. Locke would say the walking made the prices his. The law where Tomas works says otherwise. The Philippines’ copyright law⁵ leaves “mere data as such” unprotected. The United States Supreme Court gave its reason for the same rule in 1991. Facts “may not be copyrighted”, it held, because nobody is the author of a fact. The old “sweat of the brow” doctrine, which protected effort alone, “flouted basic copyright principles”. The European Union adds a separate, narrower right for databases. It protects the maker of a database who has made a substantial investment in obtaining, checking or presenting its contents. That maker can stop others from taking or re-using a substantial part of it. None of these says what Tomas is owed when an assistant repeats a week’s prices and sends nobody to him. In the law’s sense the prices are not his. In this chapter’s sense they are, because he can say where they came from and stand behind them: his team checked them on foot. That claim is real, and it earns him nothing by itself. Whether those who profit from his checking owe him a link, a fee or credit is a question of fairness, and copyright leaves it open.

Thomas Jefferson²⁰, writing in 1813 about a quarrel over a patent, disagreed with Locke about ideas. “He who receives an idea from me, receives instruction himself without lessening mine; as he who lights his taper at mine, receives light without darkening me.” A taper is a thin candle. On Jefferson’s view, nothing was taken from Jana. She still has her eye. The machine lit its taper at hers. Yet an idea can be shared without losing any light and still cost its maker money. If the machine did learn from Jana, her licence sales show what that can cost. Neither fact settles what the law should allow.

An Indian story stands between them, and it is not comfortable. In the Mahābhārata, the great Sanskrit epic, Ekalavya, a prince of the Niṣādas, a forest people, asks the teacher Droṇa to teach him archery. In Kisari Mohan Ganguli’s translation²¹, Droṇa, “cognisant of all rules of morality”, refused him, “seeing that he was a Nishada who might (in time) excel all his high-born pupils”. The Niṣādas stood outside the ranked order of Vedic society. The boy went back to the forest, and “there he made a clay-image of Drona, and began to worship it respectfully, as if it was his real preceptor”. He practised before the image until he shot better than anyone. Droṇa’s favourite pupil, Arjuna, found out and complained: “Thou hadst lovingly told me, clasping me, to thy bosom, that no pupil of thine should be equal to me.” So Droṇa went to the forest and asked for his fee, what the tradition calls the dakṣiṇā a pupil owes a teacher. “I should like then to have the thumb of thy right hand.” The boy “with a cheerful face and an unafflicted heart cut off without ado his thumb, and gave it unto Drona”. Afterwards “he found, O king, that he had lost his former lightness of hand”. “And at this Arjuna became happy, the fever (of jealousy) having left him.”

The epic tells the story as an example of a pupil’s devotion. Many modern readers, especially Dalit and Adivasi readers, hear it as a story of exclusion. The economist Narendra Jadhav²² wrote about the education of Dalits and Adivasis. Ekalavya, he said, “was denied empowerment through education by Guru Dronacharya, who embodied the high-born and hence the establishment”. This book refuses Droṇa’s reason and his fee. The story is about a teacher, a forest prince and a promise to a favourite pupil, not about property in ideas. I borrow only its question: what does a learner owe the one he learned from?

Set beside Jefferson, the story shows something exact. Droṇa lost nothing of his archery when Ekalavya learned. Jefferson is right about the light. What Droṇa feared to lose was a position: his promise that Arjuna would be first. The thumb was taken to protect a place, not knowledge. The two lamps address different questions. Learning takes no light from anyone. It can take away a place, and that is a real loss.

Two American judges restated the same split within two days in June 2025. Both cases were about machines trained on books. Judge William Alsup²³ compared training to teaching. The complaint “is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works,” he wrote. He added: “The Act seeks to advance original works of authorship, not to protect authors against competition.” That is Jefferson’s question: is anything taken by learning? Judge Vince Chhabria²⁴ disagreed. On the market, teaching children, he wrote, “is not remotely like using books to create a product that a single individual could employ to generate countless competing works”. That is the question behind Droṇa’s demand: what happens to the master’s place? Neither ruling settled the law. Chhabria wrote that his ruling “does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful”. In September 2026 a federal appeals court²⁵ ruled that a legal research company’s copying was not fair use. Its machine, the court said, could not generate new text, and the court left the generative cases open. How the copies were obtained is a separate question. Alsup held that using pirated copies to build a central library of books was not fair use, and the case was later settled for $1.5 billion.

European courts were asked a different question: does a model hold what it learned? On 4 November 2025 the High Court in London²⁶ held that the image model Stable Diffusion did not store or reproduce Getty Images’ photographs. So the model was not an “infringing copy”. Getty’s appeal was still pending in October 2026. A week later a court in Munich²⁷ held that song lyrics memorised by a language model were reproduced inside it; OpenAI appealed. The two courts were partly asking about different things. A lyric can be memorised word for word. The court in London found no copy of the photographs in the model. Whether Jana’s eye is inside a machine depends on what the machine kept, and that differs from model to model.

Whatever the courts decide, the remedy should not be the thumb. A rule that stopped learning would protect a few places by darkening many tapers. The honest question is who shares the gain from what was learned. In December 2025 a committee of India’s Department for Promotion of Industry and Internal Trade²⁸ proposed a licence for training. Developers could train on lawfully obtained works without asking, and would pay royalties into a central fund. As of October 2026 it was still a proposal, not law. In July 2026 the Delhi High Court²⁹ refused the news agency ANI an interim injunction against OpenAI. It took the view, for now, that training was fair dealing. ANI has appealed, and the case continues. Neither position is settled. Both say that the question is about sharing, not about thumbs.

There are people inside the machine, too, and they are paid to be there. In August 2026 Sapna sits at a desk in an office in Noida, outside Delhi. She left her phone at the door, as the client requires. Through the day she draws boxes round cars, people and animals in video frames. On some days she ranks a chatbot’s replies in Hindi: which is more accurate, which is kinder, which is wrong. She earns about ₹14,000 a month. Reporters³⁰ found pay of ₹10,000 to ₹15,000 a month at one such firm in Noida in 2026, and their headline called such jobs low-paying. The industry body NASSCOM expects up to a million such workers in India by 2030. Her judgement, thousands of times a day, teaches a model what a correct reply looks like. The model will never carry her name. Mary Gray and Siddharth Suri³¹ called this kind of hidden human labour “ghost work”.

Locke speaks to Sapna’s case, and not in the way his admirers expect. In the next section, Locke lists what becomes his own: “the grass my horse has bit; the turfs my servant has cut; and the ore I have digged”. All of it, he says, becomes “my property, without the assignation or consent of any body”. The servant cuts the turf, and the turf is the master’s. On a common reading, Locke’s theory gives Sapna’s labour to whoever pays her wage. The political theorist C. B. Macpherson read Locke’s servant as a wage labourer whose work his master buys; James Tully disputed that reading.³² On Macpherson’s reading she has been paid, and the judgement she put into the model is someone else’s. I am applying a reading of Locke to a case he never imagined. What her contract allows is one question. What a fair arrangement owes her is another. It has four parts that need not rise together: fair pay for the work, credit for it, a say in how it is reused, and a share of later gains. A worker may be owed the first without the last, or want a say more than a royalty.

Other arrangements are possible. In 2023 TIME³³ reported on Karya, a non-profit in India that pays workers to record speech for training. It pays “at least $5.00 per hour”, and it “pays them again every time a company licenses it to build a new AI”. The questions to ask of such a scheme are plain: whether the workers understand the terms, whether they can refuse a use, and who controls the relicensing. Karya is one design among several.

20.3 When the Model Encodes Your Judgement

Yamamoto Shigeru is a senior partner at a law firm in Kobe. For thirty-two years the firm has handled disputes between Japanese manufacturers and their distributors abroad. The disputes are decided under contract law and in commercial arbitration, before private panels the parties choose.

In 2025 the firm trained a model on its own archive: analyses, memos and assessments written over three decades. Given a new dispute, it drafted a first assessment in under an hour. The assessments were good. Once, it showed the partners a pattern they had not seen: how tribunals had read one common clause about delivery dates over many years. The model was right, and the partners changed their advice. It also carried the firm’s old habits, including advice the firm would no longer give.

The partners saw the model as the firm’s expertise made usable by a machine. Nobody asked whose expertise. Three groups have a claim, and none was asked.

The first are the lawyers who wrote the memos. In early 2026 a senior associate who had written many of them left for a rival firm. Her skill went with her. A copy of her judgement stayed behind, in a model she can no longer use. Apprentices have always taken their masters’ judgement without paying for it; that is how crafts survive. Here the direction is reversed. The firm keeps a working copy of what its people learned, after they leave. The firm can reply that it paid for the memos: she wrote them for its clients, in its time. That settles who keeps the files. It does not settle what is owed for the judgement in them, any more than Sapna’s wage did.

The second are the clients. Their confidential disputes fill the archive. In Japan, as in most countries, a lawyer’s duty to keep a client’s secrets is set in law; the Attorney Act³⁴ makes it both a right and a duty. Training a model inside the firm may not break that duty. Possessing a client’s files is not a licence to use them in every way. Whether clients should be asked before their disputes teach a machine is a different question. In spring 2026 the partners had to decide it, and the decision was theirs to make and to answer for.

The third are the juniors who are not there. First assessments were how young lawyers learned. Chapter 4 called this the missing rung, and the firm’s model sits on it. The firm keeps its old conclusions and may lose the people who knew how to question them.

When your decisions train a model you do not own, your judgement becomes someone else’s asset. That is true of Sapna, of the associate who left, and in a smaller way of Jana. It does not make the model wicked. It means that someone decided whose judgement it would hold, and someone must answer for that decision.

The same question reaches records that nobody has written yet. Ishikawa Kenji has made steel dies in a toolmaking shop in Nagoya for twenty-six years. Twice he has seen a faint ridge on a new die become a crack in a customer’s press. When the shop hears of a failure, nobody writes the reason next to the die’s number, which is partly why that knowledge lives in his memory. If the shop began to write it, a junior could learn from the record, and so could a model. The shop would gain, and so would its customers and whoever owned the model. A junior would not need twenty-six years. Kenji would lose something: he would no longer look like the only person who could know. What he is owed is the question this chapter asked about Sapna, and its four parts need not rise together. He might want credit, and a say in who learns from the record, more than a fee. Smaller records kept at work raise it too: a supervisor’s one-line reasons for overruling a camera, or the log of a quality manager’s flags. The employer usually keeps them. What they hold is a person’s reasons.

Arjun’s ways of instructing the machine are not property. Nor is his habit of testing each design against three arguments in the voices of those who would live with it. Anyone can learn them. The habit came from a website he designed for a hotel chain in Goa, rejected in 2024. It is a habit of work, and it is worth more as a habit than it would be as an asset.

20.4 Defensive and Offensive Strategy

The print edition divided strategy into defence and offence. Defence was patents and trademarks. Offence meant “creating so well and so specifically that the competition can’t keep up”, because rivals “lack the judgement, the taste”. No one can promise that. A rival with the same tools may have more judgement, or none and better luck. In law, defence is protecting what you have registered or made. Offence is asserting your rights against others. Both cost money, and they are not equally open to everyone.

You can’t sue your way out of a statistical distribution. That holds for Arjun. When two designers converge because their tools pull the same way, there is nobody to sue. It does not hold for everyone. Organisations that hold large catalogues have sued, and some have been paid.

Lekan Oyerinde is a music producer in Ibadan. He builds the rhythms that singers sing over: five or six patterns of drums and percussion, locked together so that they sound simple. He has kept a production diary since 2012, with dated stems, the separate tracks of each recording, and handwritten notes. In 2025 he began to hear machine-made tracks with rhythms he recognised. He could not prove where they came from, and he could not afford to sue in another country. Similarity is not proof of copying, and its absence is not proof that nothing was used. Usually only the toolmaker knows in detail what it trained on, so the cost of proof should not fall automatically on the party least able to get the evidence.

Others could. In November 2025 Warner Music Group³⁵ settled its lawsuit with Suno, a company whose tools make songs. Suno promised “fully licensed models”, and the label’s artists and songwriters were offered an opt-in for the use of their “name, image, likeness, voice, and compositions”. On 31 July 2026 a court in Munich³⁶ ruled against Suno in a case brought by GEMA, Germany’s society that collects royalties for composers; the judgment was not final. These settlements and judgments protect catalogues held by labels and collecting societies. Lekan’s rhythms were never registered with either. He is outside both the lawsuits and the deals. His diary shows his practice to anyone who reads it. So far no institution reads it. And if one did, recognition is not one currency. Credit, a fee and the right to refuse a use are different things, and he might want one without the others.

Arjun’s position was like Lekan’s on a smaller scale. His contract said that the work was original. Nobody had asked what that means when a tool may have seen everything. Some of the companies that sell these tools promise to defend their customers. In September 2023 Microsoft³⁷ announced that if a commercial customer was “challenged on copyright grounds, we will assume responsibility for the potential legal risks involved”. The promise covered paying commercial customers who had used the built-in safeguards. A designer working alone carries the risk himself.

So who should carry it? Chapter 13 warned that advice to adapt can move a cost from the institutions that chose a change to the people it was done to. Advice to “keep records” can do the same. Toolmakers could keep a full version history by default. Clients could pay for a record of how the work was made, and say in the contract who pays if the tool turns out to have copied. Until they do, the smallest party pays for the toolmaker’s convergence with his own hours.

There is also offence of a gentler kind: deciding what to give away. Arjun’s curve came partly from a radio grille whose designer nobody can name. The dusty rose of his label for a rose sharbat came from a sharbat maker on Mohammed Ali Road and from his mother’s silence. He owed them nothing in law. He owed them the truth about where his work came from, and saying what he borrowed costs nothing.

20.5 The Patent Illusion

Kang Hae-won founded a company in Daejeon, the South Korean city of research institutes. Its software predicts how much energy a building will use before it is built. From 2021 to 2023 her team developed new ways of modelling heat in buildings. Between 2023 and 2024 her patent attorney secured seven patents on them.

By 2025 three competitors had products with comparable accuracy. None infringed her patents. They had reached the same place by other mathematical routes. She believes that machines helped them search for those routes faster. That may be true; nobody has measured it. Designing around a patent is as old as patents. What she still had was what no patent covered. She had energy data from 1,400 buildings across South Korea, with their materials, occupants and weather. She had the architects whose projects produced the data. And her team knew which variables mattered and which were noise. Data is not protection merely because it was hard to gather. It can grow stale, carry the habits of the buildings it happened to measure, or be matched by someone else’s.

The print edition’s lesson comes down to one sentence: a patent protects what can be written down, and the valuable part increasingly cannot be. The first half is exact. A patent protects the claims written in it, and nothing else. The second half needs a limit. In Kang’s field, as she found it, the valuable part was mostly data and relationships, which a patent claim cannot hold. Whether that is increasingly true in every field is a claim, not a finding.

Why do patents exist at all? Jefferson gave his reason in the same letter of 1813: “Inventions then cannot, in nature, be a subject of property.” Then: “Society may give an exclusive right to the profits arising from them, as an encouragement to men to pursue ideas which may produce utility”. But this, he added, “may or may not be done, according to the will and convenience of the society, without claim or complaint from anybody”. On his view, a patent is a bargain made by a society, not a natural right. The inventor publishes the invention and gets a limited monopoly in return. If machines make some kinds of invention cheaper, the bargain may be worth less to society in those fields. That is a question for legislators, and it carries a date.

The bargain still needs a human on the inventor’s side. In December 2023 the UK Supreme Court³⁸ held that “an inventor within the meaning of the 1977 Act must be a natural person”. In November 2025 the United States Patent and Trademark Office³⁹ repeated that “only natural persons can be inventors”. It compared machines to “laboratory equipment, computer software, research databases”. Here again the law asks for a person, not for awareness.

Patents still matter for licensing, for trading rights between companies, and where an invention is slow to reach the market. A new drug must pass years of trials, and a machine that finds a different molecule does not shorten them. But the print edition went too far when it said that Qualcomm’s “entire business model” depends on licensing. In the year to September 2025, licensing earned Qualcomm⁴⁰ $5.6 billion of its $44.3 billion revenue. That is about an eighth of its revenue, though about a quarter of its two main businesses’ earnings before tax. Licensing matters to Qualcomm. It is not the whole company.

Kang kept filing patents. She stopped calling them her protection. Her investors had asked for patents, partly as a sign to other investors that the company was serious, and she had obeyed. A patent can be a signal and still guard little. The illusion is not that patents stopped mattering. It is that owning one guarantees control of an invention’s future.

20.6 Provenance and the One Who Answers

Provenance is the history of a thing: where it came from, who made it, and whose hands it passed through. The print edition said that records of a maker’s process were “awareness made visible”. This edition holds something smaller and more useful. A record shows where a work came from. It is testimony, and like all testimony it can be false.

It can also be true and miss the point. A record can show that a person was present throughout without showing that she decided the result. It can hold a thousand prompts and miss the one judgement that mattered: the choice that changed the work, the error caught, the version refused for a reason. Activity is not contribution. Records help establish a contribution; they do not create it.

Arjun started a project journal in June 2026, six months after the dispute. The audio company’s lawyers had recommended it. He bought a notebook and wrote after each session what he had explored, what he had rejected and why he had chosen what he chose.

For three weeks, the journal was exactly what the cynical reading would predict: a performance. He wrote decisions he had already made in words chosen to sound deliberate. He left out the doubt and the false starts. The entries were competent and dishonest in the ordinary way that reasons written afterwards are dishonest. A court in Beijing had refused exactly such recreations in September 2025.

In July 2026 he was designing packaging for a new tea brand, traditional chai for younger buyers. He had forty-odd directions from the machine, and one held him: a green so dark it was almost black. He liked it and could not say why. He reached for the journal to write his reason and found that he had none. So he wrote a question: why this green?

The next morning, before he opened his tools, he looked at the green again, and a memory came. His grandmother’s kitchen in Jaipur, where his mother grew up. The steel tin where she kept loose tea, green so dark it was almost black, worn at the edges to a lighter shade beneath. The smell of cardamom.

The print edition said that the journal had discovered where the green came from. Perhaps it had. The psychologists Richard Nisbett and Timothy Wilson⁴¹ showed that people often give confident reasons for choices that had little to do with what moved them. The tin may be where the green came from. It may also be a story his memory supplied when the page asked for one. The journal cannot tell which. What it did do is plain. It made him ask. Over the next weeks he caught himself, more than once, about to accept a machine’s direction because it was good enough, and the empty line in the notebook stopped him. The tea brand chose the dark green. That does not prove the journal right; a good outcome does not show a good decision.

This journal has a different purpose from the notebooks earlier in this book. Priya Chandran, a test lead in Bengaluru, kept a page of decisions in Chapter 5 so that she could answer for them. Kavitha Sharma, a science teacher in Sagar, kept a notebook in Chapter 10 to test her eye against a machine’s. The paragraph in Chapter 11 was sealed so that a decision could be judged by what was known when it was made. Arjun’s journal is a record of where a work came from, made to be shown to someone else, and a check on the story he would tell later.

Machines now keep such records too. Content credentials are signed records, attached to a file, of how it was made and changed. They follow a standard written by an industry group, the Coalition for Content Provenance and Authenticity (C2PA). In September 2025 Google⁴² said that its Pixel 10 phones were “the first to have Content Credentials built in across every photo created by Pixel Camera”. The EU’s AI Act was amended by Regulation (EU) 2026/1744⁴³, in force 27 July 2026. Under it, the duty in Article 50(2) to mark machine-made content in a machine-readable way applies to systems already on the market from 2 December 2026. Jana now attaches credentials to her own pictures.

A credential shows a history signed by a device or a program. It does not show that the history is true of the world, or who should be called the author. The standard’s own list of common questions⁴⁴ admits that “metadata can be removed inadvertently or intentionally”. Google itself wrote that “C2PA Content Credentials are not the sole solution for identifying the provenance of digital media”. A missing credential proves nothing either, and a credential that nobody may challenge is theatre.

Two old ideas of where authority comes from help here, and they disagree. In an essay first drafted in 1935, the German critic Walter Benjamin⁴⁵ wrote that “even the most perfect reproduction of a work of art is lacking in one element”. In Harry Zohn’s translation, that element is “its presence in time and space, its unique existence at the place where it happens to be”. Authenticity, he wrote, is “the essence of all that is transmissible from its beginning”, including “its testimony to the history which it has experienced”. For Benjamin, authority lives in the one original and the history it carries. A copy has none.

The Vedic tradition offers another way to think about copies. Its texts were passed from mouth to mouth for centuries, and every recitation was a copy. Yet their authority was not lost in copying, because it travelled with the people who passed them on. The Bṛhadāraṇyaka Upaniṣad ends with such a line of teachers. In F. Max Müller’s translation⁴⁶ it begins “Now follows the stem” and names each teacher as the pupil of another, back through the generations. It ends: “Brahman is Svayambhu, self-existent. Adoration to Brahman!” This line of teachers is called a vaṃśa, a lineage. Chapter 8 met the view of the Mīmāṃsā school that the Veda has no human author. The vaṃśa shows that a text with no author can still have a provenance. It names who passed it on, not who made it.

Digital files share with the recited Veda a condition that Benjamin’s single original does not face: many faithful copies. Both make us ask what a record keeps and who can vouch for it; neither settles authorship. A content credential is a vaṃśa of devices and keys: it records each step, and it can give a machine’s output a provenance without giving it an author. What it lacks is what the Upaniṣad’s list had at every human link: a person who could be asked.

That is what makes a record costly to fake, and it meets the hardest objection to this section. Any system that rewards records of a process can be gamed by producing records. A machine can now write a process diary, dead ends included, in seconds. A record kept only for yourself costs almost nothing to forge. Two things raise the cost. The first is that others hold parts of the record: an email sent to a client, a launch the public saw, a camera’s signature, a registry’s stamp. To fake those, you must deceive people who do not work for you. The second is a person who can be questioned, who must explain the record, and who will lose something if it is false. Chapter 8 asked three questions of any testimony: what supports it, how reliable the source is for this kind of question, and who must answer if it is wrong. Provenance helps with the first two. Only people, and the institutions they act through, can answer the third. If a law one day named a machine as the party liable, the money and the remedy would still come from the people who made it and used it. A name without the power to put right is not yet an answer, and a record without a remedy is only a history.

A record can also be about someone who did not write it. Chapter 15 told of a ledger in Lagos in which cloth traders sign their judgements of suppliers and buyers. What the traders are owed for those judgements is the question this chapter asked about Sapna. The people the entries name are owed something else. In April 2026 the ledger gave them one right: a reply beneath each entry. The law may already owe them more. Nigeria’s Data Protection Act of 2023⁴³ requires anyone who holds personal data to keep it “accurate, complete, not misleading”. A person the data is about can ask for it to be corrected, or deleted if it cannot be corrected. She can object to its use, and then the holder must stop, unless it shows a public interest or other legitimate grounds that outweigh her rights. As I read the Act, the dyer who heard of her entry by chance, from a customer, is such a person. A reply lets her be heard. The law gives her a claim to have a wrong entry put right.

The French philosopher Michel Foucault argued in 1969 that authors were first named for an older reason than ownership. In Donald Bouchard and Sherry Simon’s translation⁴⁷, speeches and books “were assigned real authors, other than mythical or important religious figures, only when the author became subject to punishment”. That is his history, and historians argue about it. But it fits what this chapter has found. Authorship answered before it owned. In Plato’s Phaedrus⁴⁸, writing “always needs its father to help it; for it has no power to protect or help itself”. The record does not defend itself either. It needs someone to stand behind it.

Words with history

The English author comes from the Latin auctor, and in Latin the word could name someone who had made nothing. Lewis and Short’s Latin dictionary gives among its senses “a seller, vender”, so called because he warrants the buyer’s right to possess what is sold. It also gives “a voucher, bail, surety, witness”: one who stands security for something.⁴⁷ The author of a written work is another of its senses.

So the word already held both halves that this chapter finds the law pulling apart. One is the maker, whose choices shaped the work. The other is the one who stands behind it, as a seller stood behind a title to goods he had not made. Copyright now looks for the first. Answering looks for the second, and it can find someone even where no human choices show.

So the chapter’s question can be met, in two parts. What the law lets you own is decided by countries, differently, and it will change again before this book is old. What stays yours is the work you can answer for: you can say where it came from, what you chose, and what you will put right. Answering is not what is left of ownership when the law withdraws it. It is a separate obligation, and a person can owe it for work she does not own. It promises no income, as Jana and Lekan know. And the chain has an empty seat. Sapna’s judgement is inside the model, and when the model is sold or harms someone, nobody will ask her.

This is the last chapter of Part IV. The Part began by asking whether a product that survives deserves to exist. At each level, product, organisation and role, it found that what a machine does not take over is the person who answers. This chapter adds that the law will not settle ownership either. That changes what the first Volume meant by “yours”. It was never a skill or an asset someone holds. It is an act that someone must be found to do.

Ending: Two Notebooks in Parel

In October 2026 the studio in Parel looks much as it did. The first Goa website is still pinned on the wall, the one that would not work. Beside the design books there are now two notebooks, the second half full.

At the end of each project Arjun writes one more page and sends it to the client. It says what the machine produced and which of its outputs he used, and what he chose and changed. It names the borrowed things: a sharbat maker on Mohammed Ali Road, a radio grille from the 1950s, a tea tin that may be a memory or a story. No client has asked for it. Two have written back to say thank you. One asked why he bothered.

Some weeks the page is three lines long. Some weeks he forgets it, and writes it from memory a few days later, and marks it so.

The tea boxes arrived in September, dark green. He kept one on his desk. He does not think of the notebooks as property. They are the part of the work he can be asked about.

Practice

The page that leaves the desk. Choose one piece of work that will carry your name and that a machine helped to make. While you work, keep a record made at the time, not afterwards. Note what the machine produced and which outputs you used. Note what you chose, changed and rejected, one line each, with the reason if you have one and “no reason yet” if you do not. Note where any borrowed element came from: a person, an object, an older design. When the work is done, send the record to someone else: a client, a colleague, or your own email. Then a copy is held where you cannot quietly change it. If a client forbids you to keep working files, ask how a dated record will be kept, who may see it, and for how long. Then write the paragraph you would give someone who asked, “Is this yours?” Compare it with the record, and mark each sentence the record supports. The unmarked sentences are what memory added. This costs a self-employed person unpaid minutes at the end of each session; for an employee it costs the employer’s time, so ask for it. It may not protect you: a design registered first can still win, and courts in different countries ask for different things. It can also show a client or a court which parts a machine made, and in some countries those parts are not protected. It shows what you can answer for.

Words to keep

• answering: explaining a work, standing behind it and putting right what it harms; the sense of “yours” that does not change by country.

• content credentials: signed records attached to a file, showing how it was made and changed (the C2PA standard); they can be removed, and they show history, not truth or authorship.

• dakṣiṇā: the fee or gift owed to a teacher for teaching.

• ghost work: Gray and Suri’s term for hidden human labour behind machines, such as labelling data and rating answers.

• independent creation: making a work without copying another; a full answer to a claim of copying in copyright, and no answer to a registered design or patent.

• novelty: the test of patents and registered designs: whether anything like it was public before the filing date.

• originality: the test of copyright: whether a work came from its author, not whether it is new to the world.

• provenance: the history of a thing: where it came from, who made it, and whose hands it passed through.

• vaṃśa: a lineage; in the Upaniṣads, the list of teachers through whom a teaching was passed on.

Sources

1. Patent Law of the People’s Republic of China (2020 amendment), Article 11 (no exploitation of a design patent without the patentee’s authorisation) and Article 42 (design term of fifteen years), China Justice Observer, China Laws Portal; WIPO, Patent Judicial Guide: China, 4.8, “Design patents” (preliminary examination only; prior design is “any design known to the public domestically and/or abroad before the filing date”).

2. Copyright Act of Japan, Article 30-4 (added by the amendment of 2018), trans. Copyright Research and Information Center (CRIC), “Copyright Law of Japan”, chapter 2.

3. Sheldon v. Metro-Goldwyn Pictures Corp., 81 F.2d 49 (2d Cir. 1936), opinion by Learned Hand (OpenJurist).

4. Anil R. Doshi and Oliver P. Hauser, “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content,” Science Advances 10, eadn5290 (12 July 2024); see Chapter 3.

5. Agreement on Trade-Related Aspects of Intellectual Property Rights (1994), Article 9.2 (WIPO Lex). On facts: Intellectual Property Code of the Philippines (Republic Act No. 8293, 1997), s. 175: no protection extends to any “discovery or mere data as such, even if they are expressed, explained, illustrated or embodied in a work” (ChanRobles Virtual Law Library); Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340, decided 27 March 1991 (Justia): “Facts, whether alone or as part of a compilation, are not original, and therefore may not be copyrighted”; “Without a doubt, the ‘sweat of the brow’ doctrine flouted basic copyright principles”; Directive 96/9/EC on the legal protection of databases, Article 7(1): the maker of a database who shows a substantial investment in obtaining, verifying or presenting its contents may prevent extraction or re-utilisation of the whole or a substantial part of them (text on legislation.gov.uk).

6. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, released 29 January 2025; quotations from the Office’s release of that date (NewsNet Issue 1060). Part 3, on training, was released in pre-publication form on 9 May 2025.

7. Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025); certiorari denied, 2 March 2026 (No. 25-449), as reported by Mayer Brown, “Supreme Court Denies Review in AI Authorship Case,” 11 March 2026.

8. Li v. Liu, Beijing Internet Court, 27 November 2023; summarised in Wolters Kluwer Copyright Blog, “Beijing Internet Court Grants Copyright to AI-Generated Image for the First Time” (2 February 2024).

9. Changshu People’s Court (Jiangsu), the “half heart” case, announced 7 March 2025: an image made with Midjourney and refined in Photoshop held to be an original work of fine art (Society for Computers and Law, “Another Chinese court finds that AI-generated images can be protected by copyright,” March 2025).

10. Zhangjiagang People’s Court (Jiangsu), 19 March 2025, the “butterfly chairs” case, upheld on appeal by the Suzhou Intermediate People’s Court; summarised in Technology’s Legal Edge (DLA Piper), July 2025.

11. Beijing Internet Court, September 2025: copyright refused where the claimant had no original generation records and offered recreations made for the litigation; Linklaters TechInsights, October 2025, on the documentation of creative processes for AI-generated images in China.

12. Copyright, Designs and Patents Act 1988, s. 9(3) (legislation.gov.uk; unamended as of October 2026).

13. UK Government, Report on Copyright and Artificial Intelligence, 18 March 2026, proposing that the protection for computer-generated works in s. 9(3) be removed “in the absence of evidence of its ongoing value” (Herbert Smith Freehills Kramer, 18 March 2026).

14. Copyright Act, 1957 (India), s. 2(d)(vi), inserted by the Copyright (Amendment) Act 1994 (consolidated text, Centre for Internet and Society).

15. Copyright Act, 1957 (India), s. 2(d)(iii) (the author of an artistic work other than a photograph is “the artist”), s. 17 (“the author of a work shall be the first owner of the copyright therein”, subject to the Act; a design drawing is not among the commissioned works in s. 17(b)), s. 19(1) (an assignment is valid only “in writing signed by the assignor”) and s. 57(1) (the author keeps the right “to claim authorship of the work” after assignment) (consolidated text, Centre for Internet and Society). For the box as an industrial design, the Designs Act, 2000, s. 2(j)(i), makes the person who commissioned it the proprietor, and s. 15(2) of the Copyright Act ends copyright in an unregistered registrable design after fifty industrial copies. The assignment in Arjun’s contract is part of the composite.

16. Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market, Article 4(3) (EUR-Lex).

17. Kneschke v. LAION e.V., Landgericht Hamburg, 27 September 2024 (310 O 227/23); Oberlandesgericht Hamburg, 10 December 2025 (5 U 104/24), summarised in Technology’s Legal Edge (DLA Piper), December 2025; the appeal court held the copying covered by both the text-and-data-mining exception and the scientific research exception (§60d UrhG); heard by the Bundesgerichtshof on 3 September 2026 (I ZR 281/25), where the senate said a referral to the Court of Justice of the EU was possible (beck-aktuell, 3 September 2026; Bundesverband professioneller Bildanbieter, 14 September 2026); decision to be pronounced on 17 December 2026 (Bundesgerichtshof, Terminhinweise, checked 8 November 2026).

18. Regulation (EU) 2024/1689 (the Artificial Intelligence Act), Article 53(1)(c) and (d), applicable from 2 August 2025.

19. John Locke, Second Treatise of Government (1690), ch. V (“Of Property”, §§25–51), §27 (opening clause) and §28 (Project Gutenberg, from C. B. Macpherson’s edition, Hackett, 1980). Chapter 14 quotes the rest of §27.

20. Thomas Jefferson to Isaac McPherson, 13 August 1813, in The Writings of Thomas Jefferson, ed. Andrew A. Lipscomb and Albert Ellery Bergh (Washington: Thomas Jefferson Memorial Association, 1905), reproduced in The Founders’ Constitution, Art. 1, Sec. 8, Cl. 8, Doc. 12 (University of Chicago Press). Founders Online prints the manuscript spelling (“recieves”).

21. The Mahabharata of Krishna-Dwaipayana Vyasa, Book 1, Adi Parva, Section CXXXIV, trans. Kisari Mohan Ganguli (Calcutta, 1883–1896; Internet Sacred Text Archive). The episode is Mahābhārata 1.123 in the Critical Edition numbering; Droṇa’s demand for the right thumb is 1.123.35, tam abravīt tvayāṅguṣṭho dakṣiṇo dīyatāṃ mama. Droṇa’s word for the fee is vetana (1.123.33); in 1.123.35 dakṣiṇo means “right”. Dakṣiṇā, or gurudakṣiṇā, is the tradition’s name for a pupil’s gift to a teacher.

22. Narendra Jadhav, Empowerment of Dalits and Adivasis: Role of Education in the Emerging Indian Economy, CASI Working Paper 09-04 (Philadelphia: Center for the Advanced Study of India, University of Pennsylvania, November 2009).

23. Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal.), Order on Fair Use, 23 June 2025, p. 28; settlement of $1.5 billion (about $3,000 a book, over 400,000 books) given final approval on 20 July 2026 (JURIST, 22 July 2026). The same order held that using pirated copies to build a central library was not fair use (its conclusion: “The downloaded pirated copies used to build a central library were not justified by a fair use”).

24. Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal.), Order, 25 June 2025, pp. 3 and 5.

25. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. (3d Cir., 29 September 2026), summarised by Ballard Spahr, “Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved,” 2 October 2026.

26. Getty Images (US) Inc and others v Stability AI Ltd [2025] EWHC 2863 (Ch), 4 November 2025 (Joanna Smith J), para. 600; permission to appeal granted in December 2025; appeal pending as of 15 October 2026 (CMS, AI and copyright case tracker; Herbert Smith Freehills Kramer, November 2025).

27. GEMA v. OpenAI, Landgericht München I, 11 November 2025 (42 O 14139/24), summarised by CMS, 19 November 2025; appealed to the Oberlandesgericht München (GEMA, 31 July 2026).

28. Department for Promotion of Industry and Internal Trade (India), committee working paper on generative AI and copyright, Part 1 (December 2025), summarised by the European Commission’s Intellectual Property Helpdesk, 29 December 2025; consultation closed 6 February 2026; no bill introduced as of May 2026 (Intepat, 13 May 2026), and still a proposal as of 26 October 2026 (Legal 500, “Can A Publisher Say No To AI? India’s Copyright Debate After ANI v OpenAI”).

29. ANI Media Pvt Ltd v. OpenAI OPCO LLC, Delhi High Court (Amit Bansal J), interim order of 24 July 2026, refusing an interim injunction; training held prima facie to be fair dealing under s. 52(1)(a); the suit continues (CMS case tracker; Legal 500); a division bench issued notice on ANI’s appeal against the order on 15 September 2026 (LiveLaw, 15 September 2026).

30. Sagrika Kissu, “AI is creating lakhs of low-paying jobs in Gurugram, Noida. It is India’s new BPO,” ThePrint, 1 September 2026.

31. Mary L. Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Boston: Houghton Mifflin Harcourt, 2019); see also Mary L. Gray, “Paradox of Automation’s Last Mile,” ghostwork.info, 12 January 2019. On ranking model outputs as a training method: Long Ouyang et al., “Training language models to follow instructions with human feedback,” arXiv:2203.02155 (2022): “We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback”.

32. C. B. Macpherson, The Political Theory of Possessive Individualism: Hobbes to Locke (Oxford: Clarendon Press, 1962), on the servant of §28 as a wage labourer; James Tully, A Discourse on Property: John Locke and His Adversaries (Cambridge: Cambridge University Press, 1980), disputing that reading.

33. Billy Perrigo, “Manu Chopra,” TIME100 Most Influential People in AI, TIME, 7 September 2023 (“a nonprofit”; the two quotations); Billy Perrigo, “The Workers Behind AI Rarely See Its Rewards. This Indian Startup Wants to Fix That,” TIME, July 2023 (workers read text aloud to record speech for a speech model).

34. Attorney Act (Japan, Act No. 205 of 1949), Article 23, as summarised in WIPO Standing Committee on the Law of Patents, “Confidentiality of Communication between Clients and their Patent Advisors: Japan”: non-disclosure is a professional right and a professional obligation, unless a statute provides otherwise.

35. Warner Music Group and Suno, settlement and partnership announced 25 November 2025 (SiliconANGLE, 25 November 2025).

36. GEMA v. Suno, Landgericht München I, 31 July 2026 (42 O 763/25), not final (Bristows, 10 August 2026; GEMA, 31 July 2026).

37. Microsoft, “Microsoft announces new Copilot Copyright Commitment for customers,” 7 September 2023; extended as the Customer Copyright Commitment on 15 November 2023.

38. Thaler v Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49, 20 December 2023, press summary (National Archives, Find Case Law).

39. United States Patent and Trademark Office, “Revised Inventorship Guidance for AI-Assisted Inventions,” released 26 November 2025, Federal Register, 28 November 2025, rescinding the guidance of February 2024; the notice restates the Federal Circuit’s holding that “only natural persons can be inventors” (Goodwin, December 2025).

40. Qualcomm Incorporated, fourth-quarter and fiscal 2025 results, Exhibit 99.1 to Form 8-K (SEC): fiscal year ended 28 September 2025, total revenues $44,284 million; QTL (licensing) revenues $5,582 million; QTL earnings before taxes $4,043 million against QCT’s $11,670 million.

41. Richard E. Nisbett and Timothy DeCamp Wilson, “Telling More Than We Can Know: Verbal Reports on Mental Processes,” Psychological Review 84, no. 3 (1977): 231–259; see Chapter 1.

42. Google, “How Pixel and Android are bringing a new level of trust to your images with C2PA Content Credentials,” 10 September 2025.

43. Regulation (EU) 2026/1744 (the Digital Omnibus on AI), amending Regulation (EU) 2024/1689: in force 27 July 2026; Annex III high-risk duties from 2 December 2027; product-embedded from 2 August 2028; Article 50(2) for systems already on the market from 2 December 2026 (White & Case, July 2026). Chapter 25 develops the AI Act. On records about people: Nigeria Data Protection Act, 2023, s. 24(1)(e) (personal data must be “accurate, complete, not misleading, and, where necessary, kept up to date”), s. 34(1)(c) (“the correction or, if correction is not feasible or suitable, deletion” of personal data that is inaccurate, out of date, incomplete or misleading), s. 36(1) (“the right to object to the processing of personal data relating to the data subject”), s. 36(2) (processing stops “unless the data controller demonstrates a public interest or other legitimate grounds” that override the data subject’s rights) and s. 65 (a data subject is “an individual to whom personal data relates”); text as passed by both Houses and certified by the Clerk to the National Assembly (DataGuidance copy; the assent date is illegible in that scan); signed into law on 12 June 2023 (KPMG Nigeria, June 2023).

44. Coalition for Content Provenance and Authenticity, “FAQs” (c2pa.org).

45. Walter Benjamin, “The Work of Art in the Age of Mechanical Reproduction” (first drafted 1935; Zohn translates a later version), section II, trans. Harry Zohn, in Illuminations, ed. Hannah Arendt (New York: Schocken, 1968) (Marxists Internet Archive).

46. Bṛhadāraṇyaka Upaniṣad VI.5 (6.5.1–4), trans. F. Max Müller, The Upanishads, Part II, Sacred Books of the East 15 (Oxford: Clarendon Press, 1884).

47. Michel Foucault, “What Is an Author?” (1969), trans. Donald F. Bouchard and Sherry Simon, in Language, Counter-Memory, Practice (Ithaca: Cornell University Press, 1977), pp. 124–127. For “Words with history” in 20.6: Charlton T. Lewis and Charles Short, A Latin Dictionary (Oxford: Clarendon Press, 1879), s.v. “auctor”: “A seller, vender”, glossed “inasmuch as he warrants the right of possession of the thing to be sold”; “One that becomes security for something, a voucher, bail, surety, witness”; “An author of scientific or literary productions” (text as in the Berlin-Brandenburg Academy’s copy, cld.bbaw.de, checked 4 November 2026).

48. Chapter 8, citing Plato, Phaedrus 275d–e, trans. Harold North Fowler, Plato in Twelve Volumes, vol. 1 (Cambridge, MA: Harvard University Press, 1914).

Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.

Revised from the print edition (2026)

• What the print edition claimed: that intellectual property now belongs to “the age of consciousness”; that what survives is “the presence of a specific person, paying specific attention, for specific reasons”, “the one thing no AI could replicate”; that records are “evidence of the kind of engagement that only awareness produces” and “awareness made visible”, and that “awareness deepened through practice produces something that no shortcut can replicate”; that legal frameworks “are converging”; that taste is “the most valuable IP asset”; that “the workflow is the moat”; that prompting methods are “IP in all but legal name”; that rivals cannot keep up because they “lack the judgement, the taste”; that AI “generates workarounds faster than the patent system processes applications”; that Qualcomm’s “entire business model” depends on licensing; that the Osaka firm’s disputes turned on “WTO regulations”, and that its model was “its expertise made machine-legible”; that content credentials prove “that your work is yours”; that the Lagos producer’s rate “rose by nearly half” and that “he’d been minting it all along”; and that “Building unconsciously is building disposably.” It also had Arjun prove “independent creation” against a Chinese design registration, took the sharbat colour from brass at Chor Bazaar, called the sharbat packaging “famous”, called the Goa rejection a “disaster”, let the journal discover the green as fact, timed the photographer’s light by a “forty-minute window”, ended on a summary of the Part and a bridge to Part V, and named Arjun’s friend Deepak, the producer Kofi Asante, the photographer Lena, the lawyer Yamamoto Kenji in Osaka and the founder Park Jiyeon in Seoul.

• What I now hold: “yours” means three things: what the law lets you stop others doing (ownership, which differs by country and changes every year), where a work came from (origin), and who must explain it, stand behind it and put it right (answering). No law tests awareness; the laws ask for a human whose choices show, or a person who made the arrangements, and they diverge. Copyright asks whether a work came from you; patents and registered designs ask whether it was new, and independent creation is no answer to them. A machine that may have seen everything makes independent creation hard to witness. The law has split the author: machine output can stay without an owner, but rarely for long without someone to answer. What helped produce a work, whose choices shaped it, whom the law names as author and who must put it right need not be the same party. Convergence (similar work from shared tools) is a different problem from influence (a model shaped by one creator without copying her expression), and copyright’s answer to copying does not settle what is owed for training. Learning takes no light from anyone (Jefferson), but it can take a place (Droṇa’s fear) and cost its maker money; neither fact settles what the law should allow, and the remedy should share gains, not take thumbs. On a common reading (Macpherson’s, disputed by Tully), Locke’s theory gives the servant’s labour to the master; what a contract allows is not what a fair arrangement owes, and the latter has four parts that need not rise together (pay, credit, a say in reuse, a share of later gains). Recognition is not one currency, and the cost of proof should not fall automatically on the party least able to get the evidence. A model trained on a firm’s work holds the judgement of people who were not asked; the firm may keep its old conclusions and lose the people who could question them; possessing client files is not a licence for every use. Data is not protection merely because it was hard to gather. A patent protects what is written in its claims; in some fields the valuable part is data and relationships, and whether that is increasingly true everywhere is a claim. Provenance is testimony: costly to fake when parts are held by others and a person can be questioned. Activity is not contribution: a record can show presence without showing who decided the result. Only people, and the institutions they act through, can answer for a record; a name without the power to put right is not yet an answer, and a record without a remedy is only a history. A contract that demands secrecy during the work and proof of process in a dispute, with no safe place for the records, asks for both sides of a contradiction. Answering is not what is left of ownership; it is a separate obligation. The print edition’s title is read as its question. What Tomas is owed for prices an assistant repeats is a question of fairness that copyright leaves open. A firm that paid for its memos keeps the files; that does not settle what is owed for the judgement in them. A record that would let a junior or a model learn from Kenji’s years gains the shop, its customers and the model’s owner; what he is owed has the same four parts. A law that named a machine liable would still need people’s money and remedy. A person named in a record may be owed correction, not only a reply (Nigeria’s Data Protection Act, 2023). An honest record of process can also show which parts are unprotected. How training copies were obtained is a separate question from training (Alsup’s ruling on pirated copies).

• What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do; the law itself, in 2025 and 2026, which asks for persons and choices, not awareness; the split between the judges in San Francisco and between the courts in London and Munich; the reviewers’ objections (October 2026) that a record of activity is not a record of contribution, that a company can answer through its people, and that a wage is a contract, not a moral settlement; the audit’s corrections (the Chinese design, the WTO, Qualcomm, the sharbat colour); Chapter 1’s evidence that people’s reasons are often made afterwards; Chapter 8’s account of testimony and its “father”; Chapter 13’s warning that advice to individuals can move an institution’s cost.

• What survives: the 312 files and “The work speaks for itself”; “certainty ... is not the same as proof”; “So now what?”; the photographer’s eye and “selling the view”; “You can’t sue your way out of a statistical distribution”, now with its limit; the patent illusion, narrowed; the uses patents keep in licensing and in trading rights between companies; the firm whose judgement is in a model; “For three weeks, the journal was exactly what the cynical reading would predict: a performance.”; “why this green?” and the tea tin in Jaipur, now uncertain; provenance as the question of ownership when making is cheap.

• Section titles in the print edition: 20.6 was “Consciousness as Provenance”. The other five keep their print titles.

Next in the series: Part V: Conscious Strategy (overview).