Creative Work and Authorship

Two claims sit inside this fear and they are usually stated together. The first is economic: work that used to be commissioned from people is now generated, so the paid market for spot illustration, stock imagery, commercial translation, background music, corporate copy and voice work shrinks. The second is legal: the material used to train the systems was largely taken without licence or payment, so the people whose work made the systems possible are the people the systems compete against.

The United States Copyright Office gave the economic claim an unusually precise formulation in the third part of its Copyright and Artificial Intelligence report, published in pre-publication form on 9 May 2025. It argued that the speed and scale at which generative systems produce material can dilute the market for works of the same kind as those in the training data, even where no individual output copies any individual input, and that the copying involved in training therefore threatens harm to the market for or value of copyrighted works.^[1] That theory of market dilution is contested, and it goes beyond the usual fair use question of whether a specific output substitutes for a specific work.

The authorship strand is narrower and better settled. The Copyright Office concluded in January 2025 that human control over creative expression is required for registration, that prompts alone do not generally supply it, and that copyright can attach where human authored material is perceptible in the output or where a person makes creative arrangements or modifications.^[2] The Court of Appeals for the District of Columbia Circuit held on 18 March 2025 in Thaler v. Perlmutter that the Copyright Act requires a human author as a matter of statutory law.^[3] The Supreme Court denied certiorari on 2 March 2026, leaving that holding in place.^[4]

The most direct evidence comes from surveys of the affected workers. The Society of Authors ran a survey in January 2024 and received 787 responses from its 12,500 members and other authors. Among respondents, 26% of illustrators and 36% of translators said they had already lost work to generative AI, while 37% of illustrators and 43% of translators said the income from their work had fallen in value because of it. Looking forward, 78% of illustrators and 77% of translators expected generative AI to reduce their future creative income. Around 5% of illustrators and 8% of translators reported using generative AI because a publisher or commissioning organisation had asked them to.^[5]

Platform data points the same way. Xiang Hui, Oren Reshef and Luofeng Zhou examined a large online freelance marketplace after the release of ChatGPT, DALL-E 2 and Midjourney and found declines in both the number of jobs and total earnings for writing and image related freelancers, with suggestive evidence that highly rated freelancers were disproportionately affected rather than protected.^[6] The World Economic Forum's Future of Jobs Report 2025, built on employer surveys, moved graphic designers from a moderately growing category in its 2023 edition into its list of fastest declining roles, attributing the change to generative AI's growing capacity for knowledge work.^[9]

Music is the sector where volume is easiest to count. Deezer reported on 21 July 2026 that around 90,000 fully AI generated tracks were being uploaded each day and that these had passed half of all new uploads at their June 2026 peak, after what the company described as a steady rise in fully synthetic content since January 2025. The same report put AI generated music at 1% to 3% of actual streams, with up to 85% of those streams judged fraudulent during 2025 against streaming fraud of 8% across the whole catalogue that year.^[7]

Now that half of all daily uploads are AI-generated tracks, we are taking additional steps to safeguard the rights of artists and songwriters, while maintaining focus on music that fans actually love.

The displacement is concentrated at the bottom of each market, which is where careers start. Spot illustration, stock photography, product description copy, subtitling and library music are the commissions that let people build portfolios and pay rent while developing the work they are known for. Removing them does not immediately reduce the number of celebrated artists; it removes the ladder. The Society of Authors data shows the effect landing hardest on translators and illustrators, two occupations that are almost entirely freelance.^[5]

The second mechanism is dilution rather than substitution. If half of new music uploads are machine generated, human recordings compete for attention in a catalogue growing far faster than listening time, and per track royalties fall even if total revenue rises. Deezer's figures show the shape of this: enormous upload volume, small stream share, and a large fraud component attached to the machine generated portion.^[7] The Copyright Office's market dilution theory describes the same dynamic for text and images.^[1]

The third is consent and replication of a specific person. Voice, likeness and performance style can now be reproduced closely enough to substitute for a hire. That is why the digital replica provisions in performer agreements have become the main bargaining subject in entertainment: the object being protected is not a work but a person's identifiable performance.^[16]

The fourth is that legal outcomes may not track economic harm. Getty Images abandoned its primary copyright claim in the United Kingdom after accepting there was no evidence the training took place there, and the High Court dismissed the secondary infringement claim.^[12] If training can be sited in whichever jurisdiction is most permissive, national copyright reform has limited reach over it.

The economist Joel Waldfogel documented that the last technological upheaval in creative industries produced the opposite of the predicted collapse. In Digital Renaissance he showed that as digitisation cut the costs of production, distribution and promotion, the quantity of books, songs, films and television produced rose sharply, and that measures of quality drawn from bestseller and critical lists did not fall.^[17] The industries whose gatekeepers were bypassed predicted ruin; the output data did not support them. That is a base rate worth holding against current predictions, not a proof that this time will be the same.

Sector revenue also keeps rising. IFPI reported global recorded music revenues of 31.7 billion US dollars in 2025, up 6.4% and the eleventh consecutive year of growth, with 837 million paid subscription accounts.^[8] Aggregate revenue is not the same as creator income, and the two can move apart, but a sector in revenue decline would be much easier evidence for the fear than a sector at a record high.

On the legal side, several scholars argue that training is largely lawful and that framing it as theft misdescribes what happens. Matthew Sag has argued that training on text is predominantly non expressive use, closer to indexing than to copying expression, and therefore within fair use in most circumstances, while accepting that memorisation and output substitution raise separate problems.^[18] The courts have not endorsed the maximal rights holder position either. Judge William Alsup held in Bartz that training on lawfully acquired books was fair use while retaining pirated copies for a permanent library was not, and the settlement that followed releases only past conduct, leaving claims based on AI outputs, claims about future conduct and claims over works absent from the works list intact.^[11] In the United Kingdom the High Court went further against the rights holder: Getty's secondary infringement claim failed because the model weights were not held to be an infringing copy.^[12]

Finally, supply is not demand. Deezer's own data shows machine generated tracks passing half of daily uploads while accounting for 1% to 3% of streams, and a large share of even that listening being fraudulent rather than human.^[7] Flooding a catalogue is cheap. Being listened to is not. The same asymmetry applies to text and images, and it limits how much of the market machine generated material has actually taken as opposed to occupied.

Collective bargaining has produced the most specific protections. The 2023 Writers Guild of America minimum basic agreement states that neither traditional nor generative AI is a writer, so no material produced by it counts as literary material; that a company may not require a writer to use AI; that companies must disclose when material given to a writer was generated by or incorporates AI; and that the Guild reserves the right to assert that using writers' material to train AI is prohibited.^[15] SAG-AFTRA members ratified a 2025 Interactive Media Agreement in July 2025 requiring separate, written, reasonably specific consent before a performer's digital replica is created or used, with consent invalidated if the use departs from the description given.^[16]

Regulators have moved on transparency rather than on payment. Article 53 of the EU AI Act requires providers of general purpose models to put in place a policy for complying with Union copyright law, including honouring text and data mining reservations, and to publish a sufficiently detailed summary of the content used for training using a template supplied by the AI Office.^[13] The European Commission published that template on 24 July 2025.^[19]

The United Kingdom consulted and then declined to choose. Its progress statement of 15 December 2025 under section 137 of the Data (Use and Access) Act 2025 reported that 88% of consultation respondents supported requiring licences in all cases, while 7% preferred no change and only 3% supported the data mining exception with rights reservation that the government had named as its own preferred option. The statement endorsed none of the reform options, committing instead to four expert working groups covering control and technical standards, information and transparency, licensing, and wider support for creatives, and to an economic impact assessment required by section 135 of the same Act.^[14]

Compensation has so far come through litigation rather than legislation. Judge Araceli Martinez-Olguin granted final approval to the Bartz settlement on 20 July 2026, at 1.5 billion US dollars and roughly 3,000 dollars per work, which the Authors Guild noted is four times the 750 dollar statutory minimum that is the most common award in ordinary copyright cases.^[11] Platforms have added their own measures: Deezer tags AI generated tracks and excludes them from algorithmic recommendations and editorial playlists, and reports a true positive detection rate of 99.8%.^[7]

Separate the labour question from the legal question, because they can resolve in opposite directions. A court can find training lawful while commercial illustration income keeps falling, and a settlement can pay authors for past scraping without changing what commissions exist next year. The Society of Authors data is about income. The Copyright Office reports and the Getty judgment are about rights. Neither answers the other.

Watch commissioned rates and the number of people earning a living from creative work, rather than sector revenue. Recorded music revenue at a record high and translators losing a third of their commissions are compatible facts, because revenue accrues to catalogues and platforms while commissions accrue to individuals. Where a study reports growth, check whether the unit is money, output volume or people paid.

Treat volume statistics carefully in both directions. Half of daily music uploads being machine generated says a great deal about the cost of production and very little about what audiences choose. The measure that matters for creative livelihoods is the share of paid work, and in most creative fields nobody currently publishes it.