A live test case for who captures AI’s value
Most of this manual’s labor-displacement discussion runs on exposure indices and employer surveys, because economy-wide effects are still mostly forecasts. Creative work is different: it already has dated survey data, signed contracts, a $1.5 billion court settlement, and binding disclosure law. That makes the creator economy one of the few places where “what has AI actually done to a profession” can be answered with named sources rather than projection. It is not a uniform story — some professions report clear income loss, others report growth alongside loss, and the money flowing back to creators is real but unevenly distributed and often contested even after it is committed.
Voice actors: measurable losses, real legislative pushback
The National Association of Voice Actors’ “State of Voiceover 2026” survey of 1,379 members found 21% had lost work directly to AI in 2026, up from 14% the year before, while 9% discovered an unauthorized synthetic clone of their own voice in a professional project. Income effects split: 41% of respondents reported growth, 30% a decline. NAVA presented the findings in Washington in May 2026 while lobbying for the federal NO FAKES Act, which as of this writing remains pending legislation, not enacted law.
The response has been concrete rather than merely rhetorical. SAG-AFTRA’s 2025 commercials contracts require clear consent plus minimum payments for any AI recreation of a performer’s voice, following the 2024–2025 video-game voice-actor strike that won a 15–15.7% initial wage increase and new digital-replica consent terms. Outside the US, Rest of World reports that Mexico has banned unauthorized AI dubbing, South Korean and Brazilian voice actors have pushed for statutory protections, and Watch the Skies (2025) became the first AI-dubbed feature released theatrically in the US, using Flawless AI.
Illustrators: the sharpest numbers in this dataset
The UK’s Association of Illustrators surveyed 6,844 members ahead of the government’s Copyright and AI consultation, which closed February 25, 2025. Its results are the starkest in this file: 32.4% had already lost work to AI, at an average loss of £9,262 per affected illustrator, and more than 99% wanted compensation for past unauthorized use of their work in AI training. That is a self-reported survey from an interested professional body, not a controlled study — but its scale and specificity make it harder to dismiss than anecdote. Separately, freelance-platform data shows Upwork ending 2025 with roughly 47,000 fewer active clients than the year before, its largest annual contraction, while Fiverr’s active buyers fell 13.6% year over year; Upwork’s own materials, unsurprisingly self-interested, frame this differently, reporting AI-related work on its platform growing 60% year over year.
Musicians: labels settle with AI startups, artists sue the labels
Spotify and Universal Music Group announced licensing agreements in May 2026 enabling a fan-made AI covers-and-remixes tool with direct artist and songwriter revenue share. That followed a wave of 2025 label settlements with AI music generators: Warner Music Group settled with Suno in November 2025 around a licensed-model partnership with artist opt-in; Universal Music Group and, separately, Warner settled with Udio the same month. Sony Music has not settled with either company and continues litigating — in May 2026 it moved to expand its claims in the Udio case from 560 to over 61,000 recordings after discovery reportedly showed training on millions of tracks.
Those settlements created a new fight: over who gets the money. The American Federation of Musicians has sued Universal, Warner, and Atlantic, alleging the labels licensed members’ recordings to Suno and Udio for training without musician compensation or credit, and are withholding settlement and future licensing revenue from the artists whose recordings were licensed. No damages figure has been disclosed. Separately, a coalition including the RIAA, IFPI, and SAG-AFTRA has proposed voluntary “AI-generated” and “AI-assisted” streaming tags — a labeling effort running in parallel to the licensing fights, not a resolution of them.
Writers and journalism: compensation demanded, causation contested
The Authors Guild’s long-standing survey of more than 1,700 writers found 90% believe authors should be compensated when their books train generative models, and 65% support a collective licensing system — figures the Guild continues to cite as its baseline. On the employment side, Media Copilot’s tracker counted more than 2,300 US and UK newsroom job cuts in just the first half of 2026, already roughly 34% ahead of all of 2025. The tracker itself calls the causal link to AI murky rather than measured: AI is blamed both for collapsing referral traffic, as readers turn to AI chat and search instead of visiting sites, and for production tools that let remaining staff do more with fewer people — but both effects are entangled with a broader collapse in digital advertising revenue. Report this as AI-linked, not AI-caused; it is one more data point for the occupational-exposure picture in labor and economic displacement, not a standalone causal finding.
The publisher-licensing map
OpenAI has assembled the largest visible portfolio: roughly 20 verified news-publisher partnerships covering more than 160 outlets in 20-plus languages as of June 2026, beginning with the Associated Press in 2023 and expanding through News Corp, Axel Springer, Vox Media, Condé Nast, Axios, and The Guardian, among others. Widely reported dollar figures for some of these deals — such as News Corp’s reported up-to-$250 million, five-year agreement — trace to secondary reporting that could not be independently reverified against a primary outlet in this research and should be treated as reported, not confirmed.
Other companies have built parallel arrangements rather than joining OpenAI’s. Perplexity’s Comet Plus, launched in August 2025, is a $5/month subscription seeded with a $42.5 million pool sharing 80% of revenue with publisher partners. Microsoft launched a “Publisher Content Marketplace” in February 2026 letting publishers set their own licensing terms; Google, Mistral, and Meta each struck separate deals through 2025 and early 2026. A parallel layer of collective licensing is also forming — the Copyright Clearance Center has been building an AI-training collective license for publishers — though Brookings has argued these marketplaces mostly recreate existing gatekeeping rather than reaching small or independent creators. Notably, Anthropic has signed no publicly disclosed publisher-licensing deal of this kind — and, distinctively among frontier labs, has also not been sued by any news publisher.
Anthropic’s $1.5 billion settlement, and who actually gets paid
Anthropic’s exposure instead came from authors. Bartz v. Anthropic — the largest copyright class-action settlement on record — received final court approval on July 20, 2026, covering roughly 500,000 books at an estimated $3,000 per work, and requiring Anthropic to destroy pirated files originally sourced from Library Genesis and Pirate Library Mirror. Settlement, though, is not the same as payment. By September 2026, authors were disputing publisher and literary-agency claims filed against the payout pool: author April Henry found HarperCollins had claimed payment for a book whose rights had reverted to her roughly seventeen years earlier, and had listed HarperCollins as her “employer” without authorization. Writer Beware documented similar cases of publishers claiming works they no longer held rights to. Authors Guild CEO Mary Rasenberger attributed this to record-keeping confusion rather than deliberate misconduct — a real, unresolved dispute over a settlement often described as already closed.
Labeling AI output: provenance meets binding law
Compensation and disclosure are separate problems, and the disclosure layer is maturing faster. OpenAI’s own provenance policy describes it as a C2PA steering-committee member and states that, as of July 31, 2026, it extended Google DeepMind’s SynthID watermarking to its audio outputs and launched a public verification tool. This manual’s deepfakes and shared-facts page covers the mechanics and limits of C2PA and SynthID in more depth; the point relevant here is that adoption by a major model provider gives creators a technical, if imperfect, way to assert authorship. The legal layer is now binding, not aspirational: the European Commission confirms that AI Act Article 50 transparency duties requiring machine-readable labeling became fully enforceable on August 2, 2026, with fines reaching up to €15 million or 3% of global turnover for relevant violations — one instance of the broader legal and regulatory leverage this manual tracks across AI governance generally.
What this doesn’t resolve
None of this settles whether creators come out ahead. Licensing deals compensate the organizations that hold rights, which is not always the individual who created the work; settlement funds can sit disputed for months after “final approval”; voluntary labeling tags do not require a bad actor to use them; and survey-reported losses coexist with survey-reported income growth in the same professions. What has changed since 2023 is that the fight now runs through courts, contracts, and enforceable law rather than only through public argument — which makes it slower, but for the first time, auditable.