The impact of AI art on artists looks different depending on which evidence you read. Surveys of working artists report lost work, lower income and fewer income opportunities. Labour market data for the United States show no broad drop in artists’ earnings so far. Artists report real harm, but it has not yet shown up in aggregate US pay. Generative AI is changing how many artists work more clearly than it is replacing artists as a group.
The impact of AI art on artists: two kinds of evidence, two different answers
Artists’ own accounts are stark. In a survey of 378 verified professional visual artists, 80% said they compete with generative AI, 90% said it had diminished their income opportunities and 75% said it had hurt job security or client stability. When the Society of Authors surveyed UK creators in January 2024 (787 responses), 26% of illustrators said they had already lost work to generative AI and 37% said the income from their work had decreased in value.
The labour statistics tell a calmer story. A study in the Journal of Cultural Economics, reported by Gallup, used Bureau of Labor Statistics data from 2017 to 2024 and found that earnings trends for artistic occupations more exposed to generative AI look broadly similar to those for less exposed ones, with slightly positive estimates that are not statistically distinguishable from zero. Census Bureau survey data showed a modest earnings rise for more exposed artists in 2023 that faded somewhat in 2024, while total hours worked rose from 2022 and stayed elevated.
The two kinds of evidence measure different things.
What each kind of evidence can and cannot show
| Evidence | What it measures | What it can show | What it cannot show |
|---|---|---|---|
| Artist surveys (378 visual artists; SoA UK, 787 respondents) | Self-reported experience and expectation | Where artists feel pressure, lost clients or devalued rates | Whether the losses are representative of all artists, or caused by AI rather than other market changes |
| National labour data (BLS 2017-2024, Census survey) | Wages, employment and hours by occupation | That aggregate earnings in exposed occupations have not fallen sharply | Losses concentrated among freelancers, early-career artists or narrow niches that average out inside a broad occupation code |
| Regional reporting (China’s game industry, 2023) | Specific studios and freelancers | What displacement looks like where adoption is fast | How typical that market is elsewhere |
One detail matters for interpreting the reassuring result: the exposure index in the Gallup-reported study scores occupations against the capabilities of large language models. Our reading is that an index built that way may understate exposure for image-heavy work such as illustration, where the displacing tools are image generators rather than text models. That is a reason to treat “no broad earnings drop yet” as a finding about aggregates and about now, not a forecast.
Is AI replacing artists, or changing how they work?
In the US data, AI is not yet reducing artists’ employment, but it is changing how they work. The research also found that even the most AI-exposed creators were working slightly more hours in 2024 than in 2022, which does not suggest a shrinking workload. Roughly one in four occupation-defined artists say they use AI frequently, against about one in five workers across the wider economy. They are more likely than other workers to use it for idea generation and creative exploration, and also to automate small tasks, consolidate information and support collaboration.
Some artists describe genuine creative gains: Cosmo Wenman, working with blind researchers on Midjourney images, called the technology “essentially moveable type for visual rhetoric”. Where these tools fit in real production, and where they do not, is covered in our piece on AI art use cases and creative workflows.
Replacement does happen, though, as reporting from China’s game industry shows. Rest of World reported that a Chinese game studio laid off five of its 15 character-design illustrators; Xu, at the studio, believed AI image generators were partly to blame, saying “two people could potentially do the work that used to be done by 10”. A freelance illustrator there who had earned 3,000 to 7,000 yuan per game poster said she was now offered small fixes, such as correcting lighting and skewed body parts, for a tenth of her original rate. A Hangzhou recruiter put the fall in illustrator jobs at about 70% over a year, attributing it to regulatory pressure and a slowing economy as well as AI.
Which artist roles are most exposed to generative AI?
Entry-level positions and the conceptual phase of projects are singled out as particularly at risk. The survey literature identifies entry-level positions in film, TV and gaming as particularly at risk, with the concept phase a popular point at which to substitute AI, and notes that traditional niches such as book-cover illustration can now be filled by generated output. The film side is covered in our article on AI in cinematography.
Occupation-level exposure scores range from structured creative production to embodied performance:
- Music directors and composers score about 0.70, because much of their work is composition and arrangement that AI tools can help draft or modify.
- Special effects artists and animators score around 0.54; art directors, producers and directors cluster near 0.50.
- Craft artists and choreographers sit around 0.27 to 0.28, actors around 0.18, and dancers near 0.04, because live presence and physical skill are hard to substitute.
Exposure is not the same as displacement. A high score says a large share of tasks overlaps with what the tools can do; whether that turns into fewer jobs or faster drafts depends on how employers use the time saved.
What are the negative effects of AI art on working artists?
The reported harms go beyond lost income. Among the 378 professional visual artists surveyed, 61% reported disruption to their employment or business, 57% said individual negotiating power had fallen, and responses on career growth were 77% negative against 1% positive. Some artists described being fired, or planning to leave the field entirely.
Pressure to adopt is a reported harm of its own. Artists in the same study reported bosses telling them that “employees who want a future in this company need to embrace AI”, and of the third who saw a change in turnaround times, almost all said it had become shorter. In China’s game industry, one artist said AI had made her team “more productive but also more exhausted”, as people who once drew one scene or character a day were now producing 40 for managers to choose from.
Then there is suspicion. Artists report being asked to prove their work is not AI-generated, and some now record their creative process as evidence. Style mimicry is a widespread worry: 86% of the Society of Authors’ respondents said they were concerned about their style, voice and likeness being reproduced in AI output. The authors of the 378-artist study also warn that the harm may fall disproportionately on marginalised groups, since freelance art can be a more accessible career for disabled and LGBTQ+ people.
Why the AI art controversy runs so deep among artists
The anger centres on consent and who profits. Generative image models are trained on artists’ work scraped without explicit consent, and their outputs then compete with those artists in the marketplace.
Artists are not of one mind about this. In a survey of 459 artists, 61.87% agreed that AI models are a threat to art workers and 44.88% agreed they are a positive development, with 22.03% holding both views at once. About 48% of respondents had used AI models themselves. What they largely agree on is transparency: 80.17% said model creators should be required to disclose in detail what art they train on. On payment, 50.97% said they did not need compensation but cared who profited from their work, and the most chosen option was “I don’t need to profit, but I don’t want for-profit companies to profit from my art”.
One public flashpoint was an open letter, signed by nearly 4,000 people, asking Christie’s New York to cancel its first sale dedicated solely to AI-created art over models trained on copyrighted work. Buyers care too: Rest of World found game players who said they would pay less for AI-made skins, and a studio that pledged against AI artwork after players called out a poster.
Can AI companies legally train models on artists’ work?
In the United States the early rulings do not point in one direction. In Bartz v. Anthropic the court held that training on books was transformative and fair use, but that downloading and retaining pirated copies was infringing. That piracy holding underlies Anthropic’s $1.5 billion settlement in September 2025, roughly $3,000 per work. In Kadrey v. Meta, the court ruled for Meta because the plaintiffs failed to prove market dilution, while stressing that a stronger record could change the outcome.
Both of those cases concerned books. For visual artists, the case is Andersen v. Stability AI, a class action by artists including Sarah Andersen, Karla Ortiz and Greg Rutkowski over Stable Diffusion’s training on their work. In an August 2024 order the court let the direct infringement claims proceed, finding it plausible that Stable Diffusion was built to a significant extent on copyrighted works and created to facilitate infringement by design, while dismissing the DMCA claims with prejudice. Whether fair use applies has to be tested on the evidence, which is a later stage than that order.
The US Copyright Office frames the stakes directly: some warn that licensing requirements would throttle the technology, while others fear unlicensed training will use artists’ entire bodies of work to produce content that competes with them. The Office’s pre-publication analysis treats training as less transformative when a model is built to produce content that shares the purpose of appealing to a particular audience. Our reading is that this is the argument most relevant to image generators trained to output illustration, and that it will turn on market-harm evidence of the kind the artist surveys describe.
For teams building generative AI features, the book rulings already make clean acquisition of source material a central legal risk variable, so provenance belongs on the checklist alongside model quality when moving a generative AI prototype into production.
Can you copyright art made with AI?
Only the human part. The Copyright Office’s 2025 report on copyrightability concludes that generative AI outputs are protected only where a human author has determined sufficient expressive elements, which can include perceptible human-authored work or creative arrangement and modification, but not the mere provision of prompts. Using AI to assist, or including AI material in a larger human work, does not bar protection. “Purely AI-generated material” remains ineligible.
AI creativity vs human creativity: what the tests actually show
On the creativity tests studied, AI beats the average person and loses to the best. A study published in Scientific Reports in January 2026 compared more than 100,000 people with today’s most advanced AI systems and found that some, including GPT-4, exceeded average human scores on divergent linguistic creativity, measured with the Divergent Association Task. The most creative half of participants still outscored every AI model tested, and the gap widened among the top 10%, especially on richer work such as poetry and storytelling.
An earlier study by Koivisto and Grassini (Scientific Reports, 2023) found the same shape using the Alternate Uses Task with 256 participants: chatbots outperformed humans on average, yet the best human ideas matched or exceeded theirs. The authors identify humans’ clearest weakness as poor-quality ideas, which were absent from the chatbots’ responses.
Two limits apply before anyone extends these results to art. The tests measure divergent thinking, one facet of a multifaceted ability. And they are verbal tasks, listing words or uses, so they say nothing direct about composition, draughtsmanship or visual judgement. The fair summary is that AI raises the floor of idea generation while the ceiling stays human, which is consistent with the survey literature’s finding that entry-level positions are particularly at risk.
For working artists, a model’s score on a word test matters less than how courts weigh market harm from training without consent. The answer will come from cases like Andersen rather than from benchmarks.
