Beyond Ansel Adams: How AI Tools Exploit Artists Without Consent

When photographer Ansel Adams' iconic image of Half Dome was displayed in a gallery, it was theft. When that same image—along with millions of others—becomes training data for an AI generator without the photographer's knowledge or permission, most tech companies call it progress. The distinction matters less than you'd think, and it reveals a seismic shift in how creative work is appropriated at scale: instead of exploiting individual artists, generative AI companies are systematizing the exploitation of entire creative classes. The lawsuits are piling up, the stakes are in the billions, and the legal frameworks designed to protect creators are struggling to keep pace with technology that moves faster than courtrooms.

The training data behind today's most powerful generative AI image tools reads like a greatest hits collection of internet culture, minus one crucial element: permission. DALL-E, Midjourney, and Stable Diffusion were all built by scraping millions of copyrighted images from across the web—pulling from art portfolio sites, Instagram, ArtStation, and datasets like LAION-5B, which alone contains approximately 5.85 billion image-text pairs harvested without artist consent. These aren't small or obscure datasets; they're foundational infrastructure built on what amounts to industrial-scale copyright infringement, normalized and executed by well-funded technology companies operating in a legal gray zone.

Stable Diffusion's training data came largely from the LAION dataset, which scraped images indiscriminately from the internet. Wired's investigation into AI data laundering exposed how these datasets perpetuate a cycle where artists' work circulates through multiple systems without compensation, each time building more value for tech companies. Midjourney doesn't publicly disclose its exact training sources, but the company has acknowledged using internet images. DALL-E, OpenAI's flagship tool, was trained on data including images from the web and licensed datasets, though the specifics remain opaque—a pattern across the industry where transparency about training data is treated as proprietary information.

What makes this different from past copyright disputes isn't just the scale; it's the systematic denial of agency. When a photographer's work is stolen and resold, there's a clear victim and perpetrator. When millions of artists' works are used to train a system that then generates new images, the liability becomes abstract. The companies argue they're engaging in fair use—that training AI models is transformative enough to fall under the doctrine. Artists counter that the transformation isn't theirs to authorize, and that fair use has limits when applied to commercial products generating billions in value.

The Legal Battlefield: Lawsuits, Uncertainty, and Collective Action

The legal reckoning is underway, though far from resolved. In 2023, a group of artists filed a class-action lawsuit against Midjourney, Stability AI, and DeviantArt, alleging copyright infringement and unjust enrichment. Similar suits followed against other AI companies. These aren't boutique cases—they represent organized collective action by artists who've watched their styles replicated, their techniques extracted, and their economic value redistributed to AI shareholders.

The New York Times sued OpenAI and Microsoft in late 2023, claiming their models were trained on millions of Times articles without permission or compensation. That case matters because it's not about individual artists fighting a collective battle; it's a major media organization with resources to litigate, establishing that even acknowledged copyright holders see AI training as infringement. Yet the outcome remains uncertain. The doctrine of fair use is deliberately broad, designed to accommodate innovation, and courts have historically been reluctant to restrict technology's development based on copyright concerns. When Google scanned millions of books without publisher permission to create Google Books, courts ruled it was fair use. The precedent cuts both ways for AI companies.

What distinguishes these new lawsuits from past copyright cases is the scale and the business model. Previous disputes involved specific instances of infringement. AI training involves wholesale appropriation—millions of works used simultaneously to build products that directly compete with the artists whose work trained them. A digital artist who spent years developing a distinctive style can watch an AI generator produce variations of that style in seconds, at a fraction of the cost. That's not just copyright infringement in the traditional sense; it's economically devastating in ways copyright law wasn't designed to address.

Fair Use as a Crumbling Shield

The tech industry's defense rests on fair use, the legal doctrine allowing limited use of copyrighted material without permission. The argument follows logic: AI companies are using copyrighted images for training, which is transformative; they're not copying images to sell them; and they're creating new tools that benefit society. Courts have accepted similar arguments before. Google Books, despite mass digitization of copyrighted works, was deemed fair use because the scanning served a public interest and didn't directly substitute for the original works.

But there are critical differences. Google Books doesn't allow you to search for "books written in the style of Stephen King" and generate new content styled after his work. It's transformative in a specific, limited way. Generative AI doesn't just transform copyrighted works; it replicates their essential qualities—style, technique, aesthetic choices—and commodifies those replications. An artist's years of work developing a visual signature can be distilled into parameters that produce infinite variations, each one potentially undercutting the original artist's market.

The fair use doctrine has four factors: purpose and character of use, nature of the copyrighted work, amount and substantiality of the portion used, and effect on the market for the original. Generative AI companies struggle on at least two of those points. The effect on the market is measurably negative for professional artists whose work is in the training data—they're being undercut by cheaper AI alternatives. And while the companies argue their use is transformative, the substantiality question remains: they're using entire copyrighted works, in their entirety, to extract artistic value. It's not sampling a passage; it's extracting the DNA of millions of artworks.

The Electronic Frontier Foundation has examined fair use in AI contexts, acknowledging both the legitimate innovation concerns and the real harms to creators. But EFF's balanced approach—which tends to defend broad fair use interpretations—may not reflect how courts ultimately rule when confronted with the scale of AI training and its direct economic impact on the artists whose work was used.

Building Alternatives: Technical and Legal Solutions Still in Flux

If lawsuits alone won't solve the problem—and legal uncertainty suggests they won't, at least not quickly—what would? The industry has proposed and begun implementing several alternatives, though none have achieved consensus or become standard practice.

Opt-out registries represent one approach: artists could register their work, and AI companies would exclude it from training data. Stability AI created an opt-out tool for artists, but participation is voluntary and the company still retains already-trained models. It's a gesture toward consent that doesn't actually prevent past infringement. More fundamentally, opt-out puts the burden on millions of individual artists rather than on companies with resources to implement opt-in systems.

Licensing agreements offer another path. Some AI companies are beginning to negotiate with rights holders, paying for access to training data. Getty Images signed a deal with Nvidia to license images for AI training. These arrangements could become industry-standard, but they risk creating a tiered system where only established, well-known artists get compensated while the vast majority of working creatives are left out.

Legislative solutions are emerging globally but unevenly. The European Union's AI Act imposes transparency requirements for AI training data, though enforcement mechanisms remain weak. The U.S. has proposed bills addressing AI copyright issues, but none have passed, and tech industry lobbying remains formidable. What's missing is an international framework that would establish clear liability for AI companies, mandate licensing, and create compensation mechanisms for artists whose work was used without permission.

The current moment feels transitional. The technology has outpaced the law, companies have built billion-dollar businesses on uncertain legal ground, and artists are organizing to demand recognition of their contributions. The outcome will reshape not just AI development but the relationship between technology companies and creative workers for decades to come.