The internet was built on a simple, unspoken bargain: creators publish, audiences consume, and everyone benefits from the free flow of information. That bargain is now breaking apart. And the fight over what comes next will define the economics of artificial intelligence for a generation.
At the center of the dispute is a deceptively simple question — can AI companies scrape the open web to train their models without paying the people who created the content? The answer, depending on whom you ask, is either an obvious yes rooted in fair use doctrine or a brazen act of industrial-scale copyright infringement. Courts in multiple jurisdictions are now being asked to settle the matter. The outcomes will reverberate far beyond the technology sector.
BBC News reported that a growing coalition of creators, publishers, and rights holders is pushing back against what they describe as the wholesale appropriation of their work. The frustration isn’t abstract. Writers, journalists, musicians, visual artists, and software developers have watched their output get ingested by large language models — systems that can then generate text, images, and code that competes directly with the humans whose labor made the technology possible in the first place.
It’s a collision that was years in the making.
The large language models powering tools like OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude were trained on massive datasets assembled from books, news articles, academic papers, social media posts, forum discussions, and open-source code repositories. Common Crawl, a nonprofit that archives petabytes of web data, has served as a foundational resource for many of these efforts. But Common Crawl doesn’t own the content it indexes. Neither do the AI companies that build on top of it. The original creators do — at least in theory.
In practice, enforcement has been nearly impossible. The scale of ingestion is staggering. OpenAI’s GPT-4 was reportedly trained on hundreds of billions of tokens drawn from across the internet. No individual author, no single newsroom, no lone photographer could have anticipated that their work would be fed into a statistical model designed to predict the next word in a sequence — and that the resulting product would be valued at tens of billions of dollars.
The legal challenges are now stacking up. The New York Times filed suit against OpenAI and Microsoft in December 2023, alleging that millions of its articles were used without permission to train AI systems. The lawsuit, one of the most closely watched intellectual property cases in years, argues that OpenAI’s models can reproduce Times content nearly verbatim — a claim the company disputes. OpenAI has countered that its use of publicly available material falls under fair use, a defense that permits limited use of copyrighted work for purposes like commentary, criticism, and research.
Fair use. Two words carrying extraordinary weight right now.
The doctrine has always been fact-specific and unpredictable. Courts weigh four factors: the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on the market for the original. AI training arguably touches all four in complicated ways. The use is commercial. The works are creative. The amount used is essentially everything. And the market effect is hotly contested — does an AI summary of a news article substitute for reading the original, or does it drive traffic back to the publisher?
There’s no consensus yet. But the legal machinery is grinding forward. Authors including John Grisham, Jodi Picoult, and George R.R. Martin have joined class-action complaints. Getty Images sued Stability AI in both U.S. and U.K. courts over the use of its photographs to train image generators. Music labels have raised alarms about AI-generated songs that mimic the voices and styles of specific artists. The Recording Industry Association of America has warned that without clear protections, the creative industries face an existential threat.
Some AI companies have tried to get ahead of the backlash. OpenAI has struck licensing deals with publishers including the Associated Press, Axel Springer, and Le Monde. Google has pursued similar agreements. These deals typically involve payments in exchange for the right to use archived and ongoing content for training purposes. But critics argue the sums involved are a fraction of the value being extracted — and that the deals create a two-tier system where large publishers can negotiate from a position of strength while independent creators get nothing.
That disparity matters. A freelance journalist whose work appeared in dozens of online publications has no practical way to determine whether her articles were included in a training dataset, let alone seek compensation. A photographer who uploaded images to a stock platform years ago may find that AI-generated imitations of his style now flood the market. A novelist whose backlist was digitized and scraped has no mechanism for opting out after the fact. The asymmetry of information and power is enormous.
And the technology keeps advancing. Each new generation of models requires more data — or at least higher-quality data — to improve performance. Synthetic data, generated by AI systems themselves, has been proposed as a partial solution, but researchers have found that training models on their own output can lead to a phenomenon called model collapse, where quality degrades over successive generations. Human-created content remains the essential fuel.
This creates a paradox. The AI industry needs creators to keep producing original work. But if creators can’t sustain themselves economically because AI-generated alternatives are flooding the market and undercutting prices, the supply of high-quality training data will eventually shrink. It’s a feedback loop that could undermine the very foundation these models depend on.
Some voices in the technology sector reject this framing entirely. They argue that AI training is analogous to a human reading books in a library — learning from existing knowledge and synthesizing it into something new. Under this view, requiring licenses for training data would be like requiring authors to pay royalties to every writer who influenced them. The analogy has intuitive appeal but breaks down under scrutiny. A human reader doesn’t memorize and reproduce entire passages on demand. A large language model, under certain conditions, demonstrably can.
The international picture adds further complexity. The European Union’s AI Act and its copyright directive take a different approach than U.S. law. The EU permits text and data mining for research purposes but allows rights holders to opt out of commercial mining — a framework that places the burden on creators to actively protect their work rather than requiring AI companies to seek permission upfront. Japan has taken an even more permissive stance, broadly allowing the use of copyrighted material for AI training. The U.K. considered a similar approach but pulled back after fierce opposition from the creative sector, as BBC News detailed in its coverage of the ongoing policy debate.
The patchwork of global rules creates opportunities for regulatory arbitrage. An AI company could theoretically train its models in a jurisdiction with looser rules and deploy them worldwide. Enforcement across borders is notoriously difficult. And the pace of AI development far outstrips the speed of legislative action.
So where does this leave the creators?
Organizations like the Authors Guild, the News Media Alliance, and various collecting societies have been lobbying for legislative solutions. Proposals range from mandatory licensing schemes — similar to those that exist in the music industry — to transparency requirements that would force AI companies to disclose which copyrighted works were used in training. The U.S. Copyright Office launched a formal inquiry into the issue in 2023 and has received thousands of public comments, but concrete policy action remains elusive.
Meanwhile, some creators are taking matters into their own hands. Websites can deploy robots.txt files to block AI crawlers, though compliance is voluntary and not all companies honor these signals. Technical tools like Glaze, developed by researchers at the University of Chicago, allow visual artists to add imperceptible perturbations to their images that disrupt AI training. But these are defensive measures, not systemic solutions.
The financial stakes are immense. The global AI market is projected to exceed $1 trillion within the next several years, according to estimates from multiple research firms. A significant portion of that value derives from the ability to process and generate human-like content. If courts or legislators impose substantial licensing costs, the economics of foundation model development could shift dramatically. Smaller players might be priced out. The largest companies — those with the deepest pockets and the most extensive existing datasets — would consolidate their advantage.
That concentration of power worries antitrust scholars as much as it worries creators. If only a handful of companies can afford to train frontier models, the competitive dynamics of the AI industry narrow considerably. The irony is that the open web — a resource built on principles of democratized access — could end up fueling a market dominated by a few corporate giants.
There’s a deeper philosophical tension here, too. Large language models don’t understand the content they process. They identify statistical patterns in text and generate plausible continuations. Yet the outputs are often indistinguishable from human-written material. This creates a category problem that existing intellectual property law wasn’t designed to address. Copyright protects expression, not facts or ideas. But when a machine can replicate the style, tone, and substance of a specific writer’s expression, the line between protected and unprotected becomes blurry.
Courts will ultimately draw that line. The New York Times case, the Getty Images litigation, and the various class actions working through federal courts will produce precedents that shape the industry for decades. But litigation is slow, expensive, and uncertain. Many creators can’t afford to wait.
The technology companies know this. And they’re moving fast to establish facts on the ground — training models, shipping products, building user bases — before the legal and regulatory frameworks catch up. It’s a familiar playbook from Silicon Valley. Move first, negotiate later.
But the creators pushing back this time aren’t a fragmented group of individuals shouting into the void. They include some of the most powerful media organizations on the planet, well-funded trade associations, and increasingly sympathetic policymakers on both sides of the Atlantic. The political dynamics are shifting. Public opinion surveys consistently show that majorities support compensating creators whose work is used to train AI.
None of this will be resolved quickly. The cases will take years to work through the courts. Legislation will be debated, watered down, stalled, and possibly revived. International coordination will prove elusive. And the technology will continue to evolve in ways that make today’s disputes look quaint.
But the core question won’t go away. The words, images, and sounds that train the most powerful AI systems on Earth were created by human beings. Those human beings want to know: what’s their work worth? The answer will shape not just the future of artificial intelligence, but the future of creative labor itself.
WebProNews is an iEntry Publication