Generative AI Is A Crisis For Copyright Law

Ben Zhao and his team of researchers at the University of Chicago. Front row, Heather Zheng, Anna ..

. More Ha, Cathy Li, Wenxin Ding, Zhuolin Yang. Back row: Shawn Shan, Stanley Wu, Ben Zhao, Kritika Prakash.



This image has been run through Nightshade On March 25, 2025, Open AI launched a controversial image-generation feature, GPT-4o, which allows users to transform various visuals—ranging from personal selfies to iconic movie scenes—into images resembling the distinctive style of Studio Ghibli, the renowned Japanese animation studio founded in 1985 by Hayao Miyazaki, Isao Takahata, and Toshio Suzuki. Fans of have criticized this as disrespectful to Miyazaki's legacy. Studio Ghibli co-founder Miyazaki has been a vocal critic of AI-generated art, describing it as "an insult to life itself.

" He has emphasized that AI lacks the ability to understand human emotions or pain, which are central to authentic artistic expression. Miyazaki views AI art as soulless and anti-human, undermining the emotional depth and storytelling that define Ghibli's work. This event compounds the hotly debated topic of generative AI’s impact on copyright law.

The ambiguity surrounding AI-generated works' authorship and ownership has led to numerous lawsuits against AI companies. Artists have been vocal, and fear being overshadowed by AI tools that can replicate their styles quickly and cheaply, threatening their livelihoods. Recent developments in the White House’s Executive Order prioritizing global AI dominance have resurrected the debate on big tech’s “unfettered” access to the world’s data, including data protected by copyright originating from the publishing and artistic community, more recently Hollywood artists.

And while these lawsuits have yet to produce resolution it has also propelled solutions like Nightshade and Glaze to proactively protect the property of artists. When Open AI launched in 2023, it fundamentally turned the internet into its own “ plumbing system, ” transforming an environment that was designed for exploration and discovery into one that encourages direct engagement through predefined AI interfaces. As of late 2024, there were at least 30 major copyright lawsuits against generative AI companies related to image copyright infringement .

In Getty Images v. Stability AI : Getty Images sued Stability AI in both the US and UK, for $1.7Billion alleging misuse of over 12 million copyrighted photos to train AI models.

In Andersen v. Stability AI, Midjourney, DeviantArt, and Runway AI, a group of visual artists sued these companies for allegedly using their artworks without permission to train AI image generators. And in April 2024, Zhang v.

Google LLC, visual artists sued Google, claiming their copyrighted images were used to train the Imagen model without consent. More recently, in a landmark ruling, Thomson Reuters won a significant victory in its copyright infringement lawsuit against Ross Intelligence, a legal AI startup was found to have infringed Thomson Reuters' copyright by using their proprietary headnotes and structure to train its AI-powered legal research tool. Ross Intelligence eventually shuttered its doors because of legal costs.

This was precedence-setting and the first to address the issue of fair use in the context of AI. In a recent call for submission by the White House Office of Science and Technology Policy, Google and Open AI submitted proposals for extensive policy changes that justify national security interests in securing competitive advantage over rivals like China, arguing “that limiting AI training on copyrighted materials could weaken America’s technological edge and slow down innovation.” Both Google and Open AI argued for “loosening” fair use protections, on all data, including copyrighted material.

Big Tech has argued, as in the case of the NYT lawsuit that “in order to train their technologies, AI companies should be allowed to use works under copyright protection without consent.” This was met with swift response from Hollywood artists in a 400-signature open letter to the White House, which stated, ..

. this issue goes well beyond the entertainment industry, as the right to train AI on all copyright-protected content impacts all of America’s knowledge industries. The letter goes on to argue that both Google and Open AI should not be given a government exception to “freely exploit America’s creative and knowledge industries, despite their substantial revenues and available funds.

..” adding, “There is no reason to weaken or eliminate the copyright protections that have helped America flourish.

Not when AI companies can use our copyrighted material by simply doing what the law requires: negotiating appropriate licenses with copyright holders — just as every other industry does.” I happened upon Joseph Gordon Levitt’s substack , where he claimed, “AI Companies want to Legalize Theft” and in it he warned, If these companies get what they want now, then we’ll be living in a future where any valuable work done by any human being will become fair game for a tech company to hoover up into its AI model and monetize, while that human being gets nothing. His post goes on to unveil the crux of big tech’s strategy, rebuking, “If, however, you believe in an economy where capital and control is centralized in the hands of a few giant tech companies, then I guess you should be a fan of this new Orwellian pro-theft policy that Google has named ‘fair learning.

” Levitt is not wrong. The White House’s Executive Order, The AI Action Plan clearly states that this is the first step in “securing and advancing American AI dominance. ” The $500 billion Stargate Initiative is poised to do just that.

I spoke to Ben Zhao, who is a Neubauer Professor of Computer Science at the University of Chicago. He and his team of researchers have developed Glaze and Nightshade to address the growing concerns surrounding copyright infringement and artistic exploitation by generative AI systems. When presented with OpenAI’s submission to the White House Executive Order asserting compliance with fair use under the law, Zhao responded, You know your business is in trouble when your chief legal argument relies on gross anthropomorphization of a flawed technology as justification for why it should be allowed the break the law.

Ben Zhao, Neubauer Professor of Computer Science at the University of Chicago Ben Zhao, who was recently named to the Time100 AI List in 2024 in recognition of his contributions in the field, has focused his research on computer security, particularly protecting artificial intelligence systems and developing tools to counter malicious actors. “I've been a computer science professor for about 21 years,” says Zhao. “During much of this period, AI was primarily associated with tasks like image classification and content recognition—technologies that served beneficial purposes.

AI used to be about models that could recognize things in content." Zhao explains these technologies would highlight the potential tumor in an MRI scan or identify objects on the road for drivers, stating, “These were unambiguously good products that improved our lives." Zhao spent years safeguarding these kinds of tools, ensuring their integrity and reliability.

Since 2022, Zhao has shifted his focus toward addressing the misuse of AI technologies. "As AI’s availability became ubiquitous, it inevitably opened the door for bad actors to exploit it in harmful ways," Zhao explains. His work has included efforts to disrupt companies that train facial recognition models without users’ consent or even their knowledge.

More recently, Zhao has turned his attention to protecting individuals, especially human creators, who have been negatively impacted by generative AI technologies. "We've really focused on protecting human creators who have borne the brunt of attacks using generative AI—whether it's mimicry or other types of exploitation," he says. Zhao draws parallels between the current challenges in AI governance and the early days of the Internet, referencing a seminal paper by MIT’s David D.

Clark, Tussle in Cyberspace . " Clark wrote about how stakeholders with conflicting interests—such as those concerned with copyright, privacy, and security—needed a level playing field to resolve their differences," Zhao explains. He argues this was crucial for the Internet to thrive and to avoid more negative or harmful outcomes.

In contrast, Zhao argues that the AI landscape lacks such mechanisms today, "When we look at copyright law, regulations, and how businesses manage tensions, it's clear that major stakeholders—like big tech companies training AI models—are using tactics that are unprecedented." Zhao points to instances where companies have threatened governments or regions with economic consequences, such as withholding tax revenue, if regulations are imposed. "This is not something we've seen before.

Political lobbying is one thing, but threatening to exit with billions in tax revenue to force legal changes is entirely different," he emphasizes. Zhao reflects, I look at artists, writers, musicians—entire creative industries—that are facing not just disruption but existential crises because of what these tools enable. The harm, Zhao believes, often stems not from the capabilities of AI itself but from how people misuse these technologies.

"These negative impacts are what I’m trying to mitigate." Zhao describes machine learning security as the intersection of computer security and machine learning, encompassing a wide range of issues such as attacks, defenses, privacy concerns, and misuse. Over the past decade, he has dedicated significant time to building tools that protect machine learning models from these threats.

However, Zhao acknowledges the dual nature of these technologies. "These systems are always, as they say, a double-edged sword," he notes. One of the key insights from the machine learning community over the past 11 years is the realization that of the fundamental gap between how humans and AI interpret content, and that gap isn’t closing anytime soon," Zhao says.

This disparity creates vulnerabilities where attackers can design inputs that appear normal to humans but completely disrupt AI models.” While this gap poses risks—such as tricking image classifiers into making incorrect decisions—it also presents opportunities. "We’ve been able to develop tools that preserve the look and feel of content for human users while making it disruptive to AI models in ways that cause unintended consequences for AI models trained on them," Zhao explains.

Zhao highlights a classic example that demonstrated the vulnerabilities of AI image classifiers through adversarial methods. "It was a picture of a panda hanging off a bamboo trunk," Zhao explains. Researchers showed that by altering just a small number of pixels in the image—changes imperceptible to human eyes—the AI model misclassified the panda as a gibbon.

"The image looked untouched to humans, but to the AI, it was something entirely different," he adds. Adversarial methods are phenomena fundamental to image classifiers and has been a focus of research for over a decade. "The community has largely come to agree that these vulnerabilities are inherent to AI systems," Zhao notes.

While techniques can reduce the likelihood of such errors, there is no definitive solution to eliminate them entirely. Zhao emphasizes that AI systems do not truly "understand" content in a human sense. "What AI does is extract numerical patterns from data to define categories—like distinguishing a panda from another animal," he clarifies.

These numerical rules, however, are inherently flawed. "The rules AI creates always contain gaps and imperfections, leaving untrained regions that can be exploited to mislead the model into making incorrect decisions," Zhao adds. AI’s limitations with its reliance on patterns rather than genuine comprehension persists across nearly all machine learning applications.

He explains in cases like gender bias detection, AI might analyze data using statistical correlations that humans would find illogical or irrelevant. "When AI trades on content [processes information], they trade on content for very different things than what humans see," Zhao notes. This disparity creates vulnerabilities.

Zhao concludes, As long as you understand those rules, you can exploit them. You can manipulate and them and get the model to do things that you know they weren't initially designed to do. Zhao explains that while both Glaze and Nightshade exploit the perceptual gap between humans and AI, their goals and methods differ significantly.

"Glaze is specifically designed to disrupt model fine-tuning," Zhao says. This tool addresses a common issue where individuals use diffusion-based image models, such as Stable Diffusion or DALL-E, to mimic an artist’s style without their consent. "People take a model, train it on samples of an artist's work, and essentially say, 'I’m going to wear your style like a skin and control it like a puppet,'" Zhao explains.

This process often happens without the artist’s knowledge, consent, or compensation. Glaze protects artists by allowing them to subtly alter their artwork before sharing it online. "We change 80 to 90 percent of the pixels in an image in ways that are imperceptible to human eyes," Zhao notes.

Even the artists themselves often cannot distinguish the altered version from the original when viewed side by side. However, these changes confuse AI models during training, causing them to misinterpret the style. Zhao provides an example, “If someone tries to mimic a charcoal artist’s dark portraits, the AI might instead produce something resembling a Jackson Pollock or Picasso painting.

If we can achieve that level of disruption, then we’ve succeeded because the person trying to mimic the artist will fail," Zhao emphasizes. Nightshade , on the other hand, takes a broader approach. While Glaze is designed to protect individual artists from style mimicry, Nightshade aims to shift the power balance for all artists by targeting the content of images rather than just their style.

"Nightshade allows creators to alter their images in ways that are invisible to humans but completely distort how AI models perceive them," Zhao explains. "You can't see the difference before and after," Zhao says, emphasizing how imperceptible these changes are to the human eye. However, these alterations fundamentally change how AI systems interpret the image.

He illustrates that an AI model might see all the visual features of a cat in what is actually a photo of a dog. Over time, this "poisoning" of training data can lead AI systems to develop flawed understandings of objects. Zhao likens Nightshade’s approach to adding hot sauce to food: "It’s not particularly spicy for you, but it’s spicy enough to cause trouble for anyone who steals your lunch.

" By increasing the cost and complexity of training AI models on stolen content, Nightshade incentivizes companies to properly license content from creators rather than resorting to unauthorized data scraping. The goal is to push back against exploitation and ensure artists are fairly compensated for their work. When asked how he responds to critics who argue that tools like Nightshade and Glaze might harm the development of beneficial AI, Zhao challenges the premise, "I would ask, what beneficial AI?" he says.

He argues that much of the current hype around generative AI is overblown, with real-world applications often falling short of their promises. "Once you look past the marketing hype and ask for tools that can generate text or images, you find that text generators hallucinate randomly and unpredictably, making up facts, while image generators often produce content that infringes on copyrights, " Zhao explains. He questions whether these tools genuinely help society or simply aim to replace human roles for cost efficiency.

"When I ask people to show me revenue-producing applications that truly benefit society, there's often a long pause," Zhao remarks. He notes that the answers typically revolve around replacing customer service representatives with chatbots. "Sure, there's such a thing as efficiency, but even in those cases, we've seen recorded failures where chatbots say things companies don't want them to say," he adds.

Zhao cites real-world examples of chatbot errors, such as Air Canada being forced to honor a bereavement discount due to a chatbot hallucination. "Now imagine this happening in safety-critical scenarios—like legal translations or medical transcriptions," he warns. Recent research has shown that AI transcription systems can hallucinate during doctor-patient conversations, raising serious concerns about their reliability at scale.

Zhao also criticizes the argument that these issues can be easily fixed, contesting "People have been saying for that for years but I understand the architecture behind these systems, and they’re not so easily fixable." He believes the flaws are deeply tied to the fundamental design of current AI models. "Unless we break away and do something different, these issues are here to stay," Zhao asserts.

Zhao argues that generative AI models are approaching a plateau in their ability to improve accuracy and reliability, which is creating significant challenges for both adoption and return on investment (ROI). "When you look at these models, the first 80% of getting something to work correctly is relatively easy," Zhao explains. "But getting that last bit—eliminating the remaining errors—is extremely difficult because those errors occur in corner cases, where things interact in unexpected ways.

" Despite significant investments in advancing AI technology, Zhao questions whether meaningful solutions are achievable. "Every company is throwing money at the problem, but there’s a real question of whether we can actually make these models more accurate," he notes. An alternative approach gaining traction is retrieval-augmented generation (RAG), which integrates external databases into the AI workflow.

Zhao explains: With RAG, instead of relying solely on the model, you add facts or external information in real-time. Taken to extreme, it can mean that you basically hollow out the internals of large language models (LLMs), so they are no longer answering you, but they are basically a translator—it takes your query, converts it into a database search, retrieves the results, and translates them back into conversational language. While RAG systems is an approach to address the limitations of LLMs, they fundamentally change the role of these models.

"For decades, we've relied on systems with real, concrete information that we can trust," Zhao says. "This makes the output easier to understand for users," Zhao notes. "At that point, you're not training LLMs to reason or possess knowledge," he says.

"And no one's going to pay millions of dollars for a fancy translator that essentially acts as an intermediary between you and a database. Zhao points out that the hype around generative AI often overshadows its limitations. "The value lies in the actual content and real data inside databases—not in these massive models themselves," he asserts.

This plateau in performance raises serious questions about whether generative AI can deliver on its promises or justify the billions being invested in its development. With increasing industry claims of artificial general intelligence (AGI) on the horizon, Zhao cautions about AI’s current limitations: "These models are extremely good at mimicry, but there is no intelligence, no intention of any kind," Zhao explains. He dismisses concerns about AI "lying" as a form of anthropomorphism, noting that such behavior is simply a fallback mechanism when the model encounters errors.

"There’s no proof that these systems are gaining sentience or intention," he states. Instead, Zhao argues, The control we’re losing is not toward AI—it’s toward money and those with extreme resources. Society is sacrificing creativity, truth in media and trust online for the promise of future economic gains.

Drawing parallels to the early 2000s copyright battles over music sharing and peer-to-peer networks, Zhao notes a significant shift in power dynamics. "Back then, the stakeholders harmed by technology—like recording companies—had armies of lawyers and resources to fight back," he explains. "Today, those with the resources are often the ones breaking copyright laws, leaving creators and smaller groups with far fewer resources to defend themselves.

" Today, OpenAI and Google, with the executive branch of government, have the vast resources to fend off, for the moment, these copyright lawsuits. In the coming months or year, there will be legal resolution on the question of copyright. More broadly, Zhao’s reflections serve as a cautionary reminder: You can’t possibly expect to win the global AI race without sacrificing the sound legal structures, equitable resource distribution and the embedded societal mechanisms created to ensure a free and fair society for all.

At that point, no on wins..