AI and Copyright in 2026: The Evidence on Fair Use, Authorship, and Liability

By mid-2026, the rapid deployment of generative artificial intelligence has collided forcefully with centuries-old copyright frameworks. With the U.S. Supreme Court declining to intervene in a landmark authorship case and multiple federal district courts issuing split decisions on training data, the legal boundaries of artificial intelligence are finally solidifying. This evidence pack evaluates the current legal precedents governing AI and copyright, examining the copyrightability of AI-generated outputs, the legality of ingesting copyrighted works for model training, and the shifting liability landscape for enterprise users. The outcomes of these legal battles will dictate the future of digital creation.[8]

The stakes are existential for both the technology sector and the global creative economy. If training data requires explicit licensing, the economics of frontier model development will fundamentally change, potentially locking out open-source developers. Conversely, if AI outputs cannot be protected and training is deemed universally fair use, businesses face unprecedented risks regarding intellectual property ownership, and human creators face systemic devaluation.[8]

**Claim 1: Purely AI-generated works cannot be copyrighted.** The evidence for this claim is currently definitive and backed by the highest courts. In March 2026, the U.S. Supreme Court denied a petition for certiorari in Thaler v. Perlmutter, letting stand a D.C. Circuit ruling that copyright protection strictly requires human authorship. This decision effectively ends the debate over whether an autonomous machine can be legally recognized as a creator, establishing a firm boundary that intellectual property rights are exclusively reserved for human beings. For businesses relying heavily on automated content generation, this ruling confirms that their unedited outputs reside entirely in the public domain.[2][4][6]

Computer scientist Stephen Thaler had sought copyright protection for a visual artwork titled 'A Recent Entrance to Paradise,' which was generated autonomously by his AI system, DABUS. The courts unanimously rejected the application, affirming the U.S. Copyright Office's long-standing position that the 'traditional elements of authorship' must be executed by a human being, not a machine. The ruling emphasized that copyright law was designed to incentivize human creativity, and extending those protections to algorithms would fundamentally distort the purpose of intellectual property frameworks.[1][2][4]

However, the U.S. Copyright Office has introduced a critical nuance: AI-assisted works can receive protection if the human contribution is sufficiently creative and controls the expressive elements of the final work. Merely entering a detailed text prompt into a generative model does not meet this threshold, but extensive human editing, arrangement, and modification of AI outputs may qualify the human-authored portions for legal protection. This requires companies to meticulously document the human involvement in their AI-assisted workflows.[1][4][6]

**Claim 2: Ingesting copyrighted data for AI training is not categorically protected as Fair Use.** The evidence here is strong, though its application in court remains highly context-dependent. Technology companies have long argued that training large language models is inherently transformative—akin to a human reading books to learn concepts, facts, and linguistic styles. Under this theory, the models do not store exact copies of the ingested works, but rather mathematical weights and parameters, making the training process a non-infringing fair use of the original material.[4][5]

Regulatory bodies are increasingly rejecting this blanket defense. In its May 2025 Part 3 report on generative AI, the U.S. Copyright Office concluded that using copyrighted materials for model development may constitute prima facie infringement. The agency explicitly warned that 'transformative' arguments are not inherently valid when models possess the mechanical ability to output near-perfect copies of their training data, directly competing with the original works.[1][4]

Federal courts are beginning to align with this skeptical view. In the February 2025 ruling for Thomson Reuters v. Ross Intelligence, a Delaware federal court ruled against the automatic classification of AI training as fair use. The judge determined that the transformative nature of the AI model must be weighed against the potential market harm to the original creators, forcing developers to defend their data ingestion practices on a rigorous, case-by-case basis rather than relying on a universal legal shield.[4]

Furthermore, courts have drawn a hard line against the use of illicitly obtained data. In a mixed ruling involving Anthropic in mid-2025, a federal judge determined that while AI companies may legally use copyrighted materials for training if obtained legally, creating a permanent library from pirated sources or shadow libraries is not excused by the fair use doctrine. The ruling implies that AI companies must move toward licensed, permission-based models for their data pipelines.[4][5]

**Claim 3: Proving specific infringement and market harm remains a massive evidentiary hurdle for plaintiffs.** While the theoretical framework favors rights holders, the practical reality of litigation heavily favors AI developers. The evidence for this asymmetry is robust across multiple international jurisdictions. Even when courts acknowledge that unauthorized data scraping occurred, plaintiffs consistently struggle to prove that the ingestion of their specific works directly resulted in financial damage or that the AI model serves as a direct market substitute for their creations.[5][6]

In May 2026, Judge Vince Chhabria denied a motion brought by a group of authors against Meta. The plaintiffs alleged that Meta's Llama models were trained on their copyrighted books. However, the court ruled that the authors failed to present meaningful evidence demonstrating how the training of large language models directly impacted the market for their specific works, halting the momentum of the class-action suit.[5]

A similar dynamic played out in the United Kingdom. In the 2025 case Getty Images v. Stability AI, the court dismissed the secondary copyright infringement claim, highlighting the immense evidential challenges rights holders face in proving that their specific images were used within datasets containing billions of parameters. The court noted that without direct access to the model's underlying architecture, proving specific infringement is nearly impossible.[6]

This evidentiary gap means that while unauthorized training might technically constitute infringement, successfully suing an AI company requires a level of forensic proof regarding the model's black-box training process that most plaintiffs simply cannot access without unprecedented legal discovery. This reality has forced creators to look toward legislative solutions rather than relying solely on the courts.[5][6][8]

**Claim 4: Global regulatory frameworks are pivoting from outright bans to mandatory transparency.** With courts struggling to process infringement claims efficiently, legislative bodies are stepping in to force disclosure. The evidence for this shift is strongest in the European Union, which has abandoned attempts to ban AI training in favor of strict auditing requirements. Regulators recognize that without transparency, copyright holders cannot even begin to enforce their rights, prompting a wave of new compliance mandates aimed at opening the black box of model development.[7][8]

The EU AI Act, which reaches its primary enforcement deadline on August 2, 2026, imposes strict transparency rules on general-purpose AI models. Providers must now publish detailed summaries of the content used for training their models, respecting EU copyright law regardless of where the model was originally trained or headquartered. Failure to comply can result in massive administrative fines or the suspension of the AI system within the European market.[7]

In the United States, similar momentum is building. Bipartisan bills introduced in early 2026 seek to compel AI companies to disclose when copyrighted works are utilized in training data, aiming to give creators the visibility needed to pursue licensing agreements or targeted litigation. While a federal framework has yet to pass, the push for transparency is becoming the dominant regulatory strategy.[4]

**The Enterprise Liability Shift.** The synthesis of these legal developments points to a significant transfer of risk from AI developers to the businesses that use their tools. Because purely AI-generated content lacks copyright protection, any business publishing unedited AI outputs has no legal recourse if a competitor copies their marketing materials, software code, or strategic documents. This creates a dangerous vulnerability for companies attempting to replace human creators entirely with automated systems, as their core assets become legally unprotected.[3]

More critically, as intellectual property lawyers now warn, if an AI-generated output infringes on an existing copyright—either because the model memorized training data or because the user prompted it to mimic a specific creator—the business utilizing the tool can be held directly liable. The legal burden is shifting from the developers who built the models to the enterprise users who deploy them, making corporate AI governance a critical necessity.[3][8]

As 2026 progresses, the legal consensus is clear: the era of unchecked data scraping is ending. AI developers must navigate a tightening web of licensing requirements and transparency mandates, while enterprise users must implement rigorous human-in-the-loop oversight to secure their intellectual property and avoid catastrophic liability. The technology has matured, and the law is finally catching up.[8]