Copyright and Generative AI: Navigating Fair Use, Licensing, and Data Provenance in 2026

It is July 2026, and the legal fog surrounding Generative AI machine learning models that generate new text, images, audio, code, or video based on patterns in training data has not lifted-it has just become more complex. If you are building, buying, or regulating AI tools today, you are likely asking one burning question: Is it legal to train a model on copyrighted material? The answer used to be a simple "maybe." Now, thanks to shifting court rulings, high-profile lawsuits, and new government guidance, the answer is a nuanced "it depends on your data, your license, and your output controls."

The landscape changed significantly with the U.S. Copyright Office’s (USCO) pre-publication report released in May 2025. This document didn't give developers a free pass, nor did it ban AI training entirely. Instead, it laid out a rigorous framework for evaluating fair use, emphasizing that context matters more than ever. For businesses, this means relying on guesswork is no longer a viable strategy. You need a concrete plan covering licensing, data provenance, and risk mitigation.

The Core Problem: Human Authorship vs. Machine Output

Before diving into training data, we have to address the elephant in the room: who owns the output? Under current U.S. law, specifically 17 U.S.C. § 102, copyright protects "original works of authorship" fixed in a tangible medium. The Supreme Court clarified in Feist Publications v. Rural Telephone Service (1991) that this requires a "modicum of creativity."

Here is the hard truth for AI creators: As of 2026, the U.S. Copyright Office maintains that works created by a machine without human authorship are not eligible for copyright protection. This stance was reaffirmed in their March 2023 guidance and reinforced in the 2025 AI Training report. Cases like Thaler v. Perlmutter (2023) confirmed that only humans can be authors. So, if you prompt an AI to write a novel or paint a picture, and you do not exercise significant creative control over the final selection and arrangement, you might not own the copyright to that work. This doesn't mean the output is public domain, but it does mean you cannot exclusively claim it as your intellectual property in the traditional sense.

Fair Use: The Four Factors That Decide Your Fate

Most AI developers rely on the Fair Use Doctrine a legal principle in U.S. copyright law that allows limited use of copyrighted material without requiring permission from the rights holders codified in 17 U.S.C. § 107. The USCO’s 2025 report explicitly states that copying copyrighted works into training datasets constitutes prima facie infringement-meaning it technically checks the boxes for ownership and copying. The burden then shifts to the developer to prove fair use.

Courts evaluate four factors:

1. Purpose and Character: Is the use transformative? Is it commercial or nonprofit educational? Research-oriented, non-commercial training is viewed more favorably, akin to the Authors Guild v. Google (2015) precedent where mass scanning for search snippets was deemed fair use. Commercial training faces stricter scrutiny.
2. Nature of the Work: Creative works (novels, art) get stronger protection than factual works (news reports, databases).
3. Amount Used: Training often requires ingesting entire works. The USCO warns that wholesale copying "ordinarily weighs against fair use" unless justified by a transformative purpose that results in non-expressive outputs.
4. Market Effect: Does the AI output substitute for the original work? If your model generates articles that compete directly with paid news subscriptions, this factor weighs heavily against fair use.

The key takeaway from the 2025 report is that fair use is not automatic. It is highly context-specific. If your model "memorizes" and regurgitates verbatim text or near-identical images from its training data, you are walking into a minefield. Lawsuits like The New York Times Co. v. OpenAI (filed Dec 2023) allege exactly this-that models reproduce paywalled content, harming subscription markets.

Licensing: Buying Peace of Mind

Because fair use outcomes are uncertain, many major players are opting for licensing. This trend accelerated between 2022 and 2024. High-profile deals include:

• Shutterstock and OpenAI (Oct 2022): OpenAI licensed Shutterstock’s image library for training, and Shutterstock launched an AI generator using OpenAI’s tech, sharing revenue with contributors.
• Associated Press and OpenAI (July 2023): A deal allowing OpenAI to license part of AP’s text archive dating back to 1985.
• Axel Springer and OpenAI (Dec 2023): A multi-year partnership allowing training on and displaying snippets from Axel Springer content, with attributed links and revenue-sharing.

These deals illustrate a shift toward monetization rather than litigation. Pricing models vary: some are flat fees (often six-to-eight figures for large archives), while others are usage-based or involve revenue sharing (typically 5-20% of attributable revenue). For smaller creators, platforms like Adobe Stock offer standardized agreements where royalties are paid based on downloads or use events. If you are a startup, negotiating these deals might be tough, but understanding them helps you gauge the market value of data.

Data Provenance: The New Compliance Standard

If licensing is the shield, data provenance is the sword. Data Provenance the ability to trace the origin, history, and custody of data throughout its lifecycle is no longer optional; it is essential for legal defense and auditability. The USCO and various legal experts emphasize that "public URLs still carry copyright," and finding something online does not make it free to use.

To build a robust provenance system, implement these five controls:

1. Source Audits: Before ingestion, verify the copyright status and license terms of every dataset.
2. Written Licenses: Ensure terms of service explicitly allow for AI training. Don’t assume default web scraping rights.
3. Granular Documentation: Maintain logs with hashes, timestamps, and chain-of-custody records for every dataset version.
4. Data Minimization: Keep only what is necessary. Delete non-essential files to reduce exposure.
5. Periodic Reevaluation: Licenses expire, and laws change. Regularly review your data sources.

Technical tools can help here. Use automated detectors to scan generated outputs against known corpora for similarity. Implement human review checkpoints for public-facing outputs. And log all prompts and responses to reconstruct events if a claim arises. This isn't just about avoiding lawsuits; it's about building trust with users who increasingly demand transparency.

Global Perspectives: EU, UK, and Japan

The rules aren't uniform worldwide. If you operate globally, you face a patchwork of regulations:

In the EU, the Digital Single Market (DSM) Directive allows text-and-data-mining (TDM) for scientific research and, crucially, for commercial entities unless rights holders opt out via machine-readable signals like robots.txt. The EU AI Act, adopted in 2024, adds a layer of transparency, requiring providers of general-purpose AI models to publish detailed summaries of their training data. This makes unlicensed scraping harder to hide. In contrast, Japan offers a very broad exception for data analysis, effectively permitting most AI training scenarios. The UK remains cautious, limiting exceptions to non-commercial research, which pushes commercial developers toward licensing.

Practical Steps for 2026

So, what should you do right now? First, map your data sources. Classify them by risk: internal data (low risk), licensed data (medium risk), and public web scrapes (high risk). Second, implement governance policies. Restrict training on subscription databases whose terms prohibit scraping. Third, train your teams. Legal and engineering staff need joint workshops to understand how copyright basics intersect with model architecture. Finally, build review loops. Treat AI drafts as raw material that requires editorial sign-off before publication.

The legal debate is far from over. Ongoing cases like Andersen v. Stability AI and Doe v. GitHub will shape future precedents. But waiting for clarity is a luxury few can afford. By prioritizing licensing, provenance, and transparent practices, you position your organization not just to survive the current wave of litigation, but to thrive in the next era of responsible AI innovation.