---
The technology of Generative AI has outpaced the law of copyright by at least a decade. In the absence of legislative clarity, the rules of the road are currently being written in courtrooms.
For enterprises, the risk is twofold:
1. Input Risk: Using models trained on copyrighted data without license (potential infringement).
2. Output Risk: Creating code or content that inadvertently reproduces protected works (potential liability).
The Core Conflict: Training Data
The fundamental legal question is whether scraping the internet to train a model constitutes "Fair Use" (in the US) or "Text and Data Mining" (in the EU/UK).
The "Fair Use" Defense (US)
AI labs argue that training is transformative. They claim the model is not "copying" books, but "learning" the patterns of language from them, similar to a human student reading a library.
Case to Watch: New York Times vs. OpenAI*. The NYT alleges that ChatGPT can recite its articles verbatim, undermining the "transformative" argument. Case to Watch: Andersen vs. Stability AI*. Artists argue that image generators are merely high-tech collage tools violating their rights.The "Opt-Out" Reality
While courts deliberate, the market is moving toward an opt-out model.
- robots.txt: Major AI scrapers (GPTBot, CCBot) now respect `disallow` directives, though this is a voluntary standard, not a law.
- Do Not Train: New standards like C2PA are attempting to embed "do not train" metadata directly into files.
Output Rights: Who Owns the Prompt Result?
If you use Midjourney to create a logo, do you own it?
US Position: No. The US Copyright Office has repeatedly ruled (e.g., Thaler, Kashtanova*) that works created without "human authorship" cannot be copyrighted. You cannot trademark a purely AI-generated logo.- EU Position: More ambiguous, but generally leans toward requiring human creative input.
Key Lawsuits Tracking Table
| Case | Plaintiff | Defendant | Core Allegation | Status |
|---|---|---|---|---|
| NYT v. OpenAI | New York Times | OpenAI / Microsoft | Large-scale copyright infringement & "recitation". | Active |
| Getty v. Stability | Getty Images | Stability AI | Training on watermarked images (Trademark). | Active (UK/US) |
| Doe v. GitHub | Developers | GitHub / Microsoft | Copilot reproduces open-source code without attribution. | Active |
| Authors Guild v. OpenAI | GRR Martin, etc. | OpenAI | "Systematic theft on a mass scale." | Active |
Enterprise Action Plan: Managing IP Risk
Until the Supreme Court rules (likely 2026+), uncertainty is the only certainty.
1. Indemnification is King: Only use enterprise AI models (ChatGPT Enterprise, Azure, Copilot) that offer IP Indemnification. This shifts the legal risk from your balance sheet to Microsoft/Google's.
2. Human-in-the-Loop: To ensure copyrightability of your output, maintain a "chain of custody" showing significant human editing of AI drafts.
3. Clean Training: If building internal models, consider "clean" datasets (e.g., Adobe Firefly, Shutterstock) that certify full licensure of training data.
4. Audit GitHub Copilot: Ensure filters are enabled to block code suggestions that match public code, avoiding "viral" open-source license infection (GPL) in your proprietary codebase.
For liability questions beyond copyright, see Liability Frameworks. To understand the open source angle, see Open Source vs. Regulation.