## Sources

1. [Who Owns the Code Claude Wrote?](https://www.oreilly.com/radar/who-owns-the-code-claude-wrote/)

---

### Who Owns the Code Claude Wrote? – O’Reilly by Sena Evren

**Main Arguments**

*   **Human Authorship Requirement:** The core legal baseline for copyright is that it only protects work created by a human [1]. The US Copyright Office and the DC Circuit (upheld in the *Thaler* case) have established that works predominantly generated by AI without meaningful human authorship are ineligible for protection [1].
*   **Ambiguity in AI-Assisted Work:** While pure AI generation is uncopyrightable, the legal status of AI-assisted code—where a human provides prompts or edits—remains a "middle ground" that is actively being litigated [2, 3].
*   **Employer Dominance via Contract:** Regardless of copyrightability, most developers do not own the code they produce at work because employment contracts and the "work-for-hire" doctrine assign ownership of all work product to the employer [4, 5].
*   **The Hidden Risk of Copyleft Contamination:** AI tools trained on open-source data may reproduce substantial verbatim portions of GPL-licensed code [6, 7]. Incorporating this "contaminated" code into a commercial product can trigger legal obligations to release the entire source code, even if the developer was unaware of the original license [7].

**Key Takeaways**

*   **"Meaningful Human Authorship" is the Threshold:** To secure copyright, a developer must prove they made genuine creative decisions, such as choosing the architecture, restructuring output, or rejecting specific AI suggestions [8, 9]. Merely providing a prompt or an objective is generally insufficient [9].
*   **Side Projects are Vulnerable:** Broad IP assignment clauses in employment contracts may allow employers to claim ownership of personal side projects if company-licensed AI tools (like Claude Code or Cursor) were used, or if the project was built on company equipment [5, 10].
*   **Context Awareness Complicates Ownership:** Some employers argue that because AI tools are "context-aware" of a company’s private codebase, any output they generate is a derivative work of company IP, though this specific argument is still legally tenuous [11].
*   **Indemnification Varies by Plan:** Anthropic's commercial and API terms offer broader IP indemnification (defending users against copyright claims) than their consumer/Pro plans [12]. However, this does not protect against downstream GPL violations caused by license contamination [12].

**Important Details**

*   **The Claude Code Leak:** On March 31, 2026, Anthropic accidentally leaked 512,000 lines of Claude Code’s source code [13]. A Python rewrite of this code reached 100,000 GitHub stars in one day, leading to DMCA takedown requests that highlighted the uncertainty of whether Anthropic even owns code that was predominantly AI-written [13, 14].
*   **The *Zarya of the Dawn* Precedent:** This graphic novel case serves as a model for developers; the Copyright Office granted protection for human-authored text but denied it for Midjourney-generated images [15]. Similarly, a developer's architecture docs and prompt logs may be protectable even if the code itself is not [15].
*   **The *Chardet* Dispute:** In early 2026, a developer used Claude to rewrite an LGPL-licensed library and rereleased it under an MIT license, claiming it was a "clean room" implementation [16]. This raised unresolved questions about whether AI-generated code that mimics training data patterns constitutes verbatim copying [16].
*   **Four Recommended Actions for Developers:**
    1.  **Run license scans** using tools like FOSSA, Snyk, or Black Duck to identify hidden open-source code [17].
    2.  **Document creative contributions** by preserving prompt logs, design documents, and detailed commit messages that explain *why* changes were made [18].
    3.  **Review IP clauses** in employment contracts to understand the risks of building independent projects [19, 20].
    4.  **Verify Anthropic plans** to ensure commercial use is covered by appropriate API or enterprise agreements [12].
*   **Settled vs. Unsettled Law:** It is settled that human authorship is required for copyright and that the work-for-hire doctrine is robust [21]. It remains unsettled exactly how much human direction is needed for "meaningful authorship" and whether AI patterns count as verbatim copying of training data [21].