Copyright Protection for AI-Generated Works Under U.S. Law

The intersection of artificial intelligence and copyright law presents one of the most contested doctrinal questions in U.S. intellectual property today. This page examines the current legal framework governing copyright eligibility for AI-generated and AI-assisted works, the U.S. Copyright Office's formal position, and the classification boundaries that determine when human authorship is sufficient to support a valid copyright claim. Understanding these boundaries matters for creators, developers, and rights holders whose workflows involve generative AI tools.

• Definition and Scope
• Core Mechanics or Structure
• Causal Relationships or Drivers
• Classification Boundaries
• Tradeoffs and Tensions
• Common Misconceptions
• Checklist or Steps (Non-Advisory)
• Reference Table or Matrix

Definition and Scope

Under U.S. law, copyright protects "original works of authorship fixed in any tangible medium of expression" (17 U.S.C. § 102(a)). The operative concept in the AI context is authorship — a term the statute does not define, but which courts and the Copyright Office have interpreted to require a human being as the creative source.

The U.S. Copyright Office's Copyright and Artificial Intelligence Policy Statement (Part 1, August 2023) formally declared that copyright "does not protect works produced by machines or mere mechanical processes that operate without any creative input or intervention from a human author." This position extends decades of administrative and judicial precedent, including the Office's longstanding refusal to register works claimed to be authored by non-humans, animals, or autonomous systems.

The scope of this page covers three distinct work categories: (1) purely AI-generated outputs with no human creative input, (2) AI-assisted works where a human makes expressive choices and uses AI as a tool, and (3) works where AI and human contributions are intertwined and difficult to separate. Each category receives different treatment under U.S. copyright doctrine, as explained in the copyright ownership and authorship reference.

Core Mechanics or Structure

The human authorship requirement operates through the Copyright Office's registration and examination process. When an applicant seeks registration for a work involving AI-generated content, the Office applies a two-part inquiry: (1) whether a human authored the work, and (2) whether the human's contribution was sufficiently original and creative to meet the minimum threshold established in Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991).

The February 2023 registration of Zarya of the Dawn — a comic book by Kristina Kashtanova — provided the first published example of the Office's analytical framework. The Office permitted registration of the human-authored text and the selection and arrangement of images but cancelled registration of the individual AI-generated images (produced by Midjourney) because the applicant "did not select or arrange those images within the meaning of copyright law." (U.S. Copyright Office, February 21, 2023 Letter to Van Lindberg.)

The copyright registration process requires applicants to disclose AI-generated content on registration applications. The Office's March 2023 guidance (Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16190) specifies that applicants must identify AI-generated elements and disclaim them. Failure to disclose known AI content can invalidate a registration.

Causal Relationships or Drivers

The human authorship requirement traces to Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), in which the Supreme Court upheld photographic copyright by reasoning that the photographer's creative choices — pose, lighting, arrangement — were expressions of a human mind. This reasoning established the causal link: copyright flows from human mental conception, not from mechanical production.

The rise of large language models and image diffusion systems has stressed this causal chain. With systems such as DALL-E, Stable Diffusion, and Midjourney, the generative process involves probabilistic sampling across billions of training parameters in response to a user-supplied prompt. The Copyright Office and reviewing courts have treated prompt input alone as insufficient human authorship because the user does not control the specific expressive elements of the output — the tool's internal process dictates the final form.

The idea-expression dichotomy is a parallel driver: even where a human conceives an idea and communicates it via a prompt, the idea itself receives no copyright protection. Only the fixed, expressive form does — and if that form is generated autonomously by the machine, the human's contribution remains on the unprotectable idea side of the line.

Legislative inertia is a third driver. Congress has not amended 17 U.S.C. § 102 to address AI authorship, leaving the Copyright Office and federal courts to interpret an analog-era statute in a generative-AI environment. The Office's ongoing AI study (initiated 2023) is examining whether legislative recommendations are warranted, but no statutory amendment had been enacted as of the Office's published record.

Classification Boundaries

The Copyright Office applies a spectrum model rather than a binary rule:

Category A — No Human Authorship (Unregistrable): Output generated entirely by an AI system in response to a prompt, with no human selection, arrangement, modification, or creative post-processing. Example: a single text passage or image generated by typing a prompt with no further editing. No copyright attaches; the work enters the public domain immediately upon creation. See public domain works for implications.

Category B — Mixed Works with Separable Elements: A human-authored text combined with AI-generated illustrations, or a human-composed musical arrangement accompanied by AI-generated lyrics. The human-authored elements can receive copyright protection; the AI-generated elements cannot. Registration must disclaim the AI-generated portions.

Category C — AI-Assisted Works with Substantial Human Creative Control: A human who uses AI as a drafting or rendering tool but exercises creative judgment over selection, arrangement, editing, sequencing, and expression may establish sufficient authorship. The determinative question is whether the human's choices, independent of the AI's autonomous output, constitute original expression. Works in this category are registrable for the human-controlled elements.

Category D — Sufficiently Transformed AI Output: Where a human substantially modifies AI-generated content through original expressive choices — altering, combining, selecting, and rearranging at a level of creative control — the resulting derivative work may qualify for copyright protection in the human-added elements. See derivative works and compilations.

Tradeoffs and Tensions

The human authorship rule creates a structural asymmetry. Developers who invest substantial resources in training and deploying generative AI models receive no copyright in the outputs those models produce. This has prompted interest in alternative forms of protection — trade secret law for model weights, contract law for output licensing, and sui generis database rights (unavailable in the U.S., though operative in the European Union under Directive 96/9/EC).

A second tension involves the prompt-as-authorship argument. Advocates contend that a highly specific, iterative prompt — particularly one produced through a sequence of rejections, refinements, and parameter adjustments — reflects meaningful human creative control. The Copyright Office has declined to adopt this position categorically, noting that even detailed prompt engineering does not give the prompter control over the specific expressive elements the system generates.

A third tension is competitive and economic. If AI-generated content enters the public domain instantly, any party can reproduce, sell, or modify it without licensing obligations. This reduces the economic incentive structure that copyright is designed to create, potentially affecting investment in AI-assisted creative industries — a concern the Copyright Office acknowledged in its 2023 AI policy notice.

The copyright remedies and damages framework also creates a downstream tension: a rights holder who registers only the human-authored portions of a mixed work and later discovers that AI-generated portions were inadvertently included in the registration may face challenges to damages calculations if the infringing material coincides with the disclaimed portions.

Common Misconceptions

Misconception 1: Registering a work that contains AI-generated elements grants copyright over those elements. Incorrect. Registration does not expand protectable subject matter. The Copyright Office examines the human authorship question separately from the registration formality. Elements generated by AI remain unprotectable regardless of whether they are included in a registration application, and applicants are required to disclaim them. (88 Fed. Reg. 16190.)

Misconception 2: Copyright protection attaches automatically to all fixed works. Partially correct but misleading in the AI context. Under copyright law fundamentals, copyright attaches automatically to original works of human authorship upon fixation — but fixation alone does not satisfy the authorship requirement. A work fixed in a tangible medium by an autonomous AI process does not receive automatic copyright protection.

Misconception 3: The company that owns or licenses the AI tool automatically owns copyright in the output. Incorrect under U.S. law. Corporate authorship requires that the output qualify as a work made for hire under 17 U.S.C. § 101, which in turn presupposes a human author whose work was prepared within an employment or qualifying commission relationship. If no human authored the work, there is no authorship to transfer or assign to any entity, including the AI developer.

Misconception 4: Training an AI on copyrighted works automatically creates a licensable right in the outputs. Unsupported by current doctrine. The copyright status of training data and the copyright status of model outputs are legally distinct questions. Courts are currently examining training data infringement claims (see, e.g., ongoing litigation in Andersen v. Stability AI Ltd., N.D. Cal.), but even if training is found lawful, that finding would not create copyright ownership in outputs.

Checklist or Steps (Non-Advisory)

The following steps reflect the Copyright Office's published registration framework for works involving AI-generated content (88 Fed. Reg. 16190):

1. Identify all AI-generated elements in the work prior to registration — text, images, audio, video, or other content generated by any AI system without direct human expressive authorship.
2. Assess human authorship contributions — document the specific creative choices made by human authors: selection, arrangement, editing, sequencing, or transformation of AI-generated elements.
3. Determine separability — evaluate whether AI-generated and human-authored elements can be described as distinct components in the registration application.
4. Prepare the registration application — complete the standard copyright registration form through the Copyright Office's online portal (copyright.gov/registration).
5. Disclose AI-generated content — include a statement of limitation in the application's "Limitation of Copyright Claim" field identifying the AI-generated material being disclaimed.
6. Document the creative process — retain records of human authorship decisions, including prompt iterations, editorial changes, and selection criteria, as supporting evidence if the registration is challenged or questioned.
7. Evaluate previously registered works — if prior registrations include AI-generated content that was not disclosed, assess whether supplementary registration or correction is appropriate under 37 C.F.R. § 202.6.

Reference Table or Matrix

References

• U.S. Copyright Office — Copyright and Artificial Intelligence
• U.S. Copyright Office — Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16190 (March 16, 2023)
• U.S. Copyright Office — Copyright and Artificial Intelligence Policy Statement Part 1 (August 30, 2023), 88 Fed. Reg. 59942
• U.S. Copyright Office — Zarya of the Dawn Registration Decision Letter (February 21, 2023)
• 17 U.S.C. § 102 — Subject Matter of Copyright: In General
• 37 C.F.R. § 202.6 — Supplementary Registration (eCFR)
• Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991) — Supreme Court of the United States
• Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884) — Supreme Court of the United States
• U.S. Copyright Office — Registration Portal