AI is changing how businesses build. Copyright needs to catch up.

The work did not begin with the final prompt

Businesses need clearer U.S. copyright rules for software built with AI. Existing law already recognizes human authorship in some AI-assisted work. The unresolved practical question is how a company demonstrates that authorship when the finished product emerges through repeated direction, editing, code generation and human decisions.

In our work, a useful application rarely arrives because someone asked for a website and accepted the first result. We study the business, decide what the product needs to do, shape the experience, reject weak implementations and keep revising. The finished work reflects knowledge accumulated long before the model entered the conversation.

That is why the one-prompt view of AI feels so incomplete. AI can produce a quick draft. Building something specific enough to run a company requires judgment about what belongs, what fails and what should change.

Those decisions have commercial value. Which ones amount to copyrightable authorship is a narrower legal question—and one businesses deserve to be able to answer before a dispute.

Using AI does not automatically erase copyright

The U.S. Copyright Office says AI assistance does not bar copyright protection. Human-authored expression that remains visible in the result, creative modifications, and sufficiently creative selection or arrangement may qualify. Purely generated material does not receive protection just because someone requested it. That distinction appears in the Office’s January 2025 explanation of AI copyrightability.

The harder boundary concerns control. In its copyrightability report, the Office concludes that prompts alone, with the technology it analyzed, do not establish sufficient human control over the output’s expressive elements. Repeating prompts or spending many hours on them does not automatically change that. Copyright rewards original authorship, rather than effort alone.

We should take that distinction seriously. A detailed instruction is not necessarily the same thing as creating the resulting expression. Approving an attractive result also does not settle who authored its contents.

But there is a substantial range of work between accepting a generated draft and personally typing every character. A person may supply original material, revise specific passages, combine elements deliberately and carry those decisions through successive versions. The question is how those contributions appear in the finished work. “Made with AI” tells us very little by itself.

Ownership becomes a business question before it becomes a lawsuit

Imagine a distributor building a custom customer portal. Its team supplies original text and designs, directs AI-assisted development, edits code and combines those pieces with existing libraries. This is an illustrative project, not a legal conclusion about a particular product.

The owner may eventually want to license the portal, sell the business or stop someone copying distinctive parts of it. “We paid for the build” does not answer every question about those rights. Section 202 of the Copyright Act distinguishes ownership of a copy from ownership of copyright. Section 204 generally requires a signed writing for transfers of copyright ownership, apart from transfers by operation of law. The ownership provisions matter even when no AI is involved.

AI adds another question to that review: which parts contain protectable human expression, and which are generated material or third-party components? A vendor’s promise that the customer owns the deliverable needs to be read alongside the rights that actually exist.

Our concern is the burden that uncertainty creates for an operating company. A team may know exactly how it shaped the product and still struggle to describe its human contribution in terms a registration process, buyer or counterparty can evaluate. That is a reason to improve the framework. It is not evidence that the entire product has become public property.

Proprietary software has more than one source of protection

Copyright has always had limits in software. The Copyright Office’s computer-program guidance distinguishes protected expression from functional elements such as algorithms, logic, functions and system design. Even fully human-written code does not create copyright ownership over the underlying business method.

Trade secrets address a different concern. Confidential code or know-how may qualify when it derives economic value from secrecy and the business takes reasonable steps to keep it secret. The USPTO’s explanation makes those conditions central. Material exposed publicly cannot simply be treated as though it remained confidential.

For an owner, the useful exercise is to map the assets: human-authored code and content, confidential processes, licensed components, generated material and the agreements governing their use. Have counsel assess the applicable rights and restrictions. The answer may differ across parts of the same application.

That also makes “Can anyone steal it?” too broad a question. Copying protected expression, obtaining confidential material improperly and independently implementing similar functionality are different situations. The presence of AI does not collapse them into one rule.

Recognize demonstrable creative control, and make it practical

The Copyright Office’s 2025 report concludes that existing law is sufficient and that a case for additional protection has not been made. We think the business need for clearer application remains compelling. Our proposed changes focus on human contributions, with targeted legislative clarification where guidance cannot resolve an established gap.

• Publish software-specific examples. Show how the authorship analysis applies to an AI-assisted application developed over many iterations: original inputs, human code edits, generated components and creative arrangements. Explain what qualifies and what does not.
• Make evidence of authorship practical to preserve and submit. Version history, original drafts and specific before-and-after changes can help describe a person’s contribution. Guidance should explain how such records are evaluated without implying that a long prompt log proves authorship.
• Clarify mixed-work claims. Businesses need understandable examples of how to identify human contributions and generated material within the same product. A clear claim should describe its limits as well as its protection.
• Address genuine gaps narrowly. If demonstrable human control over original expression falls outside existing protection because of how an AI tool executes it, lawmakers should examine that boundary. Any change should preserve competition and the rights of other creators.

These are policy proposals. They are not a description of new rights already available to a company. The point is to make the human contribution legible without assuming that every person who commissions or approves an output authored it.

A stronger rule should still leave room for other people to build

There is a serious objection to broad protection for generated output: a business could produce enormous volumes of material and try to claim exclusive rights over it without making a meaningful creative contribution. We do not want a system that rewards that behavior.

Nor should a reform turn a common booking flow, pricing method or software function into exclusive property merely because someone generated it first. A competitor should remain able to build its own implementation within the applicable rules.

The case we are making depends on identifiable human expression and control. Experience matters to the product. Hours matter to the budget. Neither should substitute for evidence of authorship.

That is also why “creative director” is an analogy, not a legal shortcut. It describes the experience of steering the work. The actual contributions still need to be examined.

Keep building, and keep a record of what your people create

For a company investing in custom software now, waiting for perfect certainty has a cost too. We would build a better record alongside the product: retain original designs and text, meaningful revisions, contributor identities, third-party licenses and the agreements that govern the engagement. Preserve relevant AI-tool terms as well. Those records support review; they do not manufacture copyright.

Keep confidential material under appropriate access controls. Ask an IP lawyer to assess ownership, registration and licensing when those rights matter to the investment or a transaction. A general policy article cannot determine the protection available for a particular codebase.

In product development, handover should make it possible to understand how the software was assembled and who contributed what. That record is useful long before anyone considers litigation.

AI lets a capable team turn its judgment into working software faster. The rules should give that team a clearer way to demonstrate the original human work inside the result. Businesses deserve to know what they are building—and what they can protect.

If you are planning a custom product, talk with us about the build and handover.