When the AI Writes the Code: Can You Still Patent, or Even Own, Your Own Product?
Date: 08/18/26
When an agent writes the product, what does the company actually own?
In an earlier article I argued that coding agents collapsed the cost of cloning software, and that patents are becoming the last moat because a utility patent has no independent-creation defense. That piece looked at the competitor who rebuilds you. This one looks at a problem inside the building. The same tools that ship the product can thin the rights you expected to use when someone copies it.
An Inventor Still Has to be a Person
In Thaler v. Vidal, the Federal Circuit held that an inventor under the Patent Act must be a natural person. An AI system cannot be named as an inventor. Thaler answered the easy question.The hard one shows up in ordinary engineering work. A person and an agent build something together. Is the human contribution enough to support a patent?
For a while the answer was murkier than it should have been. The USPTO's February 2024 guidance required the human to make a "significant contribution" to every claim. That test came from disputes between human co-inventors, and was then applied claim by claim. Practitioners started to worry about a gap: no human contributed enough to be the inventor, no AI is allowed to be one, and therefore nobody can patent the invention.
In November 2025 the USPTO rescinded that framework. The same inventorship standard now applies with or without AI. The system is treated as a tool, no different legally from lab equipment or a research database. The significant-contribution test is reserved for sorting out multiple human inventors. What the human still has to do is conceive: possess a definite and permanent idea of the complete invention, understood well enough that a person of ordinary skill could reduce it to practice.
Companies can file. They still have to prove a person conceived the invention. If an agent produces a novel technical solution and the engineer typed "make it faster" and accepted the output, there is a real argument that nobody did. No valid inventor, no patent.
Build the record now, not after a challenge. Notes that capture the human work (the problem you identified, the architecture decision, the direction you gave the agent) are going to matter the way lab notebooks mattered to an earlier generation of inventors.
Copyright Covers Less of the Repo than the Certificate Implies
Copyright in software was always narrower than founders assumed. It protects the literal expression of code, not the function. The U.S. Copyright Office's 2025 report on copyrightability drew the AI line in plain terms. Purely AI-generated output is not copyrightable. Prompts alone generally do not make the prompter an author. Protection extends only to the human contribution: the selection, arrangement and modification a person actually performs.Apply that to a codebase where agents generate most of the code and humans review, prompt and stitch. Your copyright in your own product may be thinner than the registration suggests. The slice a court would protect against a literal copyist may be a fraction of the repository.
Most days this will not matter. Competitors rarely copy code verbatim. It matters in the fights where you actually need the copyright: a departing employee, a licensing dispute or a registration that has to disclaim AI-generated material. If that right is shrinking while the product remains easy to rebuild, the patents have to carry more of the load.
Secrets You Give Away and Licenses You Never Accepted
Trade secret protection depends on reasonable measures to keep information confidential. Engineers who paste proprietary algorithms, schemas or unreleased inventive concepts into consumer-grade AI tools, outside enterprise terms that restrict training and retention, hand opposing counsel an argument that the secret was not kept. The fix is dull and it works: enterprise vendor agreements, a policy the team has read and logging that shows the measures were real.Coding tools that were trained on public repositories. They can emit code that carries open-source license obligations nobody on the team knowingly accepted. Untracked copyleft in the core of an agent-built product is a diligence problem waiting for a term sheet. Acquirers and investors have already put AI-provenance questions on their checklists. What percentage of the codebase is AI-generated, under what tools and terms, is becoming as standard as the open-source audit.
What To Do Now
For startups and other enterprises building with agents:Keep a short invention record that names the person who defined the technical problem and the approach, before the agent built it. Prompts, design docs and architecture decisions are inventorship evidence later.
File provisional patent applications on the human-conceived innovations. The November 2025 guidance lowered the panic around AI-assisted filings. The priority date still goes to whoever files first.
Use enterprise AI tools on enterprise terms. No training on your inputs, no retention you cannot control and a policy people have actually read.
Scan agent-generated code for license problems from day one. The scan is cheap. Cleaning copyleft out of a live deal is not.
For companies that already have a portfolio:
- Update employment and IP-assignment agreements so they reach AI-assisted output. Inventorship determinations should follow the conception standard, not automatically name the project lead.
- Map how much of the current codebase is machine-generated, then adjust what you register and what you expect to enforce.
- Treat the AI-tooling policy, vendor terms and provenance records as a diligence exhibit. They will soon be as expected in a data room as the cap table.
- If copyright and secrecy are doing less work, put more weight on patents that protect the functional concept, regardless of who or what wrote the implementation.
Show that the People Were There
The Patent Office has said that using AI does not disqualify the invention. Congress is still debating broader eligibility for software and AI. None of that substitutes for a record.If you cannot show who conceived the invention, or what a person actually contributed to the code, you will have a product that ships and a thin set of rights when you need to defend it. Trade secrets fail the same way if the company cannot show it kept them.
About Jeffrey R. Schell
Jeffrey R. Schell is a registered patent attorney and Managing Partner of the Mountain West practice at Whiteford, where he counsels technology companies on intellectual property strategy, AI governance and venture growth. A former multi-exit software founder and trained machine learning engineer, he advises companies from startup through exit on building defensible IP positions.
The information contained here is not intended to provide legal advice or opinion and should not be acted upon without consulting an attorney. Counsel should not be selected based on advertising materials, and we recommend that you conduct further investigation when seeking legal representation.