Businesses protect AI algorithms through a combination of patents, copyrights, trade secrets and trademarks, supported by contracts that confirm the company owns what its employees and contractors create. Patents can cover novel technical applications of AI, copyright protects source code, trade secret law often protects training data, model weights and tuning methods, and trademarks protect the product brand. The right mix depends on what is valuable and how easily competitors could copy it.
A startup’s AI advantage usually comes from several layers: the model architecture, the training data, the code around it and the know-how of the team. Each layer may be protected differently, and some may not be protectable at all unless the company takes deliberate steps early.
Patent Protection
A U.S. utility patent gives its owner the right to exclude others from making, using or selling the claimed invention, generally for 20 years from filing. AI inventions face a particular hurdle: subject matter eligibility. Under the Supreme Court’s Alice framework, claims directed to abstract ideas, including mathematical algorithms and mental processes, are not patentable unless they add an inventive concept that amounts to significantly more. USPTO guidance in recent years has emphasized that claims reflecting a specific improvement to computer functionality or another technology are more likely to qualify.
Practical points for AI patent applications:
- Describe the specific technical problem and how the invention solves it, such as reduced computing resources, improved training efficiency or a new way of processing particular data.
- Avoid claims that merely recite applying machine learning to a business problem in general terms.
- Meet the novelty and nonobviousness requirements and describe the invention in enough detail that a skilled person could make and use it.
- File before public disclosure where possible. The U.S. is a first-inventor-to-file system with a limited one-year grace period for an inventor’s own disclosures, and many foreign countries offer no grace period.
- Identify human inventors. Courts have held that an AI system cannot be named as an inventor; a natural person must contribute significantly to the conception of the invention.
Copyright Protection
Copyright protects original works of authorship, including source code, documentation and certain compilations of data. It arises automatically, but it protects expression, not the underlying idea or method. A competitor that independently writes different code to implement the same algorithm generally does not infringe.
- Register important code. For U.S. works, registration is generally required before suing for infringement, and timely registration makes statutory damages and attorney’s fees available. The Copyright Office allows deposits of source code with trade secret portions redacted.
- Use notices and license terms. Copyright notices and clear license terms support enforcement and help defeat claims of innocent infringement.
- Watch AI-generated code. The Copyright Office and federal courts have held that material generated by AI without sufficient human authorship is not copyrightable. Code written largely by AI coding tools may have thinner protection, which makes trade secret protection more important.
- Model weights. Whether trained model weights are protected by copyright is unsettled. Many companies rely on trade secret law and contracts instead.
Trade Secret Protection
For many AI companies, trade secrets are the most important form of protection. Information qualifies when it derives economic value from not being generally known and the owner takes reasonable measures to keep it secret. Training datasets, data-cleaning methods, model weights, hyperparameters, prompts and evaluation methods can all qualify. Both the federal Defend Trade Secrets Act and California’s Uniform Trade Secrets Act provide remedies for misappropriation.
Reasonable measures typically include:
- Access controls, encryption and logging for code, data and model weights.
- Confidentiality agreements with employees, contractors, investors and partners before sharing sensitive information.
- Need-to-know access and clear labeling of confidential materials.
- Exit procedures that recover devices and remind departing employees of their obligations.
- Policies on what employees may enter into third-party AI tools.
Because California generally voids employee non-compete agreements, trade secret law and confidentiality agreements are the main tools for protecting this information when employees leave. Employee confidentiality agreements should include the whistleblower immunity notice described in the Defend Trade Secrets Act to preserve certain remedies. See our post on protecting company secrets from departing employees.
Trademark Protection
Trademarks protect the names, logos and slogans used to identify your AI products. Before adopting a brand, run a clearance search, then consider applying to register with the USPTO. Use the TM symbol for unregistered marks and the registered symbol only after registration issues. Monitor for confusingly similar marks. Our guide to trademark basics covers the process.
Ownership: The Foundation of Every Strategy
None of these protections helps if the company does not own the IP. Founders should assign pre-formation work to the company, employees should sign invention assignment agreements, and contractors should sign agreements that assign their work product. In California, employee invention assignment agreements must respect Labor Code section 2870, which protects certain inventions an employee develops entirely on their own time without company resources. Our post on confidential information and invention assignment agreements explains these documents.
Also review open source licenses and third-party data and model licenses. Some licenses impose conditions that can affect how you distribute your product or whether you can use a model commercially.
The Takeaway
AI algorithms are usually protected through a layered strategy: patents for novel technical improvements, copyright for code, trade secrets for data, weights and know-how, and trademarks for the brand, all built on clear ownership agreements. Decide early which assets to disclose through patents and which to keep confidential, because public disclosure can foreclose both options.
Frequently Asked Questions
Can you patent an AI algorithm?
Sometimes. An abstract mathematical algorithm cannot be patented by itself, but an AI-related invention may be patentable if the claims are directed to a specific technical improvement or practical application rather than an abstract idea, and the invention is new and nonobvious. How the application is drafted matters a great deal.
Is it better to patent an AI model or keep it a trade secret?
It depends on the technology. Patents require public disclosure in exchange for exclusive rights for a limited term and work well for inventions that competitors could discover by examining your product. Trade secrets can last indefinitely and suit things like training data, model weights and tuning methods that cannot be reverse engineered, but they offer no protection against independent development.
Can an AI system be listed as an inventor on a patent?
No. The Federal Circuit has held that an inventor under U.S. patent law must be a natural person. AI-assisted inventions may still be patentable when one or more humans made a significant contribution to the conception of the invention, and the USPTO has issued guidance on that question.
Do I need to register a copyright for my AI software?
Copyright protection exists automatically once original code is fixed in a tangible form, but registration matters. For U.S. works, registration is generally required before filing an infringement lawsuit, and timely registration makes statutory damages and attorney’s fees available.
How Sutter Law Can Help
We advise AI startups on IP ownership, confidentiality and invention assignment agreements, licensing, trade secret and trademark matters. Learn more about our intellectual property practice, or contact us for a free consultation.
This article is general information, not legal advice. Reading it does not create an attorney-client relationship.
