Artificial intelligence is changing how we build products, diagnose disease, and write code. It’s also changing how teams document, claim, and protect new ideas. If you work with models, data pipelines, or AI-enabled features, understanding the patent playbook can save time and prevent costly missteps.
This guide explains what is patentable in the United States today, how AI affects inventorship, and how to write applications that hold up under scrutiny. It also flags the tradeoffs between patents and trade secrets, plus global wrinkles you should know.
If you’re unsure what the “invention” really is in a model and data pipeline stack, Goldstein Patent Law can help you isolate the patentable technical contribution during an inventor interview and translate it into a filing strategy.
TL;DR
- Only human beings can be named as inventors. AI is treated as a tool, not an inventor.
- AI inventions are patentable if the claims fit a statutory category, avoid abstract ideas, and meet novelty, nonobviousness, and disclosure rules.
- After Amgen, broad functional claims with thin technical teaching are risky; enable what you claim.
- File early to avoid prior art problems, and be careful with public disclosures outside the U.S.
- Consider trade secrets for datasets, weights, or prompts you cannot or should not disclose.
What Counts As An AI-Related Invention
Defining an AI-related invention involves identifying any technical solution that utilizes machine learning, neural networks, or complex algorithms to improve data processing and decision-making capabilities. Recognizing these distinctions is essential for inventors to leverage current legal frameworks and secure robust patent protection for their software and hardware innovations.
Goldstein Patent Law helps software and technology teams build a practical patent portfolio strategy. This includes inventor interviews and drafting a detailed provisional to prepare a non-provisional utility application, using a transparent flat-fee structure.
Patent Eligibility Basics
To be patent eligible under U.S. law, your claims must fit one of four categories (process, machine, manufacture, or composition of matter). It must also avoid being directed to a judicial exception like an abstract idea. Most AI work involves software, so examiners apply the familiar Alice/Mayo framework.
You improve your odds by tying the claims to a concrete technical solution, such as an improvement to model training efficiency, on-device inference latency, memory footprint, or sensor fusion performance. High-level results or data processing for a business outcome, without more, often read as abstract.
Our patent lawyers can reframe AI claims around concrete computing improvements (latency, memory, training efficiency, system performance). We can also draft multiple claim types (method/system/computer-readable medium) to reduce §101 risk.
Novelty, Nonobviousness, And Disclosure
Beyond eligibility, you must also be:
- New: No single prior art reference discloses every claim element.
- Nonobvious: The differences over the art would not have been obvious to a skilled person at the effective filing date.
- Enabled and described: The specification must teach skilled readers to make and use the full scope of the claims without undue experimentation and show that you actually possessed the invention.
Amgen tightened expectations for the enablement of broad genus-style claims. If you try to claim a whole class of models, optimizers, or architectures by function alone, be ready to provide representative architectures, parameters, training regimes, and clear guidance that lets others make and use substantially the full scope.
If you’re worried the field is crowded, our patent law firm can plan a prior-art-aware drafting approach so your application highlights what’s truly new and defensible before you file.
Inventorship When AI Assists
In late November 2025, the USPTO issued revised guidance on AI-assisted inventions, confirming that the same inventorship standard applies to all inventions. Only natural persons qualify as inventors under U.S. law, whereas AI systems are treated as tools.
When several people collaborate with AI, traditional joint inventorship principles still apply. U.S. filings must list human inventor(s) who contributed to the conception of the claimed subject matter.
Only natural persons qualify as inventors. AI systems, regardless of sophistication, are tools. When several people collaborate with AI, traditional joint inventorship principles apply. If a foreign filing names an AI as an inventor, the U.S. application must still list human inventor(s) only.
What’s Changed Recently (and Why It Matters)
In the past couple of years, the USPTO has issued AI-specific guidance that affects how software and AI claims are examined. Likewise, it has clarified how inventorship works when AI tools are involved. Successful AI patents are increasingly won or lost on:
- (1) how clearly you describe a technical improvement and
- (2) whether your application teaches enough detail to support the breadth of what you claim.
If you want a quick readout on how these changes affect your model, pipeline, or AI-enabled feature, we can walk through patentability, inventorship, and a filing plan.
Patents vs. Trade Secrets: Which Shield Best Protects Your AI Innovation?
Exactly one of the hardest choices in AI is what to disclose and what to keep secret. This comparison cuts through the complexity so you know exactly what you stand to gain or lose before committing to either path.
| Factor | Patents | Trade Secrets |
| Protection scope | Exclusive rights for claimed subject matter | Protects information kept secret that has economic value |
| Duration | Utility patents generally last ~20 years from the earliest effective nonprovisional filing date (subject to certain adjustments in some cases) | Potentially indefinite while secrecy lasts |
| Disclosure | Public enablement required; published applications and issued patents can become prior art (timing varies) | No public disclosure; must take reasonable secrecy steps |
| Enforcement | Sue for infringement regardless of the copying source | Sue only for misappropriation or breach of confidence |
| Best for | Platform-level technical improvements you can describe and defend | Datasets, weights, prompt libraries, and heuristics that are hard to reverse-engineer |
Want a quick gut check on whether your AI feature is more likely to be treated as a patentable technical improvement (vs. an abstract idea)? Book a no-cost patent strategy session with Goldstein Patent Law to talk through the soft claim strategy and next steps.
Smart Claiming and Spec Tactics for AI
There are unique hurdles of artificial intelligence, and these require a sophisticated approach to how your invention is described and legally defined. Learn these tactical maneuvers to make sure your patent is broad enough to be valuable yet precise enough to withstand rigorous examination.
Draft Claims That Track Technical Benefit
Articulating the specific computational advantages of your AI model, such as speed or efficiency, is the key to overcoming modern eligibility barriers.
- Emphasize concrete computer or device improvements: Lower compute, reduced memory, faster convergence, improved numerical stability, or on-device energy savings.
- Claim multiple categories: Method, system, and non-transitory computer-readable medium.
- Avoid results-only language: Tie steps to specific architectures, layers, loss functions, or training workflows.
Write the Spec to Withstand 101 and 112
Investing time into this detail-heavy documentation gives you the peace of mind that your intellectual property is built on a foundation of scientific clarity and legal strength.
- Provide representative architectures and hyperparameter ranges. Explain why they work.
- Describe training data characteristics, preprocessing, augmentation, and evaluation metrics. You don’t need to dump code, but you must enable skilled readers to reproduce the claimed results without undue experimentation.
- Where helpful, include benchmarks or comparative results that support the asserted technical improvement, without making unnecessary statements that could narrow the claim scope.
- For learned components, describe how they interact with memory, cache, sensors, or network stacks if that is your improvement.
Timing and Filing Strategy
The fast-moving nature of the tech world means that a single day’s delay or an ill-timed public demo can permanently jeopardize your global patent rights.
- The U.S. is a first-inventor-to-file country. File early. A U.S. one-year grace period exists for your own disclosures, but many countries lack such a grace period.
- Use a detailed provisional to lock in a date for what you actually disclose, then follow with a nonprovisional within 12 months.
- Coordinate foreign filings before any public launch, paper, or demo.
We can help by conducting an inventor interview, drafting the provisional specification and drawings, filing it on time, and mapping what should be converted into a non-provisional utility application.
International Protection
Rules vary. In Europe, AI and ML can face ‘mathematical method’ objections unless the claims are tied to a technical purpose and demonstrate a technical effect beyond generic data processing.
Claims fare better when they solve a technical problem beyond generic data processing. These include improving image denoising in a specific camera pipeline or reducing memory bandwidth on defined hardware.
If you expect global users or competitors, our patent attorneys can coordinate an international protection plan alongside your U.S. filings so timing and disclosures don’t accidentally limit foreign options.
Examples
Seeing these outcomes can mean the difference between confidently moving forward and leaving your breakthrough technology unprotected.
On-Device Model Compression for AR Glasses
A team invents a pruning and quantization pipeline that keeps accuracy while cutting memory by 40% and latency by 30% on a specified edge chipset. They draft claims to the pipeline’s ordered steps and to a device executing a pruned graph with specific calibration passes.
The specification teaches representative pruning thresholds, calibration data profiles, and scheduler changes that reduce cache misses. This looks eligible and, with sufficient technical teaching, enablement is strong. Nonobviousness turns on prior art showing similar edge optimizations.
Medical Triage Scoring With LLM Assistance
Clinicians prompt an LLM to propose risk factors, and then the team builds a new triage model. The AI’s prompts and outputs guided brainstorming, but humans selected predictors, designed a feature encoder, and validated thresholds that reduce false negatives on a specified dataset.
Only human inventors are named. Claims focus on the encoder architecture and thresholding method that achieve a documented sensitivity improvement on edge cases. The application includes data characteristics and training details sufficient for skilled practitioners to implement the encoder without undue experimentation.
Actionable Steps / Checklist
Each item on this checklist represents a safeguard that keeps your AI invention protected from the moment of conception through every stage of commercialization. To turn this checklist into a filing-ready plan, our legal team can help you document inventorship, draft a defensible provisional or non-provisional, and build a portfolio roadmap aligned to product releases and funding milestones.
- Capture human contributions: Keep dated notes showing who conceived each claimed feature and why.
- Lock down disclosures: Before preprints, demos, or launches, file at least a thorough provisional.
- Pick what to patent vs. keep secret: Consider patents for platform-level improvements; keep weights, datasets, and prompt libraries secret where practical.
- Draft for eligibility: Frame the invention as a technical improvement to computing or another technology, not just data analysis.
- Enable the scope: Provide representative models, parameter ranges, training workflows, and evaluation evidence that support the breadth of each claim.
- Claim in multiple forms: Method, system, and computer-readable medium.
- Plan internationally: Align timing for PCT or direct foreign filings before any public disclosure outside U.S. grace rules.
- Avoid naming AI as an inventor: List human inventors only, consistent with U.S. law and USPTO guidance.
- Monitor the record: File information disclosure statements and track related publications to avoid surprises.
Glossary
These definitions aren’t just vocabulary; they’re the foundation for making informed decisions that could determine whether your AI innovation is protected or vulnerable.
- Abstract Idea: A nontechnical concept like organizing human activity or basic math that is not patent eligible unless integrated into a practical application.
- Conception: The moment an inventor forms a definite and permanent idea of the complete invention.
- Enablement: A requirement that the specification teach skilled readers to make and use the full scope of the claims without undue experimentation.
- Nonobviousness: The invention would not have been obvious to a skilled person in view of prior art.
- Prior Art: Public information that existed before the effective filing date and can be used against patentability.
- Provisional Application: A lower-formality U.S. filing that secures a date and must be followed by a nonprovisional within 12 months.
- Trade Secret: Valuable information protected by secrecy measures and enforceable against misappropriation.
- USPTO: The United States Patent and Trademark Office, which examines and issues patents.
FAQ
Q: Do I have to name the AI model as a co‑inventor if it generated ideas or code?
A: Only natural persons can be inventors in the U.S. AI systems are treated as tools. Name the human contributors who conceived the claimed subject matter.
Q: Do I need to disclose training data or source code?
A: You don’t have to submit code, but you must enable skilled readers to practice the full scope of the claims. That often means describing architectures, training procedures, and data characteristics with enough detail to reproduce results.
Q: Are prompts, weights, or datasets patentable?
A: Sometimes, but prompts, weights, and datasets are often better protected as trade secrets than patents. If you can claim a technical improvement tied to how prompts, weights, or data interact with computing systems, consider patents. If disclosure would undercut value and reverse engineering is hard, consider secrecy.
Q: We already presented our paper; is it too late to file?
A: In the U.S., you may have up to a one‑year grace period for your own disclosures. Many countries do not offer a broad grace period, so international rights may already be limited.
Q: Can I patent a model trained on third-party copyrighted data?
A: Patent law focuses on technical criteria like eligibility, novelty, nonobviousness, and enablement. Data licensing and copyright are separate compliance issues you should address in parallel.
Final Thoughts
Patenting AI isn’t about sprinkling buzzwords into claims. It’s about documenting a concrete technical advance and teaching others how to achieve it.
To turn an AI feature into a filing-ready application, start with an inventor interview, a detailed provisional, and a clear path to a non-provisional. Contact Goldstein Patent Law for a predictable, flat-fee approach so you can move forward with clarity.
