We are living in an artificial intelligence (AI) gold rush. From large language models that write code to machine learning systems that optimize logistics or predict disease, AI innovations are transforming nearly every sector.
For many companies, especially startups and research-intensive enterprises, the key question is no longer whether to protect AI intellectual property (IP), but how.
On paper, patents offer a powerful form of IP protection: exclusive rights that can drive valuation, licensing, and investor confidence. In practice, AI innovators face a formidable barrier: the evolving and unpredictable doctrine of patentable subject matter under 35 U.S.C. § 101.
Since the Supreme Court’s decision in Alice Corp.诉 CLS Bank International (2014), courts have invalidated hundreds of software-related patents as “abstract ideas.”
In this uncertain environment, trade secrets have quietly become a potentially more reliable form of protection for many AI innovations. Nevertheless, the choice is not binary. A growing number of companies are turning to hybrid protection strategies that selectively patent certain technical components while keeping other innovations, 例如, core data, model weights, and training methods, secret.
Recent developments at the U.S. Patent and Trademark Office (USPTO) under its new director, John Squires, suggest a potential shift in how AI and software inventions are examined. But until the courts or Congress provide clear, durable reform, one principle holds: trade secret protection provides reliable protection for AI innovations.
§ 101 for AI Innovations
The basic idea behind § 101 is simple: you cannot patent products of nature, natural phenomena, or abstract ideas. Yet in practice, determining whether a computer-implemented invention is an “abstract idea” has become one of the most contentious issues in U.S. patent law.
After 爱丽丝, the courts adopted a two-step test asking (1) whether the claim is directed to an abstract idea, and (2) if so, whether it contains an “inventive concept” that transforms the idea into a patent-eligible application. In theory, this framework seems simple to apply. In practice, it has generated a decade of uncertainty.
The problem is particularly acute for AI. Many district courts and the Federal Circuit apply § 101 in ways that make it difficult to obtain enforceable patents for inventions that process data, learn from examples, or automate decision-making.
Consider cases like PurePredictive, Inc. v. H2O.ai, Inc., No. 17-cv-03049 (N.D. Cal. Aug. 29, 2017), where the court invalidated claims directed to an automated machine learning model-generation system as an “abstract idea” or In re Board of Trustees of the Leland Stanford Junior University, 991 F.3d 1245 (Fed. Cir. 2021), which struck down a patent application covering mathematical models for predicting genetic variation as “no more than a mental process.”
AI inventions face similar headwinds: courts often conclude they are merely directed at “collecting, analyzing, and displaying data,” a conclusion that frequently results in invalidation.
Add the three-to-five-year timeframe frequently required to obtain a patent, and many AI innovators conclude that secrecy offers both speed and stability.
Trade Secret Protection for AI Innovations
Trade secret protection, for example, under the Defend Trade Secrets Act of 2016 (18 U.S.C. § 1836 et seq.), can provide both immediate and flexible coverage.
Typically, so long as information in question derives independent economic value from not being generally known and reasonable measures are taken to keep it confidential, it will qualify for protection—no filing or examination required.
For AI innovations, trade secret protection can extend to:
-
- Proprietary training data and labeling methodologies
- Feature engineering processes
- Model architectures and weights
- Hyperparameter tuning techniques
- Deployment pipelines and inference systems
These are the components that frequently give AI companies their edge and where patents are frequently weakest.
The Upsides
-
- Automatic and immediate: Protection can begin once secrecy measures are in place.
- Covers the full stack: Data, algorithms, and workflows can all qualify.
- No § 101 risk: Eligibility doctrine doesn’t apply.
- Unlimited duration: As long as the information remains confidential, it is protected.
- Fast-moving alignment: Trade secrets can evolve with your technology in real time.
The Downsides
-
- No protection against independent discovery or reverse engineering.
- Loss of secrecy = loss of protection.
- Difficult enforcement: Proof of misappropriation can be difficult based on publicly available information.
- Employee mobility risks: Particularly acute in states like California, where non-competes are generally prohibited.
- Limited valuation visibility: Investors sometimes undervalue trade secrets in IP portfolios.
Still, when compared to the cost, delay, and potential uncertainty of obtaining enforceable patents covering AI innovations, many businesses see trade secrets as the pragmatic choice.
A Hybrid IP Strategy Can Be The Best of Both Worlds
Forward-looking AI companies are beginning to use hybrid approaches, i.e., patenting where enforceable patents are possible and keeping the rest under wraps.
1. Patent What Can Survive § 101 In Court
Target innovations with concrete, measurable technical improvements, for example:
-
- Hardware or software architectures that optimize resource use;
- Methods for solving specific engineering problems, such as latency or data integrity.
2. Keep the Secret Sauce Secret
Maintain confidentiality over:
-
- Training data and labeling processes;
- Model weights and fine-tuning protocols;
- Pipeline orchestration and monitoring systems.
3. Build a Secrecy Framework
-
- Implement NDAs, access segmentation, and encryption.
- Consider implementing digital tools to track and document trade secrets as well as access to them (for instance, 娓娓道来).
- Educate employees regularly on confidentiality practices.
4. Stay Agile
As technology matures, revisit your balance between patents and trade secrets. Early-stage technologies may favor trade secrets; later, as the risk of reverse engineering grows, patents may make more sense.
New Hope?
In September 2025, John Squires became the 60th Director of the USPTO and immediately began signaling a different approach toward AI and software inventions.
In his first month, Squires personally signed two patents in historically challenging domains (cryptocurrency infrastructure and medical diagnostics) and invoked the spirit of Samuel Morse to emphasize protecting applied technologies.
He also convened an Appeals Review Panel that reversed a prior PTAB rejection of an AI-related patent under § 101 in Ex parte Guillaume Desjardins et al.. The panel found that the claims improved “how the machine learning model itself operates,” suggesting a shift toward recognizing technical innovation within AI systems.
Squires’ moves are likely encouraging to AI innovators. But the USPTO cannot overrule Federal Circuit or Supreme Court precedent. As such, any patents covering AI innovations that issue under Squires’ new approach will still face the same uncertainty in court.
Conclusion: Pursue A Hybrid IP Strategy
For now, AI innovators should approach IP pragmatically, not doctrinally. The strongest strategies frequently combine patents and trade secrets, tailoring each to the company’s risk profile and technical realities.
-
- Use patents for demonstrable technical improvements that can withstand § 101 scrutiny.
- Rely on trade secrets for data, models, and know-how that likely will not result in an enforceable patent.
- Continuously reassess as the technology and law evolve.
In the end, success in the AI gold rush probably will not come from filing the most patents—it will likely come from knowing what to reveal, what to conceal, and when to do so.
About Ben Herbert:
本-赫伯特 is a partner with Miller Barondess, LLP and an accomplished trial lawyer specializing in patent infringement and trade secret misappropriation litigation. He is co-leader of the firm’s intellectual property practice and is adept at helping his clients resolve issues related to intellectual property through counseling and litigation.
Ben has a record of success representing plaintiffs in intellectual property trials, having been a key member of three trial teams that secured more than $1.5 billion in jury verdicts in the span of two years. He has litigated complex patent and trade secret cases involving diverse technologies, including digital radios, medical devices, computer software and hardware, graphics processors, bitcoin mining, portable power generation, and satellite technology. He also has significant experience representing clients in Patent Office proceedings, including inter partes review, as well as before the U.S. Court of Appeals for the Federal Circuit.
Follow Ben on Linkedin.
Why are programs and data considered trade secrets?
Because they contain algorithms, models, processes, and customer insights that provide competitive advantage. When exposed, the loss is immediate and irreversible.
What is the first step to protecting programs and data?
Create an inventory and identify which assets qualify as trade secrets, including code, models, datasets, prompts, and internal tools.
Why is classification important for trade secret protection?
Classification clarifies sensitivity, risk, and required controls. It strengthens governance and ensures teams understand how assets should be handled.
How does access control prevent trade secret loss?
Limiting access by role, tracking interactions, removing outdated permissions, and restricting repositories prevents internal misuse and accidental exposure.
How does AI increase the risk of trade secret leakage?
Employees often paste confidential information into public AI tools, causing it to become training data. Governance and private AI tools prevent this.
What legal agreements support trade secret protection?
NDAs, invention assignment agreements, contractor agreements, and offboarding processes ensure confidentiality obligations are clear and enforceable.
Why is employee training essential for protecting trade secrets?
Most exposure occurs through human behaviour. Training builds awareness and creates a culture of confidentiality across the organisation.
How do audits improve trade secret protection?
Audits identify weak points like outdated access, unsecured repositories, shadow AI use, and untracked datasets — key for legal defensibility.
How does Tangibly help protect programs and data?
Tangibly provides automated trade secret identification, classification, access tracking, governance workflows, and AI-driven risk detection.
