Trade Secrets and AI — Calculating the Odds of Protection
This article is part of the “Defending the Algorithm™” series and was written by Pittsburgh, Pennsylvania Business and IP Trial Lawyer Acacia B. Perko, Esq., with research and drafting assistance from OpenAI’s GPT-5. The series explores the evolving intersections of artificial intelligence, intellectual property, and the law. GPT-5 may generate errors, but the author has verified all facts and analysis for accuracy and completeness.
How Businesses Can Protect Innovation in the Age of Machine Intelligence.
In our continuing Defending the Algorithm™ series examining AI litigation and emerging risks, we now turn to trade secret protection. Following the previous analysis of the Anthropic copyright settlement, which deemed the training of LLM’s on legally obtained copyrighted materials as fair use, an important question arises in the other direction: What are the real odds that trade secret law will protect AI innovation?
A groundbreaking Massachusetts case — OpenEvidence v. Pathway Medical Inc. — is giving us the first concrete data points to help answer that question.
Trade Secrets and AI Collide in OpenEvidence v. Pathway Medical Inc., Case No. 1:25-cv-10471 (D. Mass.), in the U.S. District Court for the District of Massachusetts.
Artificial intelligence is reshaping industries — from healthcare to finance to e-commerce. It’s also reshaping how companies protect innovation.
The OpenEvidence lawsuit, filed in February 2025 (Case No. 1:25-cv-10471, U.S. District Court for the District of Massachusetts, Judge Myong J. Joun), represents the first meaningful test of whether the hidden architecture of an AI model — specifically its AI system prompts — can qualify as a trade secret. These prompts are the embedded instruction sets that shape how large language models (LLMs) reason, respond, and maintain consistency. They’re the “rules of engagement” written by developers to control tone, structure, and behavior — never visible to the end user, but critical to how the model functions.
The case also breaks new ground by asking whether so-called prompt injection attacks — queries deliberately designed to trick an AI system into disclosing those hidden instructions known as AI system prompts — amount to trade secret misappropriation. In essence, the court is being asked to decide whether manipulating an AI’s interface can constitute digital theft.
In the 143 paragraph Original Complaint, Plaintiff OpenEvidence Inc. alleges numerous federal and state claims, including misappropriation of trade secrets under 18 USC § 1836 et seq.; violation of computer fraud and abuse act under 18 USC § 1030, breach of contract, violation of digital millennium copyright act under 17 USC § 1201; and unfair competition and unfair deceptive acts in conduct of trade or commerce under Mass. G.L. ch. 93A § 11. In sum, Plaintiffclaims that Defendant Pathway Medical Inc. intentionally exploited its AI platform through a series of engineered prompts to extract proprietary instructions, later using them to accelerate development of a competing medical AI product. Defendant, in turn, argues that no trade secrets were taken — claiming that the AI system prompts, once exposed through user interactions, lose any reasonable expectation of secrecy and fall outside traditional trade secret protection.
A Modern Framework: Updating Our “Odds” of Protection
In traditional trade secret cases, protection is often viewed in binary terms: either information is secret, or it isn’t. But AI complicates that equation.
Borrowing from Bayesian reasoning — a concept foundational to modern AI — we can think of trade secret protection as a “prior probability” that evolves as new facts and case law emerge. Each decision, like OpenEvidence, provides data that helps businesses and their lawyers update their understanding of how courts are likely to treat AI-driven innovation and its intersection with data privacy.
Before OpenEvidence, the idea of protecting AI systems under trade secret law was like a model trained on incomplete data — theoretically sound, but untested in the wild. The doctrine fit AI systems well on paper, but we lacked real-world examples of how courts would treat the secret inner workings of an algorithm. Now, as judges begin to grapple with whether AI system prompts qualify as trade secrets, we’re finally seeing those theories tested in practice.
Why Trade Secrets Are Becoming a Go-To for AI Innovation
Copyrights and patents come with significant AI-specific limitations, but the law on patent eligibility in particular is rapidly evolving as the technology advances:
- The U.S. Copyright Office refuses to register works created entirely by AI.
- The USPTO requires human inventors, and courts frequently reject patents where AI contributions can’t be separated from human invention (although the trend at the USPTO may be in the opposite direction as AI becomes more ubiquitous).
By contrast, trade secret law doesn’t depend on authorship. Under both the Pennsylvania Uniform Trade Secrets Act (12 Pa.C.S. § 5301 et seq.) and the federal Defend Trade Secrets Act (18 U.S.C. § 1836), a trade secret is protectable if it:
- Derives independent economic value from remaining secret, and
- Is subject to reasonable efforts to maintain secrecy.
That flexibility makes trade secrets particularly useful for protecting data that may not qualify for copyright or patent protections, or which inventors do not wish to disclose in a public patent application or issued patent such as:
- Model architecture and algorithms
- Training data and methodologies
- System prompts and fine-tuning instructions
- Proprietary output structures and response frameworks
Teaching Point: Think of trade secrets as the “secret recipe” of AI innovation. Coca-Cola never patented its formula, yet has protected it for over a century through secrecy. AI developers can do the same — if they treat their data, prompts, and methods with the same rigor.
The OpenEvidence Case: Testing the Boundaries
OpenEvidence, Inc., based in Massachusetts, developed an AI tool that provides medical professionals with real-time answers to clinical questions. Its value rested on carefully crafted AI system prompts — hidden instructions that guided the LLM in how the AI reasoned and responded.
In its Original Complaint,PlaintiffOpenEvidence Inc. alleges that the AI system prompts have independent economic value by claiming:
“The system prompt is code that provides the LLM with its core, and critical, background and situational context. A system prompt also sets the LLM’s role, “personality,” and subject matter expertise. And it contains a set of governing rules and boundaries for interacting with users and providing responses. It is the constitutional framework of any LLM, and it is—accordingly—a proprietary and extremely valuable asset for any AI company”.
Original Complaint, paragraph 3.
Further, Plaintiff alleges that its competitor, Defendant Pathway Medical, Inc., intentionally probed its system through prompt injection attacks to extract those instructions and speed development of its own competing product.
The case alleges violations of:
Defendant’s Motion to Dismiss, filed June 16, 2025, argues that OpenEvidence failed to allege access to any nonpublic information — contending that “prompt injection” through a public interface is simply lawful reverse engineering – which can be a strong defense to claims of trade secret misappropriation.
In its Amended Complaint, filed August 21, 2025, Plaintiff OpenEvidence, Inc. reframed the dispute as an “elaborate conspiracy to steal proprietary AI technology” through unauthorized access and systematic data extraction. The case now awaits further proceedings, with Defendant’s response due October 28, 2025.
Key Legal Question: Can AI system prompts remain protected as trade secrets when they’re potentially discoverable through clever user interactions? Is “tricking” a model into revealing its internal logic an act of misappropriation — or just permissible competitive investigation?
Understanding the Risk: What OpenEvidence Tells Pennsylvania Businesses
From a practical standpoint, OpenEvidence offers a real-world data point to help Pennsylvania businesses recalibrate their expectations.
- Courts are actively considering whether prompt injection equals “improper means” under trade secret law.
- The case highlights that website Terms of Use alone may not be enough protection — technical controls and audit logs are critical.
- The likely success of similar claims depends on whether plaintiff trade secret owners can show reasonable, documentable efforts to protect AI trade secret assets.
If we think in Bayesian terms, this case nudges our posterior probability of enforceable AI trade secret protection upward — from theoretical optimism to cautious realism. The odds are improving, but the burden of proof remains somewhat confusing, and likely rather steep.
Practical Strategies for Protecting AI Trade Secrets
For Pennsylvania businesses deploying AI enterprise software, protection requires both prevention and proof. Courts will ask: What did you actually do to keep it secret?
1. Segment and Define
Identify and categorize what constitutes a trade secret within your AI ecosystem — prompts, architecture, or data. Document the economic value of that secrecy.
2. Implement Access Controls
Restrict internal and external access to AI system components. Maintain detailed logs, enforce role-based permissions, and deploy rate limits to prevent data scraping.
3. Strengthen Contracts
Ensure your Terms of Use, NDAs, and vendor agreements explicitly prohibit reverse engineering and unauthorized data extraction.
4. Monitor and Detect
Track unusual query behavior that may signal prompt injection or scraping attempts. Document each incident — contemporaneous records are key evidence of diligence.
5. Manage Training Data
Avoid training AI models on confidential data that may surface in outputs. Maintain data provenance and segregation between proprietary and public sources.
Pennsylvania Note:
Under the Pennsylvania Uniform Trade Secrets Act, courts focus on “reasonable efforts under the circumstances.” This requires more than policies — it demands demonstrable technical and procedural safeguards. Merely relying on contract terms may not meet the threshold.
Trade Secrets vs. Patents and Copyrights (Current Status)
Teaching Point: Patents are like a visible fence — everyone can see what you own. Trade secrets are buried treasure — they only remain yours as long as no one else digs them up.
Pennsylvania Trade Secret Law and AI Protection
Pennsylvania adopted the Uniform Trade Secrets Act in 2004, providing robust protection for confidential business information.
Courts evaluating AI-related trade secrets will focus on three questions:
- Were reasonable protective measures implemented by the owner?
- Was the information readily ascertainable through lawful means?
- Did the alleged misconduct cross the line from competitive intelligence to misappropriation?
Given the state’s strong precedent on confidentiality and the growing use of AI in regulated industries, Pennsylvania is poised to become an important jurisdiction in defining trade secret boundaries for AI systems.
Key Takeaways for Pennsylvania Businesses
- Trade secrets are a natural fit for AI. They can protect algorithms, training data, and prompts with or without human authorship.
- But they’re fragile. Prompt attacks, employee mobility, and interface access can all erode protection.
- Case law is developing. OpenEvidence may become a landmark on whether prompt injection counts as misappropriation.
- Layered strategies are essential. Combine trade secret protection with contracts, patents, and technical safeguards.
- Reasonable measures are the key. Pennsylvania courts will expect more than NDAs — they’ll look for tangible, documented steps.
Conclusion: Balancing AI Power with Secrecy Discipline
AI innovation operates in a paradox — systems built to generate and share information must also safeguard it.
OpenEvidence reminds us that trade secret protection is not automatic. It’s earned through proactive measures, technical rigor, and continual reassessment as technology evolves.
The Bayesian mindset — constantly updating your assumptions as new facts emerge to predict the probability of a future event— offers a practical analogy for AI governance. Each new case, each enforcement action, helps refine how businesses should evaluate and fortify their own IP protection.
Handled wisely, trade secrets can give companies a durable edge in a competitive AI-driven economy. Handled carelessly, they can become litigation liabilities.
As the OpenEvidence case develops, Pennsylvania businesses should monitor the progress of the case closely — and treat it as a living lesson in how to defend the algorithm.
This article is part of Houston Harbaugh’s ongoing “Defending the Algorithm™” series, exploring AI litigation and risk analysis through practical frameworks for businesses and innovators.
