State AI Whistleblower Laws Have a Trade Secret Problem

A December 2025 White House Executive Order directed federal agencies to evaluate whether state AI laws conflict with federal objectives. California and New York, the two US states with comprehensive frontier AI legislation, now face pressure to demonstrate that their frameworks can effectively govern emerging AI risks. Both states enacted whistleblower protections as a core regulatory mechanism. Whether those protections actually function will shape debates about federal preemption, the doctrine under which federal law can override state law.

Why AI Whistleblowers Must Disclose Technical Specifics

Criminal and civil liability address harms after they occur. The value of California’s SB 53 and New York’s RAISE Act lies in identifying dangerous AI capabilities before public deployment. AI presents an irreversibility problem: once a model is released, its capabilities propagate. Other systems train on its outputs. Removing a product after one week does not recall the capability.

Consider an employee who discovers that a model generates illegal content on demand. If that employee can report the capability with sufficient technical specificity, regulators might intervene before release. This is the framework working as intended. The employee observes a pre-regulatory risk, meaning conduct that poses danger but precedes existing regulation, reports it to a designated agency, and enables preventive action.

State agencies designated to receive these reports will face resource constraints. As AI systems enter specialized domains such as biology, cybersecurity, finance, and critical infrastructure, the technical expertise required to evaluate reports will exceed agency capacity. Whistleblowers must provide specific information demonstrating a model’s potential for widespread harm. The information that makes a report actionable is often the same information that AI companies protect as trade secrets, proprietary information that derives economic value from secrecy. Model architecture, training data composition, internal safety evaluations, and capability assessments typically qualify. Under trade secret law, disclosing this information exposes whistleblowers to civil damages and potential criminal liability.

Trade Secret Law Blocks Effective Disclosure

The federal Defend Trade Secrets Act provides immunity for disclosures made to report a “suspected violation of law.” Pre-regulatory AI risks involve capabilities that are dangerous but not yet prohibited. Federal immunity is unavailable.

Forty-eight US states have adopted the Uniform Trade Secrets Act, a model statute governing misappropriation claims. Courts have declined to recognize a public interest or public safety exception. The Uniform Law Commission considered and rejected such an exception when drafting the model statute.

California’s general whistleblower statute, Labor Code § 1102.5, explicitly excludes disclosures of trade secret information from protection. SB 53 preserves this exclusion. New York’s RAISE Act includes anti-retaliation provisions but remains silent on trade secret immunity. New York is the only US state without the Uniform Trade Secrets Act, relying instead on common law, which still provides remedies for misappropriation. The result in both states: whistleblowers face potential liability for disclosing the technical specifics that would make their reports actionable.

States Should Fix Their Own Frameworks

California and New York built whistleblower architectures that assume employees can substantiate catastrophic risk claims without disclosing protected information. This assumption likely conflicts with the reality of AI systems, where evidence of risk is often proprietary.

Solutions exist. States could enact safe harbors, statutory provisions excluding from trade secret liability disclosures of reasonably necessary information to designated bodies for reporting significant AI-related risks. States could limit immunity to confidential disclosures, require a “reasonable belief” standard, or restrict protection to specific harm categories.

This is the laboratory model working as intended. California and New York are positioned to experiment with approaches that other states can evaluate and adapt. The alternative, waiting for federal preemption of state trade secret law, would eliminate the opportunity for state-level innovation.

States claiming authority to regulate frontier AI should demonstrate that claim through frameworks that actually enable the disclosures those frameworks depend on. The current structure protects whistleblowers from employer retaliation while leaving them exposed to trade secret claims. Closing this gap is essential if state-level AI governance is to function as more than formal architecture.

Author Bio: Michael Endrias is an Internet Law and Policy Foundry Fellow and a J.D. Candidate at Howard University School of Law. He is interning with the Internet Infrastructure Coalition and the Surveillance Technology Oversight Project. He likes urban biking and open source-intelligence work. This piece draws on his pre-published paper, “California’s Trade Secret Trap,” done under SPAR.