AI Export Controls for Research: Staying Compliant with EAR, ITAR, and Grant Requirements

The AI Export Control Challenge for Research

Research institutions sit at the intersection of two competing imperatives: the open exchange of scientific knowledge, which has driven American innovation for decades, and national security concerns that restrict the international sharing of certain technologies and data. When research institutions adopt AI systems — particularly cloud-based AI services — they may inadvertently trigger export control violations by transferring controlled technical data across borders without appropriate licenses.

The problem is particularly acute in research environments where:

• Data sensitivity varies widely — Some research involves basic science with no export restrictions; other projects handle dual-use technologies (civilian and military applications), defense articles, or sensitive personal information.
• International collaboration is common — Many research projects involve collaboration with researchers from countries subject to trade sanctions or requiring special authorization.
• Federal funding imposes additional requirements — NSF, NIH, DARPA, DOE, and other federal agencies have specific data handling and reporting requirements for AI-related research.
• Cloud AI creates cross-border data flows — Cloud AI providers often store data in multiple jurisdictions, making it difficult to control where research data travels.

Federal export control regulations are complex, heavily enforced, and carry severe penalties for non-compliance. Violations can result in civil penalties up to $300,000 or twice the value of the violation per violation, criminal penalties including imprisonment, debarment from federal contracts, and reputational damage that can destroy institutional research programs.

Why AI Systems Trigger Export Concerns

• Technology transfer through data ingestion — When researchers input technical data into cloud AI systems, that data may be considered an "export" if the AI provider's personnel or infrastructure are located outside the US, or if the data will be processed in a foreign jurisdiction.
• Sanctioned country access — If AI provider personnel from sanctioned countries (China, Iran, North Korea, Russia) have access to research data, that constitutes a prohibited export to those countries.
• Dual-use technology classification — Many AI technologies themselves fall under export controls (e.g., advanced machine learning algorithms, encryption tools, cybersecurity AI). Deploying these systems requires export license review.
• Training data provenance — If AI models are trained on data containing export-controlled technology, the trained model itself may be classified as an export-controlled item.
• Personnel access — Foreign nationals working on AI projects must have appropriate clearance or training before accessing controlled technical data.

EAR Export Control Requirements

The Export Administration Regulations (EAR), administered by the Bureau of Industry and Security (BIS), control the export of "dual-use" items — commercial products and technologies that have both civilian and military applications. Most AI-related exports fall under the EAR rather than ITAR.

Key EAR Concepts for AI Research

Countries Subject to Trade Embargoes

The EAR imposes near-total embargo on exports to Cuba, Iran, North Korea, Syria, and the Crimea, Donetsk, and Luhansk regions of Ukraine. Exports to China, Russia, Venezuela, and Myanmar face significant licensing requirements. Using cloud AI systems that operate in these jurisdictions may constitute illegal exports.

• Embargoed countries — No exports permitted except for specific license exceptions (unlikely for AI).
• Country group D:1 — National security concerns (includes China, Russia, Belarus, Cuba, Iran, North Korea, Syria). Licensing requirements are strict, especially for advanced AI technologies.
• Foreign Direct Product Rule — Even if you're in the US, products made abroad using US technology may require export licenses when sold to embargoed countries.

ITAR and Defense-Related AI

The International Traffic in Arms Regulations (ITAR), administered by the Directorate of Defense Trade Controls (DDTC), controls the export of defense articles and services listed on the United States Munitions List (USML). While most AI falls under EAR, defense-related AI systems fall under ITAR and are subject to stricter controls.

When Does ITAR Apply to AI?

ITAR Specific Requirements

• Fabrication only in the US — Defense articles must be fabricated, manufactured, or produced in the United States by licensed manufacturers.
• No foreign access without registration — Only US persons (citizens or permanent residents) can access ITAR-controlled technology without specific authorization. This applies even within US facilities.
• State Department jurisdiction — Unlike EAR (BIS, Commerce), ITAR is administered by the State Department's DDTC, which has different processes and timelines for licensing.
• Registration requirement — Anyone trafficking defense articles (including exporting, importing, brokering) must register with DDTC annually and pay fees.
• Licensing complexity — ITAR licenses are harder to obtain, take longer to process, and are more likely to be denied than EAR licenses.

Dual-Use Research and AI

Dual-use research refers to biological, chemical, physical, or computational research that could be used for beneficial civilian purposes but also could be misused to cause harm. The Biden administration has issued guidance on managing dual-use research of concern (DURC), and AI amplifies the dual-use nature of many research projects.

Categories of Dual-Use Research Concern

• Biosecurity — AI models that accelerate protein folding, drug design, or pathogen characterization could enable bioweapons development. Examples include AlphaFold-like models applied to novel pathogens.
• Chemical/biological synthesis — AI systems that predict synthetic pathways for toxins or drugs of abuse could enable chemical/biological weapon production.
• Cybersecurity — AI for vulnerability discovery, automated exploit generation, or AI-powered phishing campaigns could be used for offensive cyber operations.
• Autonomous weapons — AI for target recognition, decision-making, or weapon control raises questions about lethal autonomous weapons systems (LAWS).
• Surveillance technology — AI for facial recognition, behavioral analysis, or mass surveillance raises dual-use concerns even when developed for legitimate civilian purposes.

Institutional Review Requirements

• DURC oversight — Institutions receiving federal funding must establish DURC oversight committees to assess research for potential dual-use concerns.
• IRB review — Research involving human subjects requires Institutional Review Board (IRB) approval. AI systems that process personal data may create new privacy risks that IRBs must evaluate.
• IACUC review — Research involving animals requires IACUC (Institutional Animal Care and Use Committee) approval. AI training datasets derived from animal studies may raise ethical concerns.
• Biological Safety Committee (IBC) — Recombinant or synthetic nucleic acid research requires IBC review, which increasingly includes AI-driven genome editing projects.

National Science Foundation, NIH, and DoD AI Requirements

Federal funding agencies have increasingly specific requirements for AI-related research, extending beyond standard human subjects and data protection protocols.

NSF Requirements

NIH Requirements

• HIPAA compliance — Research using health data must comply with HIPAA. AI systems processing PHI require safeguards and potentially BAAs.
• Genomic data restrictions — NIH Genomic Data Sharing policy governs how genomic data can be used in AI training. Some datasets have tiered access levels.
• Recruitment plans — AI research involving diverse populations must include plans for representative recruitment and bias mitigation.

DOD and DARPA Requirements

• Responsible AI principles — DoDI 3000.09 and DARPA AI ethics framework emphasize human control, accountability, and transparency in AI systems.
• Security certifications — AI systems supporting DOD research may require FedRAMP, CMMC, or other security certifications.
• Export control screening — All project participants must be screened for export control compliance before granting access to research.

Export Control Classification for AI Systems

Before deploying any AI system in a research environment, you must determine whether it requires export control review. The classification process involves identifying the technology, determining applicable controls, and documenting the analysis.

Common AI-Related ECCNs

• 5A992 — Encryption equipment, software, and technology. Applies to AI systems with cryptographic capabilities.
• 5D002 — Information security (cybersecurity) software and technology.
• 3A090 — Advanced computing equipment including specialized high-performance computers.
• 4A101 — Information processing equipment capable of encrypted data processing.
• 5E002 — Software for development of ECCN 5A002 or 5D002 items.