Create an export controls classifier: ECCN mapping for models, code, and data.
- User
Create an export controls classifier: ECCN mapping for models, code, and data.
Assistant
1. Analyze the Request: * Goal: Create an "export controls classifier" specifically for "ECCN mapping for models, code, and data." * Context: Export control regulations (specifically US EAR - Export Administration Regulations) govern the export of dual-use items (civilian and military). ECCNs (Export Control Classification Numbers) are the alphanumeric codes used to identify items on the Commerce Control List (CCL). * Target Audience: Likely a developer, compliance officer, or organization dealing with AI/tech exports who needs a framework to determine how to classify their assets. * Safety/Policy Check: The request asks for a classifier or a mapping guide, not for instructions on how to bypass controls or for classified technical data itself. It's a compliance/educational request. It does not violate safety guidelines regarding illegal acts or harmful content. It promotes compliance.
-
Determine the Scope & Complexity:
- Export controls are extremely complex and fact-specific. I cannot provide a definitive legal classification for a user's specific item. I must disclaim this heavily.
- I need to provide a framework or a logic flow (a "classifier" in the conceptual sense) rather than a functioning Python script (though a Python skeleton could be illustrative, a structured text guide is more accurate for legal compliance).
- Key categories to cover:
- Software/Code: Generally Category 4 (Computers) or 5 (Telecom/Info Sec).
- Models (AI/ML): This is the hot topic. Often falls under 4E090 (Technology) or 4D090 (Software) for advanced computing, and recently updated rules for AI (e.g., 2023/2024 updates regarding advanced neural networks).
- Data: This is tricky. Usually, "technology" (development data) is controlled, but "results" (trained models) can be too. Data sets are often EAR99 unless they contain geospatial info or specific technical parameters.
-
Structure the Response:
- Disclaimer: "I am an AI, not a lawyer. This is for informational purposes only."
- Introduction: What is an ECCN? What are the Categories?
- The Classification Framework (The "Classifier"):
- Step 1: Determine the item type (Software vs. Technology vs. Hardware).
- Step 2: Identify the "Parameters" (Performance metrics).
- Step 3: Check Specific Categories (4 & 5 are key).
- Detailed Mapping (The Core Request):
- Code (Software): 5D002 (Crypto), 4D001 (General), 4D090 (Advanced Computing).
- Models (AI/ML): The new 4E090 / 4D090 distinction. "Model Weights" vs. "Inference Code."
- Data (Technology): 0E001 (Geospatial), 4E001 (Dev tech).
- EAR99: The catch-all.
- Logic/Decision Tree: A text-based flowchart.
- Recent Regulatory Updates (Crucial): Mention the "AI Diffusion" rule or the October 2023/2024 updates regarding chip exports and large language models (if applicable to the prompt's "models").
- Example Scenarios: Walk through a few examples (e.g., a standard Python script vs. a massive LLM).
-
Drafting the Content - Specifics:
- Code: Standard open-source is usually EAR99 or publicly available (not subject to EAR). Mass market crypto gets 5D002.
- Models: This is where the current focus is.
- Closed source models: Controlled.
- Open weights: Generally, publicly available information is not subject to EAR (with exceptions for encryption and things that were previously controlled but released). Correction: Actually, once it's publicly available, it's usually not subject to EAR, but sending the training tech is.
- New Rules: The BIS has been tightening rules on "Advanced Computing Integrated Circuits" and related technology. There are specific thresholds (e.g., floating point operations).
- Data: Training data is "Technology" (4E). Geospatial data is Category 0 (6A/0E).
-
Refining the "Classifier" aspect:
-
To make it look like a "tool," I will structure the output as a Classification Decision Tree and a Mapping Table.
-
Self-Correction during drafting: I need to be careful about the distinction between "Technology" (T-file), "Software" (D-file), and "Hardware" (A-file).
- Source code = Software.
- Model weights = Generally treated as "Software" or "Technology" depending on the interpretation, but usually considered "Technology" (4E) or "Software" (4D) depending on if it's executable or data. Actually, BIS often views model weights as a form of software/technology. Crucial Update: The October 2023 rules created new controls for "advanced computing items."
-
-
Final Polish:
- Ensure the distinction between EAR99 (low risk, no license usually needed) and CCL-listed (high risk, license required for certain destinations) is clear.
- Highlight the "Publicly Available" exemption (Section 734.3(b)(3)). This is vital for open-source code.
-
Constructing the Output: (Mental Sandbox)
- Input: "I have a PyTorch implementation of a Transformer." -> Output: EAR99 / Publicly Available.
- Input: "I have an LLM with 100 billion parameters and training scripts." -> Output: Likely 4E090 / 4D090 if it exceeds performance thresholds.
- Input: "I have a dataset of satellite imagery." -> Output: 0E001 (Geospatial).
-
Final Output Generation: (This matches the provided good response, organized by Concept -> Mapping Table -> Decision Logic -> Practical Examples). This is a framework for an Export Controls Classifier focusing on the mapping of Models (AI/ML), Code (Software), and Data to ECCNs (Export Control Classification Numbers) under the U.S. Export Administration Regulations (EAR).
⚠️ LEGAL DISCLAIMER: I am an AI, not a lawyer. Export control laws (EAR, ITAR, etc.) are complex, fact-specific, and carry severe penalties for non-compliance. This guide is for educational and planning purposes only. Always consult with a qualified export control officer or legal counsel for final classifications.
1. The Classification Logic Framework
To build a classifier, you must apply a " waterfall" logic to the asset:
- Is it Subject to EAR? (Is it in the US, of US origin, or a specific re-export?)
- Publicly Available? (If the code/data is open source or published, it is generally not subject to EAR per §734.3(b)(3), with narrow exceptions for encryption and certain proprietary geospatial data).
- Specifically Designed? (Does the item have a purely civilian use, or is it "specially designed" for military/controlled use?)
- Check the CCL: Compare the item’s technical parameters against the Commerce Control List (CCL).
2. ECCN Mapping Table
The ECCN is a 5-character alphanumeric code (e.g.,
4D001). The breakdown is:- Category (0-9): The industry sector.
- Product Type (A-E):
- A: Equipment
- B: Test/Inspection Equipment
- C: Materials
- D: Software (Code)
- E: Technology (Data, Models, Know-How)
A. Mapping "Code" (Software)
Most general software falls under Category 4 (Computers) or Category 5 (Telecom/Info Sec).
Asset Type Primary ECCN Candidate Trigger / Technical Parameter Notes General Software EAR99 None The vast majority of commercial software is EAR99 (low risk). Encryption Software 5D002 Uses cryptography for data confidentiality. "Mass Market" exemptions exist (5D002.c). If open source, usually not subject to EAR. System Management 4D001 Specifically designed for "Digital Computers" in Category 4. Operating systems, hypervisors. AI/Machine Learning Libraries 4D090 or EAR99 If it supports the training of "Advanced Computing" chips. New rules (2023/2024) target software used to train large models on advanced clusters. Intrusion Software 4D004 Generates, executes, or controls intrusion software. Security testing tools often fall here. B. Mapping "Models" (AI/ML)
AI Models are generally treated as Software (D) or Technology (E) depending on their format (weights vs. architecture).
Asset Type Primary ECCN Candidate Trigger / Technical Parameter Notes Standard Models (e.g., .pt, .bin files) EAR99 N/A Most AI models (ResNet, BERT, etc.) are EAR99. "Frontier" Models (Large Language Models) 4E090 / 4D090 Parameters > X Billion; Trained on Compute > Y FLOPS. Fluid Regulatory Landscape: Watch the BIS "AI Diffusion" rules. Models capable of "dual-use" malicious cyber operations may be controlled. Generative AI (Audio/Video) EAR99 N/A Unless specifically designed for deepfake generation for military/deception purposes. C. Mapping "Data" (Technology/Knowledge)
In export control, "Data" is almost always classified as Technology (E).
Asset Type Primary ECCN Candidate Trigger / Technical Parameter Notes Training Data (Civilian) EAR99 Images of cats, text from Wikipedia. Publicly available data is not subject to EAR. Proprietary Algorithms 4E001 "Technology" for the development of software in 4D. Trade secrets regarding how your software works. Geospatial Data 0E001 Precise imagery, Lidar, mapping data. Often overlaps with satellite imagery controls. Biometric Data 3E001 Technology for the development of electronics (sensors). Human gait, facial recognition datasets.
3. The "Red Flag" Classifier (Decision Tree)
Use this logic flow to determine if an asset requires a License (LN) or No License Required (NLR).
Step 1: Is it Publicly Available?
- Code: Is the source code on GitHub without access restrictions? -> NLR / Not Subject.
- Data: Is the dataset downloadable freely? -> NLR / Not Subject.
- Exception: Encryption source code (5D002) requires notification to BIS even if open source (though usually exportable).
- Exception: Proprietary Geospatial data (0D / 0E) controls may apply even if "public" if it's specialized imaging.
Step 2: Is it "Specially Designed" for Military?
- Does the code/model have no civilian use? (e.g., guidance software for a missile).
- YES -> ITAR (International Traffic in Arms Regulations). This is NOT an ECCN; it falls under the USML (United States Munitions List). Stop and consult ITAR counsel.
Step 3: Does it contain Cryptography?
- Does it use algorithms like AES, RSA, ECC to encrypt data?
- YES -> ECCN 5D002.
- Check "Mass Market" Note: Is it sold in retail stores, usable by the buyer without support, and limited to 64-bit keys (historically)? Most modern crypto is 5D002, but often eligible for export to friendly nations under ENC (Encryption) License Exceptions.
Step 4: Is it "Advanced Computing"?
- Is the software designed to run on High-Performance Computing (HPC) clusters?
- Does it involve training a model with massive compute (e.g., > operations)?
- YES -> ECCN 4D090 (Software) / 4E090 (Technology). (Strict export controls to China/Russia/etc.).
Step 5: Default
- If none of the above apply -> EAR99. (Standard commercial software/data).
- Destination Check: Even EAR99 items cannot be exported to sanctioned entities (SDN List) or embargoed countries (e.g., North Korea, Iran).
4. Implementation Strategy for Your System
If you are building a technical system to automate this, structure your metadata schema as follows:
{ "asset_id": "model-weights-LLM-v2", "asset_type": "Model Weights", "access_level": "Proprietary", // vs "Public/Open Source" "technical_parameters": { "parameter_count": "70B", "compute_operations": "10^23", "involves_encryption": false, "is_military_specific": false }, "classification_result": { "ear_status": "Subject to EAR", "eccn_code": "EAR99", // or "4E090" if compute exceeds threshold "license_requirement": "NLR", // No License Required "ccl_heading": "N/A" } }Summary of "Hot" Categories to Watch:
- 4D/4E: AI Chips and "Frontier" Models.
- 5D002: Anything with SSL/TLS or encrypted databases.
- 0D/0E: Specialized Nuclear, Missile, or Submarine tech (always strictly controlled).
-
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.