ITAR/EAR export controls and AI models

U.S. export-control pages do not classify “AI” as a single license category. DDTC administers the ITAR for defense articles and defense services on the U.S. Munitions List. BIS administers the EAR, including a control on model weights of certain advanced closed-weight dual-use AI models under ECCN 4E091. This page restates those official sentences. It is not a classification or a license opinion.

Status as of 14 October 2026: 15 CFR 740.27 is still License Exception Artificial Intelligence Authorization (AIA) and still names AI Model Weights specified by ECCN 4E091. eCFR still displays title 15, up to date as of 10/09/2026 and Title 15 was last amended 9/24/2026. We found no newer public rewrite of 4E091 on that page.

ITAR is about the USML, not “AI” as a heading

22 CFR 120.31 defines a defense article as any item or technical data designated in § 121.1 — the U.S. Munitions List. The definition does not include basic marketing information on function or purpose or general system descriptions.

Technical data includes information … required for the design, development, production, manufacture, assembly, operation, repair, testing, maintenance, or modification of defense articles and, separately, software (see § 120.40(g)) directly related to defense articles.

Software includes but is not limited to the system functional design, logic flow, algorithms, application programs, operating systems, and support software. A person who intends to export only software should, unless it is specifically enumerated in § 121.1, apply for a license pursuant to part 125.

Those sentences reach an AI model when the software or technical data is directly related to defense articles or is itself designated on the USML. They do not, on the pages we cite, say that every trained model is ITAR-controlled.

Technical data does not include information concerning general scientific, mathematical, or engineering principles commonly taught in schools or information in the public domain as defined in § 120.34. Public domain means information published and which is generally accessible or available to the public through the listed channels, including fundamental research as that section defines it. Publishing weights on a hub is not, by itself, the ITAR public-domain test; the regulation lists the channels.

DDTC is the State Department office that administers those rules. If jurisdiction is unclear after reading the USML and part 120, the official path is a commodity-jurisdiction request to DDTC. This site does not file or decide those requests.

EAR: closed-weight model weights under ECCN 4E091

BIS published the Framework for Artificial Intelligence Diffusion as an interim final rule on 15 January 2025 (effective 13 January 2025). The rule adds a new control on artificial intelligence (AI) model weights for certain advanced closed-weight dual-use AI models under newly created ECCN 4E091.

BIS says it is requiring a license to export, reexport, or transfer (in-country) the model weights of any closed-weight AI model—i.e., a model with weights that are not published—that has been trained on more than 10^26 computational operations. It also says it is not imposing controls on the model weights of open-weight models.

The same rule adds License Exception Artificial Intelligence Authorization (AIA), now at 15 CFR 740.27. That exception authorizes the export, reexport, and transfer (in-country) of listed items, including AI Model Weights specified by ECCN 4E091, to entities in destinations listed in paragraph (a) of supplement no. 5, subject to the conditions in that section. It cannot be used to provide the listed compute items for training AI models specified in ECCN 4E091 to entities outside those destinations.

15 CFR 734.9(l) is the AI Model weights FDP rule. A foreign-produced item specified in ECCN 4E091 can be subject to the EAR when produced by a plant or major component that is itself subject to the EAR and specified in the listed ECCNs. Note 8 says a foreign-produced 4E091 item includes any foreign produced item that is further trained or modified via post-training techniques such as fine-tuning, quantization, or other techniques. The destination scope is any location worldwide.

Those are the official EAR sentences we found for model weights of a defined class of closed-weight models. They are not a license requirement for every AI SaaS chat, every fine-tune below the stated compute threshold, or every open-weight release. Check the live eCFR and BIS pages; this article is a snapshot of the pages cited on 14 October 2026.

What we searched for and did not find

We searched the ITAR part 120 text we fetched and the AI Diffusion rule for a sentence that says “all AI models are ITAR” or “all model weights require a BIS license.” We did not find those sentences. We also did not find, on those BIS and DDTC pages, a published table that maps every popular foundation model to an ECCN or a USML category.

Analysis, not a classification: a defense-article-related model (guidance, targeting, or other USML function) is an ITAR question. A closed-weight dual-use model above the 4E091 training-compute line is an EAR question. A generic commercial chatbot is not described by either sentence until someone classifies the item. That stacking is a reading of the two regimes, not a license.

This is general information, not legal advice. Read the ITAR and the BIS rule for your own analysis.