Is ChatGPT ITAR compliant? What engineers can paste into AI

The Morse teamPublished November 4, 202610 min read

No, ChatGPT is not ITAR compliant, and no AI tool is on its own: there is no ITAR compliance certification, and the duty to comply sits with whoever does the exporting, which means you and your employer. Microsoft's export controls page puts it in one line, "There's no ITAR compliance certification" (Microsoft). Whether pasting technical data into a chatbot counts as an export is a separate question, and for a consumer AI app the regulations' encryption safe harbor doesn't help, because the chatbot has to decrypt your text to answer it.

On November 4, 2026, we read the ITAR's definitions in 22 CFR Part 120, the EAR's in 15 CFR Part 734, the two Federal Register rules that created their encryption carve-outs, the Bureau of Industry and Security's (BIS) deemed export page, and the export control pages of Microsoft Azure, Anthropic, and Google Cloud. OpenAI's help center and its ChatGPT Gov page refused our requests that day, so for OpenAI we cite only the developer documentation we could load. This isn't legal advice; ask your export compliance office.

Can engineers use ChatGPT at work?

Yes, for most of what an engineer asks, and your employer's AI policy decides the rest. The ITAR's definition of technical data leaves out "information concerning general scientific, mathematical, or engineering principles commonly taught in schools, colleges, and universities," information in the public domain, and "basic marketing information on function or purpose or general system descriptions of defense articles" (22 CFR 120.33). On the Commerce side, BIS says "research conducted using publicly available information is also exempt from any license requirements" (BIS). A question about beam deflection, a Python traceback, or a paper you're reading usually sits on that side of the line.

Your employer's rules are a separate layer, and they can be stricter than the law. On a work account or a managed laptop, your company may also see what you type, as we lay out in can my employer see my ChatGPT.

What counts as ITAR technical data?

Information "required for the design, development, production, manufacture, assembly, operation, repair, testing, maintenance, or modification of defense articles," in the regulation's words, including "blueprints, drawings, photographs, plans, instructions, or documentation" (22 CFR 120.33). The definition also takes in classified information about defense articles, information under an invention secrecy order, and software "directly related to defense articles."

Defense articles are the items on the U.S. Munitions List. Most other controlled technology, from a commercial sensor to a machine tool, falls under the EAR, which covers commercial and dual-use items. Which list a drawing or a file belongs to is a classification your export compliance office makes. It isn't something to settle from inside a chat window.

Is pasting technical data into ChatGPT an export?

It can be, depending on where the text goes and who can read it. Under the ITAR, an export includes "an actual shipment or transmission out of the United States" and "releasing or otherwise transferring technical data to a foreign person in the United States (a deemed export)" (22 CFR 120.50). A release inside the US counts as an export "to all countries in which the foreign person has held or holds citizenship or holds permanent residency."

The EAR's definition runs parallel and names source code: releasing "technology" or source code "(but not object code) to a foreign person in the United States" is a deemed export (15 CFR 734.13). Both turn on facts you can't see from a chat box: where the provider stores and processes your text, and which of its staff and contractors can read it.

The responsibility for them is yours. Microsoft sums up the Commerce Department's position this way: "when data or software is uploaded to the cloud, the customer, not the cloud provider, is the exporter" (Microsoft). BIS lists three advisory opinions on cloud services, dated January 13, 2009, January 11, 2011, and November 24, 2014 (BIS).

Does the ITAR encryption carve-out cover AI chatbots?

No, not a chatbot that has to read your text to answer it. Since March 25, 2020, the ITAR has said that sending, taking, or storing technical data is not an export if it is unclassified, "secured using end-to-end encryption," protected by FIPS 140-2 modules or encryption at least as strong as AES-128, and not intentionally sent to or stored in a country proscribed in 22 CFR 126.1 (22 CFR 120.54). The EAR has had a matching carve-out since September 1, 2016 (15 CFR 734.18).

Both rules define end-to-end encryption the same way. The data is "not in an unencrypted form" between the originator and the intended recipient, or their in-country security boundaries, and "the means of decryption are not provided to any third party." The ITAR adds that the intended recipient "must be the originator, a U.S. person in the United States, or a person otherwise authorized to receive the technical data."

A chatbot can't meet that definition. HTTPS protects your prompt on the way to the provider, and then the provider decrypts it so the model can read it. The State Department weighed a milder case when it wrote the rule. Commenters asked it to allow data to be "clear text in route to the intended recipient, for processing via applications, such as anti-virus software or spell-check." It refused: the carve-out "requires that the technical data subject to the ITAR be continuously encrypted at all times while outside of an authorized security boundary" (84 FR 70887).

BIS gave the same reason for its version in 2016. The requirement exists "to prevent unauthorized access to data in clear text by parties other than the originator (or the originator's company or organization) and the recipient, such as external service providers," and decryption is allowed only where "no decryption takes place outside of the security boundaries of the originator and the recipient" (81 FR 35586). A language model reading a whole drawing note on a provider's server is that kind of processing, done by an external service provider.

Missing the carve-out doesn't make a paste a violation by itself. It means the ordinary test applies: did the text leave the country, and could a foreign person read it? The ITAR also counts as a release "the use of access information to cause technical data outside of the United States to be in unencrypted form" (22 CFR 120.56).

Neither rule mentions hardware enclaves, the sealed processors some AI services run models in. We found nothing in the rule text or in the two Federal Register rules behind it that treats a third party's enclave as part of your security boundary, so don't assume an enclave-based AI tool meets either carve-out.

Which AI tools offer an ITAR environment?

Cloud deployments that your employer buys and configures. As of November 4, 2026:

• Microsoft lists gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna for Azure OpenAI in Azure Government, in the usgovarizona and usgovvirginia regions (Microsoft). Azure Government limits "potential access to systems processing your data to screened US persons," and Microsoft adds that "you're responsible for designing your applications to apply end-to-end data encryption that meets ITAR requirements" (Microsoft).
• Anthropic says: "For ITAR-controlled data, use Claude through Amazon Bedrock in AWS GovCloud (US), which supports DoD IL4/IL5 workloads." The same page says Claude Enterprise "isn't designed for export-controlled data" (Anthropic).
• Google Cloud's ITAR control package lists Generative AI on Vertex AI behind a US regional endpoint, says "Gemini in BigQuery is not supported," and routes ITAR support cases "to US Persons located in the US." It adds: "If a product is not listed, that product is unsupported and has not met the control requirements for ITAR" (Google Cloud).
• OpenAI's API data controls page lists ten data residency regions, the United States among them, and doesn't mention ITAR or export controls (OpenAI). We couldn't load its ChatGPT Gov page on November 4, 2026, so we don't describe that product here.

Anthropic and Microsoft deserve credit for saying in plain words which product to use and which not to. Each of these is a contract your company signs and sets up, with your export compliance office deciding what goes in. None of them is the personal ChatGPT, Claude, or Gemini app on your phone.

What engineers paste into AI, and which rule could apply

The middle column says what could apply, not what does. Classification is your compliance office's call.

Do NDAs and trade secret law limit what you paste?

They can, even when no export rule applies. Under the Defend Trade Secrets Act, information is a trade secret only if its owner "has taken reasonable measures to keep such information secret," and the statute's list covers "engineering information" and "programs, or codes" (18 U.S.C. 1839). Pasting your company's code or designs into a service that may keep them is something the other side in a dispute could point to, and whether it would matter is a question for a lawyer. An NDA can restrict where a supplier's or customer's information goes, and our guide to private AI for startups covers how those clauses meet consumer AI terms. For coding assistants, see whether GitHub Copilot trains on your code.

How Morse keeps an engineer's everyday questions off company servers

Every rule above asks the same thing: who besides you can read the text, and where. For controlled technical data, that question belongs to your export compliance office and the tool it approves. Morse makes no ITAR, EAR, or other compliance claim, and its E2EE mode decrypts inside an enclave run by a third-party provider, so don't treat any Morse mode as meeting either encryption carve-out. For the rest of an engineer's day, the code question, the paper, and the design doc with names and specs taken out, Morse is built so your chat history never sits on its servers. You pick a privacy mode for each chat:

• E2EE, for scrubbed code and drafts you'd rather no server read. Pick an open-weight model like GLM, Kimi, or DeepSeek. Your browser checks the secure enclave's hardware attestation, then encrypts the message so only that enclave can read it, and the reply is decrypted only in your browser. Morse's servers relay ciphertext they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed, which our guide to verifying an AI enclave's attestation walks through. E2EE has no web search, and an E2EE chat can only continue in E2EE. E2EE chats come with Pro, at $19 a month.
• Private, for the day's ordinary engineering questions. It's the default for a new chat, with open-weight models like DeepSeek, Kimi, and GLM, plus Grok. Morse has zero data retention agreements with its inference, search, and other providers, so they keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
• Anonymous, to ask GPT, Claude, or Gemini a textbook question without the provider knowing it's you. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider may keep the message for a limited time to check for abuse, under its own policy, so leave your employer, program, and part numbers out of it.

In every mode, your chat history is stored in your browser, encrypted, and unlocking it takes two keys: one that never leaves your device and one Morse's server gives you when you sign in. Our servers don't keep your messages or log what you write. Morse doesn't train AI on your content, sell your data, or show ads (privacy policy), and the Free plan includes coding, file, and other chat tools. Clearing your browser or deleting your account deletes your chats, and Morse can't recover them, so copy the fix you found into your own notes.

Ask your next textbook question from the Morse home page without an account, or start free and move to Pro when you want E2EE.