Unleash the Power of AI for Export Classification Determination…or Maybe Not Yet
With the January 2025 announcement of China’s generative AI tool, DeepSeek, and the attention that it brought to the AI industry in the US, the engineer in me got curious and wanted to run an experiment. I wanted to answer this question:
Can AI tools accurately determine the export classification of software?
Before we dive into this little AI experiment, let’s get grounded on what we are talking about. First, let’s briefly talk about the two main types of AI: generative AI and traditional AI. These three quotes from a Forbes article by Bernard Marr succinctly explain the difference.
“[T]raditional AI – it’s like a master strategist who can make smart decisions within a specific set of rules…These AIs have been trained to follow specific rules, do a particular job, and do it well, but they don’t create anything new.”
“Generative AI models are trained on a set of data and learn the underlying patterns to generate new data that mirrors the training set.”
“Traditional AI systems are primarily used to analyze data and make predictions, while generative AI goes a step further by creating new data similar to its training data.”
So Open AI’s ChatGPT, Microsoft’s CoPilot, and High-Flyer’s DeepSeek are all generative AI models. How well can these generative AI tools determine the export controls classification number (ECCN) of an item? To test this, I crafted a simple experiment using ChatGPT and CoPilot.
The “Item”
Let’s first say we have software that falls under export control classification number (ECCN) 3D001 associated with hardware described by ECCN 3A090.a. (Check out the EAR’s Commerce Control List in 15 CFR Part 774 Supplement No. 1 for the full ECCN descriptions.)
For this experiment, we are going to work backwards. We are starting with a specific software ECCN–and the “hardware” ECCN with which its associated–in order to craft input for these AI tools. Then we’ll see how well the AI performs.
ECCN 3D001
“Software” “specially designed” for the “development” or “production” of commodities controlled by 3A001.b to 3A002.h, 3A090, or 3B (except 3B001.a.4, c, d, f.1, f.5, k to n, p.2, p.4, r, 3B002.c, 3B903, 3B904, 3B991, 3B992, 3B993, or 3B994).
ECCN 3A090.a
Integrated circuits as follows (see List of Items Controlled)….
a. Integrated circuits having one or more digital processing units having
either of the following:
a.1. A ‘total processing performance’ of 4800 or more; or
a.2. A ‘total processing performance’ of 1600 or more and a ‘performance density’ of 5.92 or more.
The Input
Next, I wrote a description of my hypothetical software that was written against the specifications of 3D001 for 3A090.a items.
We crafted two versions–one detailed description using the technical specifications within the ECCNs and one broad description. For the sake of simplicity, we’ve intentionally left out any triggers for the term “specially designed” and the implications of its complex “catch” and “release” definition.
Detailed Description
I have software that is for developing integrated circuits. The integrated circuits include a digital processing unit with a total processing performance of 4800. Do I need to be concerned about US export controls? If so, which ECCNs might control the software?
Broad Description
I have software related to integrated circuits. Do I need to be concerned about US export controls?
How do each of these AI platforms respond to each of these prompts?
ChatGPT Output Given Detailed Input
Again, below is what we asked ChatGPT.
ChatGPT gave a decent answer. The response starts with the following: “Yes, you may need to be concerned about U.S. export controls…”
It then directed me to five ECCNs that might be relevant, one of which is indeed 3D001. We’ve shown a portion of the response below.
I also agree with the guidance given on “steps to take.”
Is anything missing?
In terms of classification determination, ChatGPT did not direct me to the sub-ECCN 3A090.a, which can be critical for assessing exactly which reasons for control apply, which license exceptions can be used, and when a Bureau of Industry and Security (BIS) license is required due to the ECCN.
Note: A BIS license might be required for other reasons, such as the recipient being on a restricted party list. And we are intentionally assuming the Department of Commerce jurisdiction and the Export Administration Regulations for this experiment.
EAR Updates in Advanced Computing, Semiconductors, and AI
Grab an infographic that summarizes changes in the EAR, including new ECCNs, license exceptions, foreign direct product rules, and more!
Copilot Output Given Detailed Input
Again, below is what we asked Copilot.
Interestingly, Copilot was a little more adamant…I “should be concerned” about US export controls. Not simply “may need to be concerned” as ChatGPT had advised. But then Copilot pointed me to only two ECCNs–and neither one was 3D001.
In this limited assessment, this advice is not the most comprehensive and is potentially misleading for those that do not investigate further.
ChatGPT Output Given Broad Input
Given ChatGPT almost “passed” my experiment when given a more detailed technical description of my hypothetical software, I wanted to further test its abilities when the technical description was lacking specifics. This latter situation is often the case when an export compliance leader is first starting to triage a potential research endeavor, trying to determine if and how export controls apply. (You know, before that critical step of connecting with the researchers or engineers.)
To do this, I gave ChatGPT the following broad description.
ChatGPT’s response, shown below, validated the notion of garbage in/garbage out. Well, that is probably being too harsh. The reality is that the input given to ChatGPT was broad…so the output was also broad…and not particularly helpful.
Which ECCNs should you focus on? ChatGPT cannot guide you without more technical information. The same goes for human beings who conduct export classification determinations.
Bottom Line
Thus, the conclusion of this simple experiment is that accessible generative AI tools can probably provide good pointers in the right direction. But it’s probably not wise (just yet!) to have them serve as your sole export compliance resource for export jurisdiction or classification determination.
P.S. – The results of running this same experiment with DeepSeek and Alphabet’s Gemini are quite interesting. I’ll let you give that a try if you’re curious!
P.P.S. – Given the rate of change with AI development, results may vary if you run this or a similar experiment on your own!
Leave a Comment
About
Traliance provides export compliance consulting services to help research universities and technology companies comply with U.S export controls regulations.
