U.S. government limits exports of artificial intelligence software

When the U.S. government limits exports of artificial intelligence software, it's not just a headline for defense contractors. It's a new layer of compliance that can trip up SaaS founders, AI researchers, and even marketing teams using geospatial tools. The short version: the Bureau of Industry and Security (BIS) now controls the export of certain AI software designed to extract objects from satellite imagery and geospatial data. But the ripple effects go deeper than that.

I've spent the last few weeks digging into the Federal Register notices and talking to a couple of export compliance lawyers. Most of the coverage misses the practical stuff. So let's fix that.

What Exactly Is Being Restricted?

The rule, which took effect in January 2020, targets "geospatial imagery software specially designed for training a Deep Convolutional Neural Network to automate the analysis of geospatial imagery and point clouds." That's a mouthful. Here's what it actually means.

Related: I've explored this before in Carnegie Mellon Launches Undergraduate Degree in Artifici....

If your software can take a satellite photo and automatically identify buildings, vehicles, or terrain features using deep learning, it might need an export license. The key word is "specially designed." A general-purpose AI framework like PyTorch isn't covered. But a specialized application built on top of it that does automated feature extraction from satellite data? That's squarely in scope.

The BIS didn't pull this out of thin air. According to their rulemaking document, the concern is that this technology could give adversaries military-grade intelligence capabilities — identifying troop movements, missile sites, or naval deployments from commercially available satellite imagery. It's a classic dual-use problem. The same tool that helps a logistics company optimize shipping routes can help a military plan a strike.

Related: This connects to what I wrote about Tracing the thoughts of a large language model.

3 Scenarios Where This Actually Matters

Most people read "export controls" and think it only applies to shipping physical products to China. Wrong. Here are three scenarios I've seen trip people up.

1. You're a SaaS Founder with Overseas Customers

Say you've built a platform that processes drone imagery for agricultural monitoring. Farmers upload photos, your AI identifies crop stress patterns, and you charge a subscription fee. You've got customers in Brazil, India, and Germany.

Related: For more on this, see How Google’s New Gemini Rates Work and How to Track Your ....

Here's the catch: providing access to your software to a foreign person can be considered an "export" — even if your servers are in Ohio. It's called a deemed export. If your AI model was trained on geospatial data and does automated feature extraction, you might need a license for customers in certain countries. India and Brazil are generally fine. But if a researcher in China or Russia accesses your tool, you've got a problem.

I've seen this happen. A drone analytics startup I advised last year had to scramble when their compliance team flagged that their "automated building detection" feature fell under the new controls. They ended up geofencing their API and adding export control language to their terms of service. It wasn't a business-killer, but it was a headache they didn't see coming.

2. You're an AI Researcher Collaborating Internationally

Academic collaboration is where this gets messy. The rule includes an exception for "fundamental research" — work that's published and publicly available. But the line between fundamental research and controlled technology is blurry.

If you're a PhD student at Stanford working on a novel CNN architecture for satellite imagery analysis, and you publish your code on GitHub, you're probably fine. But if you're working under a DARPA grant with publication restrictions, and you share your model weights with a collaborator at Tsinghua University, you might need a license. Even if the collaboration is purely academic.

The BIS has been clear: the fundamental research exclusion doesn't apply if there are access controls or publication restrictions. I've talked to researchers who are now adding export control clauses to their collaboration agreements. It's tedious, but it's the new reality.

3. You're Using Third-Party AI Tools That Incorporate Geospatial Data

This is the scenario most people overlook. You might not be building geospatial AI software. But what if you're using a third-party tool that does?

Consider a marketing team using an AI-powered location intelligence platform to analyze foot traffic patterns for retail site selection. The platform uses satellite imagery and automated feature extraction. If that platform is hosted in the U.S. but accessed by employees in a restricted country, the company providing the platform — and potentially the company using it — could have export control obligations.

I'm not saying every marketing team needs an export compliance officer. But if your business operates globally and uses AI tools that touch geospatial data, it's worth asking your vendors about their export control posture. A quick question now beats a nasty surprise later.

Why This Rule Is Just the Beginning

The geospatial imagery rule is narrow. But the thinking behind it is expanding. In October 2022, the BIS implemented sweeping export controls on advanced computing chips and semiconductor manufacturing equipment. In 2023 and 2024, those controls were tightened further, targeting chips like the NVIDIA A100 and H100 that are used to train large AI models.

The pattern is clear. The U.S. government is increasingly treating advanced AI as a strategic technology that needs to be protected, not just from military adversaries but from economic competitors. The geospatial imagery rule was an early test case. The semiconductor controls are the main event. And I'd bet good money that controls on large language models — or at least the weights of models trained on massive compute clusters — are being discussed right now.

According to a 2024 report from the Center for Strategic and International Studies, the U.S. is moving toward a framework where AI capabilities are assessed not just by what the software does, but by the compute resources used to train it. That's a significant shift. It means a model trained on 10,000 GPUs might face export restrictions even if it's just a general-purpose language model, simply because of the scale of compute involved.

What You Should Actually Do About This

I'm not a lawyer, and this isn't legal advice. But here's the practical framework I use when advising clients.

First, do a quick triage. Does your product or research involve (a) satellite or aerial imagery, (b) automated feature extraction using deep learning, and (c) foreign users or collaborators? If you answered yes to all three, talk to an export compliance attorney. Seriously. The penalties for willful violations can include millions in fines and prison time.

Second, even if you're not in the geospatial space, pay attention to the broader trend. If you're building AI tools that could have dual-use applications — and honestly, that's a lot of AI tools — start thinking about export controls as part of your compliance roadmap. It's easier to build controls in from the start than to retrofit them later.

Third, document your classification decisions. If you determine that your software isn't covered by the rule, write down why. The BIS looks more favorably on companies that make good-faith compliance efforts, even if they get it wrong, than on companies that never bothered to check.

This is where tools like AI-Mind become relevant in an unexpected way. When you're drafting internal compliance documentation, explaining technical classifications to non-technical stakeholders, or creating training materials for your team on export control basics, you need clear, accurate content fast. AI-Mind's approach — where you describe what you need and pick a content type rather than wrestling with prompts — can save hours on documentation that nobody wants to write anyway. The first 30 generations are free, which is enough to draft a solid compliance memo and a team training doc. Not a bad use of a free tier.

Key Takeaways

• The U.S. restricts export of AI software that automates satellite imagery analysis using deep learning, including deemed exports via cloud access.
• SaaS platforms with overseas users, international research collaborations, and third-party AI tool usage are the three most common risk scenarios.
• Export controls on AI are expanding beyond geospatial software to include advanced chips and, likely, large AI model weights.
• Document your classification decisions even if you conclude the rule doesn't apply — good-faith efforts matter to regulators.
• Compliance documentation doesn't have to be painful; AI content tools can speed up internal memos and training materials significantly.

Sources

Bureau of Industry and Security, Addition of Software Specially Designed to Automate the Analysis of Geospatial Imagery to the Export Control Classification Number, 2020. The official Federal Register notice establishing the geospatial AI software export controls.

Center for Strategic and International Studies, The Evolution of U.S. AI Export Controls, 2024. Analysis of the expanding scope of U.S. export restrictions on artificial intelligence technologies.

Bureau of Industry and Security, Implementation of Additional Export Controls on Advanced Computing and Semiconductor Manufacturing Items, 2022. The October 2022 rule that expanded controls to advanced AI training chips.