Commerce’s 2018 AI list made export control a staffing question

In November 2018, Commerce asked which emerging technologies, from neural networks to speech recognition, needed export controls. For banks, hospitals and agencies, exposure ran through people and access: who builds, where work happens, what gets published and what suppliers pass on.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

In November 2018, Commerce asked which emerging technologies, from neural networks to speech recognition, needed export controls. For banks, hospitals and agencies, exposure ran through people and access: who builds, where work happens, what gets published and what suppliers pass on.

  1. The week On Monday, November 19, 2018, the Commerce Department’s Bureau of Industry and Security (BIS) published an advance notice of proposed rulemaking in the Federal Register.
  2. The people who build: deemed exports on the payroll The EAR define an export to include “releasing or otherwise transferring ‘technology’ or source code (but not object code) to a foreign person in the United States.”
  3. The places work happens: offshore teams and remote access The second place the notice could land was in how organizations distribute their work.
  4. The line between research and product The third place the notice could land was the boundary between research and proprietary development.
  5. The suppliers: controls that arrive through what you buy The fourth place the notice could land was in the supply chain.
  6. Objections: too broad to matter, or too narrow to bite The strongest objection is that the notice’s list was so broad that it could never have become law in that form.
  7. With hindsight The broad list of 2018 did not become a broad set of controls, at least in the actions reviewed here.

The week

On Monday, November 19, 2018, the Commerce Department’s Bureau of Industry and Security (BIS) published an advance notice of proposed rulemaking in the Federal Register. Its title was “Review of Controls for Certain Emerging Technologies.” The notice was dated November 14 and asked for comments by December 19, 2018. The Office of Management and Budget had determined it to be significant under Executive Order 12866.

The notice rested on a new statute. As part of the National Defense Authorization Act for Fiscal Year 2019, Public Law 115-232, Congress had enacted the Export Control Reform Act of 2018. Section 1758 of that act authorized Commerce to establish controls, including interim controls, on the export, reexport or in-country transfer of emerging and foundational technologies. Under the act, those were technologies essential to U.S. national security and not already described in specified provisions of the Defense Production Act of 1950.

The act set a process. An interagency process would identify the technologies, drawing on public and classified information, the Emerging Technology Technical Advisory Committee and the Committee on Foreign Investment in the United States. It had to consider development abroad, the effect controls might have on development in the United States, and whether controls would actually limit proliferation. Once a technology was identified, Commerce had discretion over the level of control. At a minimum, it had to require a license for exports to countries under a U.S. embargo, including an arms embargo.

The notice sought comment on how to define emerging technology and what criteria to apply. It offered a representative list of fourteen categories. Several were familiar from defense: hypersonics, robotics, position, navigation and timing, and advanced materials. Others described tools in daily use far from defense. The artificial intelligence category named neural networks and deep learning, computer vision, expert systems such as decision support systems, speech and audio processing, natural language processing, AI cloud technologies and AI chipsets. A data analytics category named visualization, automated analysis algorithms and context-aware computing. Biotechnology named genomic and genetic engineering and neurotech. Advanced surveillance named faceprint and voiceprint technologies. Brain-computer interfaces had their own entry, and a quantum category named quantum computing, quantum encryption and quantum sensing.

BIS explained that these categories were currently subject to the Export Administration Regulations (EAR) but controlled only to embargoed countries, countries designated as supporters of terrorism, and restricted end uses or end users. The question was whether specific emerging technologies within them were important to national security and could be controlled effectively. BIS said it did not seek to expand jurisdiction over technologies not currently subject to the EAR, such as fundamental research. It did not seek to alter controls on technology already described on the Commerce Control List. It said a separate notice on foundational technologies would follow. It also asked about the state of development in the United States and abroad. It asked what effect specific controls would have on U.S. technological leadership, and at what stage of maturity a technology should be considered for control.

The interagency process, BIS said, was expected to lead to proposed rules for new Export Control Classification Numbers. For any organization with data scientists, speech analytics, imaging research or a decision support tool, the notice raised a question that some may never have asked: whether they were, in the legal sense, exporters.

Commerce’s November 2018 notice named technologies with obvious uses far from defense. Natural language processing, speech recognition, computer vision, decision support, data analytics and genomics were all on the list. Technology leaders in those sectors may have treated export control as a matter for shipping departments and defense contractors. The notice suggested it might become their problem too.

The way it would become their problem is the subject of this article. For a regulated organization that does not ship hardware abroad, export control reaches through people and access, not freight. The EAR treat releasing controlled technology or source code to a foreign person in the United States as an export. They treat a transmission out of the country, in any manner, as an export. They exclude technology that is published or that arises from fundamental research intended to be published. And they reach organizations through the products and services their suppliers are allowed to sell them.

Those mechanisms give four separate places where the notice could have landed in a bank, a hospital or an agency. The first is the people who build: who on the staff may see controlled technology. The second is the places work happens: offshore teams, foreign affiliates and remote access. The third is the line between research and product: what an organization publishes and what it keeps proprietary. The fourth is the suppliers: what controls on chips, models and cloud services pass down to their customers. A final section takes the strongest objections.

The four do not overlap. The first is a hiring and access question owned by human resources and security. The second is an operating model question owned by technology and sourcing. The third is a governance question owned by research leaders and counsel. The fourth is a procurement question owned by vendor management. As the hindsight section shows, the controls that followed were narrower than the 2018 list. The later rules reviewed here carved deemed exports out of their new license requirements, a sign that regulators saw the workforce questions as ones to limit with care.

The people who build: deemed exports on the payroll

The EAR define an export to include “releasing or otherwise transferring ‘technology’ or source code (but not object code) to a foreign person in the United States.” The EAR call this a deemed export. The regulations add that a release in the United States is deemed an export to the foreign person’s most recent country of citizenship or permanent residency.

The definition of foreign person matters as much as the definition of export. Under the EAR, a foreign person is any natural person who is not a U.S. citizen, a lawful permanent resident, or another protected individual under the immigration statute. A green card holder is therefore not a foreign person for this purpose. A data scientist working in the United States on a temporary visa generally is.

Why the 2018 list made this a staffing question

BIS described the notice’s categories as, at the time, controlled only to embargoed countries, supporters of terrorism, and restricted end uses or end users. For a typical employer, releasing that technology to most foreign nationals on staff required no license. If an interagency process moved a specific AI or analytics technology into a new, broader classification, that could change. A release to an employee from a country newly covered by a license requirement could require a license before the employee could see the code.

Where a regulated organization employs technologists on temporary visas, the notice’s categories could overlap with the work they do. A bank’s fraud analytics team, a health system’s imaging research group and an agency’s data science unit could all find that some of what they build fell within a category. The practical question concerned the organization rather than the trustworthiness of its staff: whether it knew which technology it held, how it would be classified, and who had access to it.

What an employer would actually have to do

Managing deemed exports is an access control program with a legal trigger. An organization needs a classification for the technology and source code it develops or receives. It needs to know the citizenship and permanent residency status of the people who may access controlled items, collected lawfully and handled with care. It needs access controls that can restrict a repository or a document set to people cleared for it. And it needs a process to apply for a license when one is required, with lead time built into hiring and project plans.

Each step has a cost beyond paperwork. Classification needs people who understand both the technology and the regulations, a rare combination in a bank or hospital. Collecting citizenship information for access decisions must be done consistently with employment law, which is a matter for counsel. Restricting repositories fragments engineering teams and slows work. License lead times can delay a hire’s start on a project. Each cost is modest for a single project. Together, across an analytics program, they change how teams are organized.

Object code, source code and the tools teams use

The deemed export rule covers technology and source code, but not object code. That distinction maps onto how regulated organizations use AI. A team that runs a vendor’s compiled product, or calls a service through an interface, generally handles object code or outputs. A team that writes or modifies models, training code and the know-how to build them handles source code and technology.

The difference matters for planning. An organization that mainly uses AI built by others has a narrow exposure. Its exposure under a deemed export rule would concentrate in the smaller groups that develop models in house, such as research units, quantitative teams and innovation labs. A leader who maps where source code and development know-how actually live can scope the problem to those groups instead of the whole workforce.

A hypothetical fraud analytics team

Consider, as an illustration, a regional bank with a fraud analytics team of a dozen people. Some are citizens, some are permanent residents and some work on temporary visas. The team writes its own transaction-scoring models and keeps the code in a shared repository. A vendor supplies the platform on which the models run.

Under the rules in force in 2018, nothing in that setup was likely to require a license, because the technology was controlled only to embargoed destinations and restricted parties. Now assume a later rule had placed a technique the team used in a new classification. The bank would have needed four answers fast. It would need to know which code embodies the technique and who has read access. It would need to know which access holders are foreign persons, and from which countries. And it would need to know whether a license exception applies. A bank that already kept a repository inventory and role-based access could answer in days. One that could not would have had to freeze access while it found out.

The illustration shows why the cost of preparation is mostly the cost of good engineering hygiene. Knowing what each repository contains, who can read it and why, is something a bank should know for security and model risk reasons anyway. Export control adds one more attribute to an inventory that should already exist.

The roles who own the answer

In an exporting manufacturer, a trade compliance office may own these questions. Where a bank, health system or agency has no such office, the questions fall between four functions. Human resources holds immigration status. Information security runs access control. Legal reads the regulations. Technology leaders know what is being built. The notice was a prompt to decide which of them owns classification and who signs off on access.

A workable assignment gives legal the interpretation, technology leaders the inventory of what is built, security the enforcement of access, and human resources the lawful handling of status information. One named person should coordinate. Without that person, each function can reasonably believe another is handling it.

How the exposure differs by organization

The differences by type of organization are real. A large bank with a research group that publishes and a platform group that builds proprietary models has both exposures and the resources to manage them. A community bank or credit union that buys its analytics from vendors has little development exposure of its own. An academic medical center combines research, clinical operations and, it may be, researchers from abroad, which in our reading puts it close to the notice’s biotechnology, neurotech and AI categories. A federal agency has its own security and access rules, and counsel would need to say how the EAR apply when it releases technology to foreign nationals, including contractors’ staff.

The places work happens: offshore teams and remote access

The second place the notice could land was in how organizations distribute their work. Under the EAR, an export includes an actual shipment or transmission out of the United States, in any manner. A reexport covers movement from one foreign country to another, and a release abroad to a foreign person of a country other than the one where the release takes place. The ANPRM itself framed Commerce’s authority as covering export, reexport and transfer within a country.

A bank with a development center abroad, a health system whose vendor uses an offshore team for coding or analytics, and an agency whose contractors work from several countries all share one feature. Each arrangement can involve sending source code, design documents or model development know-how across a border.

The global delivery model under a new classification

If a technology were moved into a new controlled classification, the question for each offshore center would be whether it receives that technology and whether the destination requires a license. A model development team split between New York and another country would need to know whether sharing a repository counts as an export to that country. A vendor’s offshore team building a decision support tool for a hospital could raise the same question for the vendor and, through the contract, for the hospital.

One possible response is segmentation. Organizations can keep controlled development onshore, or in destinations that do not require a license, while routing other work offshore. That keeps the delivery model but adds coordination costs. Teams that once shared one code base may need two, with a defined interface between them. Handoffs slow, and the organization must police which repository holds what.

Remote access is a transfer question too

Remote work makes geography harder to see. An engineer who travels abroad and opens a repository from a laptop may transmit technology out of the country. An administrator abroad who can read a code base from a cloud console has access to it whether or not they use it. Classic network controls were not designed to answer where the person reading a file is located, or what their nationality is.

An organization that takes the question seriously can combine three controls. Identity attributes record a person’s eligibility for controlled repositories. Conditional access rules restrict access by location. Travel procedures cover what technology may leave the country on a device. Each control has false positives and user friction. The aim is to limit controlled access to the people and places that need it, and to keep records that show it.

Records that show control

If an examiner, auditor or the government ever asks how an organization controls access to restricted technology, the answer is in its records. The useful set is short. It starts with a list of controlled items and where they are stored, and the access list for each with the basis for each person’s eligibility. It adds logs of access by location, and the license or exception relied on for any release to a foreign person.

Keeping those records is easier when access is granted through groups tied to a project, rather than person by person. It is also easier when the classification decision is recorded once, with its reasoning, rather than rediscovered each time a new team member joins. The same records serve an internal investigation if something goes wrong, which gives security teams a reason to support the effort.

Contracts with vendors and outsourcers

For an organization that develops little in house, the largest offshore exposure is likely to be with vendors. A contract with a software developer or managed service provider should state where work is performed, which staff may access the organization’s technology, and who is responsible for export classification of deliverables. It should require notice before work on sensitive components moves to a new country or subcontractor.

These clauses protect both sides. A vendor that moves work offshore without telling its client may create an export the client did not anticipate. A client that sends a vendor controlled technology without saying so may put the vendor in breach. Plain allocation of responsibility in the contract is cheaper than discovering a problem in an audit. Counsel should review where liability falls, since the regulations, not the contract, decide who has made an export.

Agencies and their contractors

Federal agencies face a variant of this. Where development work is done by contractors whose staff include foreign nationals or who use offshore subcontractors, the same questions arise. Agency security requirements may already restrict access to a system. The EAR add a separate layer that turns on the technology, not the system. A program office commissioning an analytics capability should know whether the deliverable could fall in a controlled category and should write that into the statement of work.

The line between research and product

The third place the notice could land was the boundary between research and proprietary development. Here the EAR offer the strongest protection. Technology or software that arises during or results from fundamental research and is intended to be published is not subject to the EAR. Technology or software made available to the public without restriction is published, and it too is not subject to the EAR. Posting on a public website or submitting work for publication are among the routes the regulations list.

The ANPRM expressly preserved that line. Commerce said it did not seek to expand jurisdiction over technologies not currently subject to the EAR, such as fundamental research. For academic medical centers, university-affiliated research institutes and agency research groups, that statement was arguably the most important sentence in the notice.

How the exclusion can be lost

The fundamental research exclusion depends on the intent to publish. The EAR note that once a researcher, institution or company decides to keep research results restricted or proprietary, the technology or software becomes subject to the EAR if otherwise within scope. Prepublication review does not defeat the exclusion when it is limited, for example to protect patent rights with only a temporary delay, or to make sure a sponsor’s proprietary information is not disclosed.

In a research hospital or bank research group, that intent can be unclear. A project may begin as publishable work and become a commercial product when a vendor partnership forms. A sponsor agreement may add publication restrictions that go beyond what the exclusion allows. A team may decide to keep a promising model in house. Each change can move technology across the line without anyone noticing, which is why the decision to restrict publication needs a record and an owner.

Information released to do research

A second limit is easy to miss. The EAR note that the fundamental research exclusion does not apply to technology or software subject to the EAR that is released to conduct the research. If a hospital research group receives controlled technology from an industry partner in order to run a study, the results may be publishable, but the technology received is not thereby freed from control. Access to it is governed in the ordinary way.

For a research administration office, that means tracking what comes in as well as what goes out. Material transfer agreements, data use agreements and technology licenses should identify any controlled items. The research team should know which inputs carry restrictions and which outputs will be published. Mixing the two in one shared drive is one way access rules can fail.

Publishing as a deliberate choice

The published-information rule gives organizations a lever. Technology that has been made available to the public without restriction is not subject to the EAR. That includes posting on a public internet site and submitting a manuscript or dataset to a journal with the intention that it be made public if accepted. An organization that wants its research to be free of export restrictions can choose to publish it, accepting that competitors will see it too.

Where the choice is not made explicitly, incentives may favor keeping models proprietary, for competitive or security reasons. Keeping a fraud detection model private may be right, since publication could help fraudsters. But the choice has an export consequence, and the governance process that decides what to keep private should include that consequence among its inputs.

Datasets, models and weights

The published-information rule lists more than manuscripts. It covers the submission of a computer-readable dataset, formula, imagery, algorithms or other representation of knowledge, with the intention that it be made public if accepted for publication. That language reaches the outputs of data science work as well as papers about it.

For organizations that build models, the line between published and proprietary now runs through artifacts. A model released publicly with its weights is in a different position from one kept behind an interface. A training dataset shared openly with researchers is in a different position from one held under a data use agreement. Health systems face an extra constraint. Privacy law limits how clinical data can be shared, so models and datasets built on it may stay restricted for reasons that have nothing to do with export control.

That combination creates a specific risk in clinical AI research. The research is intended to be published, so the team assumes the fundamental research exclusion applies. But the data and some of the code stay restricted for privacy reasons, and the published paper describes the method without releasing everything. Which parts are outside the EAR and which are not is a question for counsel. A research office that sorts each project’s outputs into published, restricted for privacy, and restricted for competitive reasons will find that question easier to answer.

Who decides

In a university-affiliated health system, the research office, the technology transfer office and the compliance office may each hold part of this decision. In a bank, it may fall to a model governance committee that has never considered export control. The notice was a reason to add one item to the existing decision about whether to publish or keep proprietary. If the organization keeps the work, it should record whether the work is subject to the EAR and who may see it. Recording the answer at the time takes minutes. Reconstructing it years later, from old emails and agreements, can take weeks.

The suppliers: controls that arrive through what you buy

The fourth place the notice could land was in the supply chain. Even an organization that develops nothing controlled depends on suppliers that may. The ANPRM named AI chipsets, AI cloud technologies, microprocessor technology and advanced computing among its categories. Controls on those items would bind the companies that make and sell them, and their compliance obligations would shape what they could offer.

For a regulated buyer, a supplier’s export controls show up in contract terms, product availability and due diligence requests. A cloud provider subject to controls on certain computing capabilities could restrict which regions or customers can use them. A model developer subject to controls on its model weights may restrict where and how its models are hosted. A chip vendor may ask for certifications about end use.

Why this was the exposure most likely to materialize

In 2018, it was not clear which of the four exposures would prove largest. The hindsight section shows that the larger controls reviewed here concentrated on advanced computing chips, semiconductor manufacturing items and, in a rule rescinded before its compliance date, the weights of the most advanced closed AI models. Those are supplier-side items. For most banks, hospitals and agencies, the practical effect of the emerging technology process came through what they could buy and on what terms.

That pattern has a logic. Controlling widely used software techniques such as natural language processing would likely have been hard to enforce and costly to U.S. developers, which the statute required Commerce to consider. Controlling scarce physical inputs and a small number of frontier models is easier to define. The burden of compliance then falls first on the few companies that make those items, and reaches users as conditions of sale.

What a buyer should ask

A procurement team can manage this exposure with a few standard questions. The answers belong in vendor files and should be refreshed when rules change.

  • What export classification does the vendor assign to the product, its components and any technology delivered with it?
  • Does the vendor’s service depend on items subject to license requirements, and could a change in controls affect availability in the regions the organization uses?
  • Will the vendor require end-use or end-user certifications from the organization, and what records must the organization keep?
  • Where will the vendor’s staff and subcontractors who access the organization’s data and technology be located?

These questions do not make the buyer an export compliance expert. They make sure the buyer knows what the vendor has decided, so that an unexpected restriction does not interrupt a critical service.

The gap between a rule and a contract

Export rules can take effect quickly. Interim final rules can take effect on publication or soon after, and comments are collected afterward. Commercial contracts run for years. A buyer whose vendor is suddenly restricted may find that its contract says nothing about what happens next.

Three clauses help. A notice clause requires the vendor to tell the customer promptly when a change in export rules is likely to affect the service. A cooperation clause commits both parties to work on alternatives, such as a different region or configuration. A termination or transition clause gives the customer a way out, with data returned in a usable form, if the service cannot continue lawfully. These clauses cannot prevent a restriction. They shorten the time between the restriction and a workable response.

Concentration risk returns in a new form

Supplier-side controls also interact with concentration risk, a risk that regulated organizations already weigh in their vendor planning. If a small number of providers supply the advanced computing an organization’s analytics depend on, a change in export rules that affects those providers can affect the organization. Business continuity and exit plans for critical technology services should include regulatory events alongside outages and insolvency. A sudden license requirement is, from the buyer’s side, another way a service can become unavailable.

Differences by sector

The supplier exposure differs by sector. Agencies buying for national security purposes may face additional rules of their own, including supply chain requirements. Health systems buying imaging and genomics equipment deal with vendors whose products include advanced computing. A bank’s exposure is likely to run through the cloud and software services it buys. Each should know which of its critical vendors are most exposed to export controls and should ask those vendors how they track rule changes.

Size changes the response more than sector does. A large institution can afford to have someone read each new export rule and ask what it means for its vendors. A small one cannot, and depends on its vendors to tell it. For the small institution, the most valuable step is contractual: making that notice an obligation rather than a courtesy. For the large one, it is analytical: mapping which critical services depend on which classes of controlled item, so that a new rule can be assessed in hours.

Objections: too broad to matter, or too narrow to bite

The strongest objection is that the notice’s list was so broad that it could never have become law in that form. Controlling natural language processing or data visualization across the board would have damaged U.S. technology companies. The statute required Commerce to weigh the effect on U.S. development and the effectiveness of controls. On this view, prudent organizations could have waited for specific proposals and ignored the list.

The objection is largely right about outcomes. The controls that followed were far narrower than the categories. But waiting had a cost of its own. Some organizations did not know which technology they held, who could access it, where work happened or how their suppliers were classified. They would have had to learn all of that quickly if a specific rule had touched them. The four exposures in this article require inventories that take months to build. The notice was a cheap occasion to build them.

The objection from the research side

A second objection comes from research institutions. At least one commenter on the brain-computer interface category, as BIS later summarized the comments, warned that overly broad controls could hinder research with important health care applications. On this view, the right response to the notice was advocacy, not compliance planning.

Both were appropriate. Organizations with a stake in research had every reason to comment, and BIS’s later summary shows that at least one commenter raised health care research directly. They also had reason to make sure their own fundamental research practices were sound, because the exclusion protects only research intended to be published and not restricted. Advocacy shapes the rule. Governance determines whether the organization can rely on the rule’s exclusions.

The objection that this is a defense industry problem

A third objection holds that export control remains a defense and manufacturing issue and that banks, hospitals and agencies should not spend scarce compliance resources on it. For a community bank or credit union that develops nothing in house, that is close to correct. Their exposure is mostly through suppliers, and vendor questions cover it.

For larger organizations with development teams, research arms or offshore delivery, the objection is weaker. Where their technology work resembles that of technology companies, the regulations follow the technology rather than the industry label. The question to ask is not whether the organization is a defense company. It is whether it develops or receives technology that a rule could classify, and whether it would know if it did.

What we would tell a CIO today

The exposures in this article are easier to manage as part of programs an organization already runs than as a new compliance function. Repository inventories belong to the engineering and security teams. Location-aware access belongs to identity management. Publication decisions belong to research governance. Vendor classifications belong to third-party risk management. Each program needs one added field or question, and one person needs to see all four answers together.

The test of readiness is simple to state. Suppose Commerce published a rule tomorrow covering a technique your teams use. Within a week, you should be able to name the code and documents affected and the people with access and their eligibility. You should also be able to name the offshore sites that hold copies and the vendors whose services depend on controlled items. An organization that can do that has prepared enough for the export control questions likely to reach it.

What would change the argument

The argument of this article would be weaker if later controls had reached broadly into everyday AI development and spared suppliers. They did the reverse. Where the rules reviewed here touched the workforce question directly, they carved deemed exports out. That outcome supports reading the 2018 notice as a test of whether organizations understood their own technology and people. The organizations that answered that test had little to change when the narrower rules arrived.

With hindsight

The broad list of 2018 did not become a broad set of controls, at least in the actions reviewed here. On January 6, 2020, BIS published an interim final rule adding software specially designed to automate the analysis of geospatial imagery to ECCN 0D521. It required a license for export and reexport to all destinations except Canada. The 0Y521 series it used is a temporary holding classification that lasts one year unless extended, while the government seeks multilateral controls. Items are added to that series on a Commerce determination, with the concurrence of the Defense and State Departments and other agencies as appropriate. The determination is that they may give a significant military or intelligence advantage, or that foreign policy justifies control. If an item is not moved to a permanent classification before the holding period ends, it becomes EAR99.

On October 26, 2021, BIS asked for comment specifically on brain-computer interface technology, one of the 2018 categories. It reported that its 2018 notice had drawn approximately 13 comments on that category. One respondent said the technology had important applications in health care and assistive technologies, and that overly broad controls could hinder research in those areas.

The larger controls arrived on the supply side. An interim final rule published in October 2022 imposed controls on certain advanced computing integrated circuits, computers containing them and semiconductor manufacturing items, with new license requirements for exports to China. It said its new license requirement for those chips did not apply to deemed exports. It also said certain activities of U.S. persons that support development or production of certain chips in China require a license. That last point brought the workforce back in by a different route: through what U.S. persons do, wherever they work.

On January 15, 2025, BIS published its Framework for Artificial Intelligence Diffusion. It added a control on the model weights of closed-weight AI models trained on more than 10 to the power of 26 computational operations. That rule excluded open-weight models from the new control. Its license requirement for model weights did not apply to deemed exports for people employed by entities headquartered in the United States or in specified destinations. Compliance was due May 15, 2025. BIS said at the time that no open-weight model was known to have been trained above the threshold, and that it would not require a license for the weights of open-weight models. On May 13, 2025, Commerce announced it was rescinding the rule, that it would not enforce it, and that it planned a regulation formalizing the rescission and a replacement rule in the future. Its announcement said the rule would have imposed burdensome requirements on companies and downgraded dozens of countries to second-tier status, and it announced additional steps on semiconductor controls the same day.

What held up from 2018 was the map of exposures. What did not hold up was the fear that everyday AI tools would require licenses. Where the later rules reviewed here reached employees, they carved out deemed exports: fully for the 2022 chip license requirement, and in 2025 for employees of entities headquartered in specified destinations. The burden landed first on chip makers, model developers and their customers’ contracts.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • export controls
  • deemed exports
  • artificial intelligence
  • fundamental research
  • Bureau of Industry and Security