In a previous Bulletin, the authors broadly outlined the legal framework that applies to export controls, as well as the challenges surrounding large language models in artificial intelligence in an era of knowledge sharing. Given that a number of legal and geopolitical developments covering various aspects of this topic took place in 2025, a brief overview is timely on the potential implications for the development of your AI projects, with a special mention of generative AI (or “GenAI”), as the new year begins. What are export controls? Export controls establish rules designed to curb the risk of transferring military, strategic and dual-use (civilian and military) goods and technologies to destinations deemed contrary to national security interests. Such technologies can take on various forms, ranging from physical hardware to technical information. In Canada, export controls are based on a licensing system, under which permits are given based on a series of items listed on the Export Control List (“ECL”) under the Export and Import Permits Act (“EIPA”). To find out if parts of your AI projects are subject to export controls, you should primarily (but not exclusively) refer to that list and to the guide prepared to better understand the list. Key events in 2025 Order SOR/2025-89  On March 7, 2025, an Order amending the ECL was published in the Canada Gazette, in an effort to include emerging technologies that are increasingly faster and more scalable, the capabilities of which raise concerns about potential adversarial military applications.1 Of particular interest in this context, subitem 5506(1) of the schedule to the ECL has been replaced by a number of paragraphs and subparagraphs. But what do these changes mean for AI projects in practice? The amendments made to subitem 5506(1) do not target AI applications (algorithms, models, data), but rather: extreme ultraviolet (“EUV”) lithography equipment, namely EUV masks and reticles making it possible to use this technology to manufacture advanced integrated circuits; cryogenic cooling equipment and ultra-sensitive amplifiers for quantum computers; advanced semiconductor materials; development and production softwares related to certain of the foregoing technologies.2 In other words, subitem 5506(1) targets the industrial toolbox used to build advanced computers, in particular through its inclusion of EUV lithography, which is used for cutting-edge integrated circuits and quantum computers that are revolutionizing the world of advanced computing. It can therefore be said that these rules affect the AI industry because of a form of hardware dependence, since tight control over these infrastructure manufacturing technologies necessarily affect the ability of a country or company to develop and operate advanced AI. In sum, these latest amendments are simply the continuation of those made in the previous year’s Order, which targeted the fields of quantum computing and advanced semiconductor manufacturing in particular (GAAFETs, representing next-generation integrated circuits).3 It has yet to be ascertained how the aforementioned orders will directly affect typical GenAI projects (model development, AI SaaS services, etc.). Those who will experience the more direct repercussions are suppliers of advanced computing equipment and businesses doing R&D on semiconductors, integrated circuits and quantum computing. Notice to Exporters No. 1159 Apart from the technical components, a certain complexity arises when we understand that the definition of a “technology” subject to export controls within the meaning of the law is meant to be broad, and that it includes technical data, technical assistance and information necessary for the development, production or use of an item appearing on the ECL. In other words, the scope of the technologies concerned goes beyond simple physical components or equipment. This is especially true given the proliferation of often cross-border cloud-based solutions, which make technical knowledge accessible digitally and circulate it far and wide. Given this context, it is appropriate to read the Guidance on the movement to and storage of controlled technology in the Cloud (Notice to Exporters No. 1159), published in November 2025 by the Government of Canada. The document was prepared to clarify instances when the use of cloud services constitutes a transfer of controlled technology under the EIPA, requiring a permit.4 In summary, the guidelines state that: it may be considered a transfer if a controlled technology is disclosed from a place inside Canada to a place outside Canada; a controlled technology is considered disclosed if it is sent from Canada and stored in a foreign location in a way that creates a reasonable possibility that a person located outside Canada would be in a position to examine that technology; a reasonable possibility means more than a mere possibility, but less than the standard of “more likely than not”; the location of servers hosting controlled technology only matters if it affects the reasonable possibility that the technology could be disclosed outside Canada; in general, it is considered a transfer when a person located outside Canada holds decryption keys or routine access rights that create more than a remote possibility that the technology may be examined, or when a cloud service provider creates an unencrypted backup copy that contains controlled technology to restore a system after an incident, and that such copy is stored on servers outside Canada where foreign administrators can access it; when cloud services are used, both the owner of the controlled technology and the cloud service provider have a degree of care and control of the technology. Thus, not only is there a risk of knowledge sharing where items directly listed on the ECL are involved (whether to manufacture them or otherwise), but the possibility of violating export controls also exists because of the interaction between cloud services and the knowledge that could be transferred (within the meaning set out above), if the cloud contains information about or relates to a controlled technology. Considerations regarding GenAI What about GenAI projects? Despite all of the above, these projects may still suffer indirect repercussions, and not only on highly technical components. You will need to exercise a certain degree of caution regarding the compliance of your GenAI projects because of the amount of information they can accumulate through the various layers of their structure. Training data There are the data used during the GenAI’s learning phase, before it is rolled out. The amount of this data can be massive, and it can be structured or unstructured. It is used to provide a knowledge base for the model and enable it to produce relevant outputs when it is given inputs. The learning phase is risky if the datasets contain controlled technical information and if the data can be regurgitated or combined when users use the GenAI. The GenAI’s weights, filters, and other operating parameters These parameters can be compared to physical control buttons—they are adjusted during the GenAI’s training and during the configuration of the solution that uses it. They determine how much each input element will influence the response and refine the model (i.e., the structure that allows the GenAI to interpret inputs and generate outputs). In the United States, weights in particular are a hot topic considering the country’s export policy, under which they can constitute key parameters for the most advanced AI models. Inputs This is the data provided by users to generate relevant outputs (e.g., text, images, structured data) when the GenAI is already rolled out. Such data is used to trigger a response or behaviour from the model. Just like with training data, inputs will be critical depending on the use made of the model and the information disclosed to obtain a response. Conditions consistent with legal requirements must be provided to prevent the model from being contaminated by sensitive data after it is rolled out, especially if it stores all the inputs provided to it for its continued learning. Outputs This is what GenAI generates in response to inputs. Outputs can be in the form of text responses or images, codes, or even data-based predictions. Given the above, it will be challenging depending on the datasets conveyed by the GenAI, to ensure that outputs do not violate export controls, as they could make it possible to indirectly obtain information the direct access to which would otherwise be prohibited. Conclusion We can imagine that the recent changes to export controls in Canada are just the beginning of an effort to address new concerns arising from this rapidly changing and ever more powerful technology. Export controls are also not devoid of a diplomatic context. For now, making AI subject to export controls seems to be the preferred mechanism to curb the exponential powers of such technology in Canada. The extent to which this will be done remains to be seen and will be interesting to follow. Government of Canada, Order Amending the Export Control List: SOR/2025-89 (March 7, 2025): Canada Gazette, Part II, Volume 159, Number 7: Order Amending the Export Control List: This is not an exhaustive list, but rather a few relevant examples that apply to advanced computing. Government of Canada, Order Amending the Export Control List: SOR/2024-112 (May 31, 2024): Canada Gazette, Part II, Volume 158, Number 13: Order Amending the Export Control List: Government of Canada, Notice to Exporters No. 1159 – Guidance on the movement to and storage of controlled technology in the Cloud (amended November 10, 2025): Notice to exporters no 1159 – Guidance on the movement to and storage of controlled technology in the Cloud

Introduction When we hear the term “export controls,” we may think it only applies to weapons and other highly sensitive technologies, but that is not the case. There are a multitude of circumstances—some unexpected—to which it is important to know that export controls apply. This is especially true if you are involved in research or in the design and development of seemingly innocuous solutions that are not necessarily tangible objects. Today, technological knowledge is shared not only through conventional partnerships between businesses or universities, but also through data sharing or access to databases that feed large language models. Artificial intelligence is, in itself, a means of sharing knowledge. Feeding such algorithms with sensitive data, or data that can become sensitive when combined, carries a risk of violating the applicable legal framework. Here are some key concepts. Overview of the federal export control framework The Export and Import Permits Act In Canada, the Export and Import Permits Act (the “EIPA”) establishes the primary framework governing the export of controlled goods and technologies. The EIPA gives the Minister of Foreign Affairs the power to issue, to any resident of Canada who applies for one, a permit authorizing the export or transfer of a wide range of items included on the Export Control List (the “ECL”) or destined for a country listed on the Area Control List. In other words, the EIPA regulates, and at times prohibits, the trade of critical goods and technologies outside Canada. The Export Control List To get the full picture of the ECL, we need to refer to the Guide to Canada's Export Control Listas published by the Department with its successive amendments, the most recent of which date back to May 2025 (the “Guide”). In summary, the Guide includes military goods and technologies, strategic goods and dual-use (civilian and military) goods and technology that are controlled in accordance with Canada’s commitments made in multilateral regimes, such as the Wassenaar Arrangement on Export Controls for Conventional Arms and Dual-Use Goods and Technologies, bilateral agreements, and certain unilateral controls implemented by Canada as part of its defence policy. The Guide also includes forest products, agricultural and food products, apparel goods and vehicles. Other laws that affect exports Also to take into account are the sanctions that Canada imposes under laws that affect exports, such as: the United Nations Act the Special Economic Measures Act the Justice for Victims of Corrupt Foreign Officials Act These sanctions against specific countries, organizations or persons include a number of measures, including restricting or prohibiting trade, financial transactions or other economic activities with Canada, or the freezing of property located in Canada.1 Finally, in order for an individual (or an organization) to transfer controlled goods outside Canada, they must register with the Controlled Goods Program (the “CGP”) to obtain an export permit, unless exempt. Key concepts Did you know? Certain goods and technologies are referred to as “dual-use” goods and technologies. This means that even though they were initially designed for civilian use or appear harmless, they may be subject to export controls if they can be used for military purposes or to produce military items. A “technology” is broadly defined to include technical data, technical assistance and information necessary for the development, production or use of an item listed on the ECL. Also included in this notion, albeit indirectly, are the technologies referred to in any of the regulations associated with the laws listed above, which make certain countries subject to specific technology transfer restrictions. A “transfer” in relation to a technology, means to dispose of it (e.g. sell it) or disclose its content in any manner from a place in Canada to a place outside Canada. This definition stems from legislative amendments to the EIPA, which expanded the scope of the law to include the mere transfer of intangible technologies by various means, thereby broadening the circumstances to which permits apply as regards transfers.2 Regarding trade relations with the United States, Canadian exporters may face additional restrictions and considerable challenges, particularly in situations where their employees or other stakeholders involved are foreign nationals.The International Traffic in Arms Regulations (“ITAR”) and the Export Administration Regulations(“EAR”) are two key sets of rules that govern exports from the United States.3 They protect both similar and distinct interests. While the ITAR aim to protect defence articles and defence services (including weapons and information), the EAR govern dual-use items.4 Both prevent exports5 in a broad sense, i.e., up to and including the transfer of information to so-called “foreign” persons, except with the permission of the authorities. It is thus quite possible that Canadian exporters will be required to comply with these American regulations, which, in addition to targeting territories, target the national origin of individuals. This is diametrically opposed to Canada’s export regime, which rather centres on prohibiting trade with a country or anyone located there. In this regard, note that Quebec’s Charter of Human Rights and Freedoms considers national origin to be a ground for discrimination. 6 A Quebec business can thus find itself struggling to balance its contractual obligations under a contract with an American company with the requirements of the Quebec Charter. Artificial intelligence: novel challenges The development of large language models in the field of artificial intelligence represents a new challenge from an export control standpoint, and a significant one at that. For example, if a large language model is trained using restricted data, a state subject to the aforementioned sanctions might attempt to use the large language model to indirectly obtain information to which it would not otherwise have had direct access. As a result, training a large language model on plans, technical specifications or textual descriptions of technologies covered by transfer restrictions (which can include knowledge transfers) can create a risk of non-compliance with the law. The same applies to accessing such data for retrieval-augmented generation, a widely used technique to expand and improve large language model responses. To limit the risk during research and development, a company that trains a large language models on such data or allows access to such data for retrieval-augmented generation will need to consider where the data will be hosted and processed. Similarly, once the artificial intelligence application is developed, it will be important to restrict access to it in a manner consistent with the law, both in terms of locating the servers on which the large language model will be installed and in terms of user access. Sanctions Any person or organization that contravenes any provision of the EIPA or its regulations commits an offence punishable by fine and/or imprisonment, as applicable. Also, failure to register with the CGPmay constitute an offence under federal laws that can lead to prosecution and substantial sanctions against the offender(s).7 Conclusion Canada’s export controls are quite complex, not only in how they are structured, but also in how they must be implemented. With the changing geopolitical and commercial landscape, it is advisable to periodically read the resources made available by the relevant authorities and put in place appropriate policies and measures, or to seek professional advice in this regard. Government of Canada, “Types of sanctions” (date modified: 2024-09-10): Types of sanctions Martha L. Harrison & Tonya Hughes, “Understanding Exports: A Primer on Canada’s Export Control Regime” (2010) 8(2) Canadian International Lawyer, 97 The ITAR and EAR are included in the Code of Federal Regulations (“CFR”). Austin D. Michel, “Hiring in the Export-Control Context: A Framework to Explain How Some Institutions of High Education Are Discriminating against Job Applicants” (2021) 106:4 Iowa L Review, 1993 The ITAR and EAR also provide for restrictions on re-exportation. See Maroine Bendaoud, “Quand la sécurité nationale américaine fait fléchir le principe de non-discrimination en droit canadien : le cas de l'International Traffic in Arms Regulations (ITAR)” (2013) Les cahiers de droit, 54 (2–3), 549 Government of Canada, “Guideline on Controlled Goods Program registration” (date modified: 2025-05-08): Guideline on Controlled Goods Program registration – Canada.ca

Blind spots to watch for when anonymizing data Anonymization has become a crucial step in unlocking the value of data for innovation, particularly in artificial intelligence. But without a properly executed anonymization process, organizations risk financial penalties, legal action and serious reputational harm, with potentially significant consequences for their operations. Understanding the anonymization process What the law says Under Quebec’s Act respecting the protection of personal information in the private sector (the “Private Sector Act”) and the Act respecting Access to documents held by public bodies and the Protection of personal information (the “Access Act”), information concerning a natural person is considered anonymized if it irreversibly no longer allows the person to be identified directly or indirectly. Since anonymized information no longer qualifies as personal information, this distinction is of crucial importance. However, beyond this definition, neither Act provides details on how anonymization should actually be performed. To fill this gap, the government adopted the Regulation respecting the anonymization of personal information (the “Regulation”), which sets out the criteria and framework for anonymization, grounded in high standards of privacy protection. What organizations need to know before starting Under the Regulation, before beginning any anonymization process, organizations must clearly define the “serious and legitimate purposes” for which the data will be used. These purposes must comply with either the Private Sector Act or the Access Act, as applicable, and any new purpose must meet the same requirement. The process must also be supervised by a qualified professional with the expertise to select and apply appropriate anonymization techniques. This supervision ensures both the proper implementation of the chosen methods and the ongoing validation of technological choices and security measures. The four key steps of data anonymization   DepersonalizationThe first step is to remove or replace all personal identifiers, such as names, addresses and phone numbers, with pseudonyms. It is essential to anticipate how different data sets might interact, in order to minimize the risk of re-identifying individuals through cross-referencing. Preliminary risk assessmentNext comes a preliminary analysis of re-identification risks. This step relies on three main criteria: individualization (inability to isolate a person within a dataset), correlation (inability to connect datasets concerning the same person) and inference (inability to infer personal information from other available information). Common anonymization techniques include aggregation, deletion, generalization and data perturbation. Organizations should also apply strong protective measures, such as advanced encryption and restrictive access controls, to minimize the likelihood of re-identification. In-depth risk analysisAfter the preliminary phase, a deeper risk analysis must be conducted. While no anonymization process can eliminate all risk, that risk must be reduced to the lowest possible level, taking into account factors such as data sensitivity, the availability of public datasets and the effort required to attempt re-identification. To sustain this low level of risk, organizations should perform periodic reassessments that account for technological advances that could make re-identification easier over time. Documentation and record-keepingFinally, organizations must keep a detailed record describing the anonymized information, its intended purposes, the techniques and security measures used, and the dates of any analyses or updates. This documentation strengthens transparency and demonstrates that the organization has fulfilled its legal obligations regarding anonymization.

While the world is focused on how the tariff war is affecting various products, it may be overlooking the risks the war is posing to information technology. Yet, many businesses rely on artificial intelligence to provide their services, and many of these technologies are powered by large language models, such as the widely-used ChatGPT. It is relevant to ask whether businesses should rely on purely US-based technology service providers. There is talk of using Chinese alternatives, such as DeepSeek, but their use raises questions about data security and the associated control over information. Back in 2023, Professor Teresa Scassa wrote that, when it comes to artificial intelligence, sovereignty can take on many forms, such as state sovereignty, community sovereignty over data and individual sovereignty.1 Others have even suggested that AI will force the recalibration of international interests.2 In our current context, how can businesses protect themselves from the volatility caused by the actions of foreign governments? We believe that it’s precisely by exercising a certain degree of sovereignty over their own affairs that businesses can guard against such volatility. A few tips: Understand Intellectual property issues: Large language models underlying the majority of artificial intelligence technologies are sometimes offered under open-source licenses, but certain technologies are distributed under restrictive commercial licenses. It is important to understand the limits imposed by the licenses under which these technologies are offered. Some language model owners reserve the right to alter or restrict the technology’s functionality without notice. Conversely, permissive open-source licenses allow a language model to be used without time restrictions. From a strategic standpoint, businesses should keep intellectual property rights over their data compilations that can be integrated into artificial intelligence solutions. Consider other options: Whenever technology is used to process personal information, a privacy impact assessment is required by law before such technology is acquired, developed or redesigned.[3] Even if a privacy impact assessment is not legally required, it is prudent to assess the risks associated with technological choices. If you are dealing with a technology that your service provider integrates, check whether there are alternatives. Would you be able to quickly migrate to one of these if you faced issues? If you are dealing with custom solution, check whether it is limited to a single large language model. Adopt a modular approach: When a business chooses an external service provider to provide a large language model, it is often because the provider offers a solution that is integrated to other applications that the business already uses, or because it provides an application programming interface developed specifically for the business. In making such a choice, you should determine whether the service provider can replace the language model or application if problems were to arise. If the technology in question is a fully integrated solution from a service provider, find out whether the provider offers sufficient guarantees that it could replace a language model if it were no longer available. If it is a custom solution, find out whether the service provider can, right from the design stage, provide for the possibility of replacing one language model with another. Make a proportionate choice: Not all applications require the most powerful language models. If your technological objective is middle-of-the-road, you can consider more possibilities, including solutions hosted on local servers that use open-source language models. As a bonus, if you choose a language model proportionate to your needs, you are helping to reduce the environmental footprint of these technologies in terms of energy consumption.  These tips each require different steps to be put into practice. Remember to take legal considerations, in addition to technological constraints, into account. Licenses, intellectual property, privacy impact assessments and limited liability clauses imposed by certain service providers are all aspects that need to be considered before making any changes. This isn’t just about being prudent—it’s about taking advantage of the opportunity our businesses have to show they are technologically innovative and exercise greater control over their futures. Scassa, T. 2023. “Sovereignty and the governance of artificial intelligence.” 71 UCLA L. Rev. Disc. 214. Xu, W., Wang, S., & Zuo, X. 2025. “Whose victory? A perspective on shifts in US-China cross-border data flow rules in the AI era.” The Pacific Review, 1–27. See in particular the Act respecting the protection of personal information in the private sector, CQLR c. P-39.1, s. 3.3.