Late last week, a new generative AI large language model called DeepSeek was publicly launched by two Chinese entities, the Hangzhou and Beijing DeepSeek Artificial Intelligence Cos. Ltd. DeepSeek is currently driving headlines claiming it represents a “Sputnik moment“ in AI development.
DeepSeek claims to perform at the same level as US-developed LLMs like ChatGPT – but to have been trained at a miniscule fraction of the cost. This was allegedly due to DeepSeek relying on older-model, less sophisticated chips, while finding innovation such as more efficient processing for prompts, different reinforcement processes for training, and a more compressed model. (We take no position on whether these claims are correct; for instance, DeepSeek’s claims about it training costs are being scrutinized by the market.) As of this week, DeepSeek appears to be one of the most-downloaded generative AI app in the US. DeepSeek even took the unusual step of open-sourcing its generative AI model and posting it on Github.
For companies, the cost and sustainability considerations associated with DeepSeek may be of primary interest. Initial reports indicate that DeepSeek offers its LLM at less than 1/15th of the cost of US-based LLMs, while claiming similar performance. These same reports suggest that DeepSeek’s energy usage is significantly less than other LLMs. These factors could lead companies to consider piloting or testing DeepSeek for cost savings, or to help with sustainability goals or ESG metrics.
As companies evaluate DeepSeek, it remains prudent to consider potential obstacles that could arise from recent data-related regulations passed in the US that are designed to broadly restrict the availability of US person data to Chinese organizations:
- First, the US Department of Justice recently finalized regulations that significantly restrict US companies’ ability to engage in “data transactions” with Chinese companies. These regulations – if they enter into effect – would complicate corporate uses of DeepSeek.
- Second, if later reporting uncovers that larger Chinese social media players are involved in creating or operating DeepSeek, the “TikTok ban” statute may grant President Trump authority broad enough to consider banning DeepSeek from the US market.
- Third, if the Trump administration views DeepSeek as a national security threat, it could leverage a 2019 EO on Information and Communications Technology Services from the first Trump administration to seek broader restrictions or prohibitions on licensing, downloading, or using DeepSeek.
This article briefly summarizes each of these considerations. Please note this article does not claim to exhaustively analyze potential risks of, or compliance strategies for, deploying DeepSeek. More is learned about DeepSeek daily and the situation is developing rapidly. For this reason, this article focuses solely on recent US statutes and regulations that could have a broad, “ban”-like effect on the deployment or use of DeepSeek by US organizations.
- DOJ’s US Data Regulation
On January 8, 2025, the US Department of Justice finalized regulations ”Preventing Access to U.S. Sensitive Personal Data and Government-Related Data by Countries of Concern or Covered Persons” (the “US Data Regulation”). The US Data Regulation prohibits or restricts transactions with companies from “countries of concern” – i.e. China, Russia, Iran, North Korea, Cuba, and Venezuela. With regards to China, broadly speaking, the US Data Regulation applies to any relationship between a US company and a Chinese company in which certain quantities of US persons’ information will be made accessible to the Chinese company. If defined quantities of US person information are made available to the Chinese counterparty, the transaction is a “covered data transaction” subject to the restrictions or prohibitions of the Regulation.
The US Data Regulation recognizes two types of transactions that might be relevant for DeepSeek:
- Restricted “Vendor Agreements:” Vendor agreements are any agreements in which a Chinese company provides goods or services to a US company in exchange for consideration. DeepSeek offers an API, which DeepSeek’s Terms of Use describe as a “Service.” Notably, vendor agreements with Chinese companies are not prohibited under the US Data Regulation. However, they are deemed “restricted.” They can be engaged in, but only if a significant catalogue of security safeguards issued by the Cybersecurity and Infrastructure Security Agency (CISA) have been implemented for all systems accessible by the Chinese vendor which may process US person information. One core requirement is a “Data Risk Assessment” by the US party that documents whether and how applicable security controls “prevent[] access to covered [US person] data” by the Chinese vendor and its personnel.
Here, DeepSeek’s Privacy Policy states that DeepSeek may collect not just usernames and inputs into its LLM, but will also assign a unique “device ID and user ID” to each user, while collecting users’ “keystroke patterns or rhythms.” This is broadly the type of data the US Data Regulation applies to when made available to a Chinese vendor. DeepSeek’s capture of keystroke rhythms could raise questions about whether DeepSeek collects “biometric identifiers” within the meaning of the Regulation, which defines biometrics as including “keyboard usage patterns” that are “enrolled in a biometric system.” DeepSeek’s Privacy Policy confirms that data it collects is stored in China. While DeepSeek’s Terms of Use suggest DeepSeek applies security controls when re-using prompts or outputs to train its model (“secure encryption technology processing, strict de-identification rendering, and irreversibility to identify specific individuals”), these do not seem to prevent DeepSeek from having general access to user data in-the-plain, including data about US users.
As a result, to enable corporate use of DeepSeek via licensing from DeepSeek, companies may need to scrutinize whether the CISA security controls can be documented as implemented by DeepSeek in the manner it offers its LLM. Engaging in a “restricted vendor transaction” with DeepSeek would also trigger recordkeeping and auditing requirements. The US company that engages DeepSeek would need to document its due diligence regarding data flows and implementation of the CISA security controls. More broadly, the company would also need to create a “Data Compliance Program,” as well as a program designed to ensure CISA controls remain implemented for its “restricted transactions.” Both of these must be audited annually.
- Prohibited “Data Brokerage.” The US Data Regulation flatly prohibits “data brokerage,” which it defines broadly as any “sale of data, licensing of access to data, or similar commercial transactions” that do not fall under the category of “vendor agreement” in which “the recipient [of US person data] did not collect … the data directly from the individuals linked or linkable to the … data.” We anticipate US companies would likely take the position that use of DeepSeek is not a “sale” or “licensing” of data, even if DeepSeek does not “collect” the data of US persons whose data is included in inputs into its platform. Additionally, the fact that DeepSeek positions itself as a “service” may enable companies to argue DeepSeek is a “vendor agreement,” which would carve it out of the “data brokerage” definition. Still, the penalties for violating the US Data Regulation may make it advisable to document an analysis that use of DeepSeek is not “data brokerage” within the meaning of the Regulation.
Although the US Data Regulation reads like a privacy statute, it was passed under the International Emergency Economic Powers Act (IEEPA) – a sanctions law. Violations thus carry both civil fines as well as criminal penalties. Executive Order 14117 was the basis for DOJ’s drafting of the Regulation, and it has not been rescinded by the Trump Administration. Instead, the Regulation is currently subject to the Trump Administration’s general regulatory freeze. If the Regulation remains in force following review by the Trump Administration, it is currently slated to enter into force on April 8, 2025. (This date could potentially get pushed back due to the regulatory freeze, or if the Trump Administration considers changes to the Regulation.)
What about downloading and self-hosting DeepSeek’s model, or licensing DeepSeek from a US provider that has downloaded DeepSeek’s model and is offering it from a US-based environment? After all, DeepSeek’s model is open-source. Localizing the DeepSeek model may represent a potential approach for avoiding the issues outlined above. Note, however, that the US Data Regulation is based on sanctions authority, so could potentially be enforced on a strict-liability basis. Testing and validation of how DeepSeek works, even when locally operated, would likely be advisable to confirm and document it does not represent a covered “data transaction.”
- The “TikTok Ban” Statute
In March 2024, the US Congress passed the “TikTok ban” statute, formally titled the “Protecting Americans from Foreign Adversary Controlled Applications Act” (PAFACA). PAFACA is initially directed specifically against the TikTok app and its corporate parent, ByteDance Ltd. It also grants the president authority to ban additional websites or apps that are “controlled by a foreign adversary.” However, for the President to issue a ban, the website/app must be offered by a “covered company.” Under the statute’s definition, a “covered company” reads like a social media network, i.e. a company that offers a website/app that:
- (i) permits a user to create an account or profile to generate, share, and view text, images, videos, real-time communications, or similar content;
- (ii) has more than 1,000,000 monthly active users with respect to at least 2 of the 3 months preceding the date on which a [“ban” decision is issued];
- (iii) enables 1 or more users to generate or distribute content that can be viewed by other users of the website, desktop application, [or] mobile application …; and
- (iv) enables 1 or more users to view content generated by other users of the website, desktop application, [or] mobile application … .
At present, there do not appear to be indications that the companies operating DeepSeek (Hangzhou and Beijing DeepSeek Artificial Intelligence Cos. Ltd.) meet these requirements. If, however, later reporting were to uncover that larger Chinese companies that also operate in China’s social media environment were involved in the funding, training, or operation of DeepSeek, this could potentially open an analysis about whether the TikTok ban statute gives President Trump authority to issue a ban.
If this occurs, it would open a discussion about how a PAFACA “ban” of DeepSeek would work in practice. PAFACA does not ban US persons from using banned applications; instead, it prohibits US persons from offering banned applications in an app store, or from providing hosting services that support the banned application’s operations. To date, there has been no interpretation addressing whether a market-standard enterprise deployment of a generative AI model like DeepSeek could fall under one of those headings.
- The 2019 EO on Information & Communications Technology Services
During President Trump’s first term in office, he passed EO 13873 on “Securing the Information and Communications Technology and Services Supply Chain” (the “ICTS EO”). The ICTS EO declared a national emergency based on the threat that information technologies developed by “foreign adversaries” (which included China) posed to US national security. It grants the Department of Commerce broad authorities to investigate foreign-developed information technologies for national security risks, and to determine that acquisition, importation, installation, or use of the technology should be restricted or prohibited. On the basis of the ICTS EO, the first Trump administration issued EOs prohibiting certain transactions involving Chinese applications like WeChat and TikTok.
The Biden administration rescinded these orders against WeChat and Tiktok, but did not rescind the ICTS EO itself, which remains in effect. As such, the ICTS EO would be available for the Trump administration to review DeepSeek and, if desired, develop restrictions on the ability of US organizations to license, install, or use DeepSeek. This would represent the first use of the ICTS EO on an open-source AI model that can be downloaded and used locally, as opposed to a proprietary app that – in all products offered to the market – is hosted by a foreign provider. That may raise interpretive questions of first impressions concerning some of the ICTS EO’s requirements for restricting transactions. But given how the first Trump administration utilized the ITCS EO, it would seem to remain a authority that might possibly be utilized in relation to DeepSeek.
————————————————————————————————————–
As stated above, additional considerations could arise from corporate use of DeepSeek that this article does not catalogue. For example, US-based financial institutions would likely evaluate GLBA Privacy and Safeguards Rule compliance; US-based healthcare entities would evaluate HIPAA Privacy and Security Rule compliance; and many companies across industries would potentially assess whether DeepSeek use entails a “sale” of data under state privacy statutes such as California’s Consumer Privacy Act. Similarly, companies subject to the EU’s General Data Protection Regulation would likely evaluate whether corporate use of DeepSeek is compliant within the GDPR’s framework; in this regard, the Italian data protection authority’s recent announcement it has issued requests for information to DeepSeek may be relevant. Also, as more is learned about the training of DeepSeek and the infrastructure on which it operates, sanctions questions could arise. The situation is dynamic and developing rapidly, and positions on these issues could change as more is learned about DeepSeek. Still, the above “ban”-like risks may remain among the larger potential obstacles to corporate use of DeepSeek as long as the US Data Regulation is permitted to enter into force in its current form.
