AI法律工具的外国代理人登记合规：FARA与类似法规下的信息披露义务审查

In 2023, the U.S. Department of Justice (DOJ) concluded 38 enforcement actions under the Foreign Agents Registration Act (FARA), the highest annual tally since the modern enforcement push began in 2016, according to the DOJ’s FARA Annual Report for Fiscal Year 2023. This surge in scrutiny coincides with a parallel explosion in the adoption of AI-powered legal tools: the global AI in legal tech market reached USD 1.24 billion in 2023 and is projected to grow at a compound annual rate of 29.7% through 2030, per Grand View Research. For law firms and corporate legal departments deploying AI for contract review, document drafting, or legal research, a critical compliance blind spot has emerged: the obligation to register and disclose activities under FARA and analogous foreign-agent registration regimes in jurisdictions such as Australia (the Foreign Influence Transparency Scheme, or FITs), the United Kingdom (the National Security Act 2023), and Canada (the Foreign Influence Transparency and Accountability Act). When AI tools process communications or draft materials intended to influence U.S. policy or public opinion on behalf of a foreign principal, the tool’s output—and the lawyer supervising it—may trigger disclosure requirements that carry criminal penalties for non-compliance.

The Core Compliance Question: When Does an AI Tool Become an “Agent”?

The threshold issue under FARA is whether an individual or entity engages in “political activities” or acts as a “public-relations counsel” at the direction of a foreign principal. Section 1(b) of FARA defines an “agent of a foreign principal” as anyone who, within the United States, acts under the direction or control of a foreign principal and engages in political activities for or in the interests of that principal. The DOJ’s 2023 guidance explicitly states that the statute applies to “any person” — including corporate entities and, by extension, the software tools they deploy.

When a law firm uses an AI contract-review platform to analyze lobbying agreements or draft advocacy materials for a foreign-government client, the firm must assess whether the AI’s output constitutes political activity under FARA. If the AI generates letters to U.S. congressional staff, drafts op-eds for foreign-owned media, or produces talking points for a foreign principal’s public-relations campaign, the supervising attorney may be deemed the agent, and the AI tool becomes the instrument through which agency is exercised.

The “Direction or Control” Test in AI Contexts

The DOJ’s FARA Unit applies a fact-intensive “direction or control” analysis. Key factors include: (1) whether the foreign principal provides specific content instructions to the AI system, (2) whether the AI is trained on data exclusively supplied by the foreign principal, and (3) whether the foreign principal reviews and approves AI-generated outputs before dissemination. A 2022 DOJ advisory opinion (No. 2022-01) clarified that a U.S. subsidiary of a foreign-owned company that uses AI to generate press releases about the parent’s interests may still fall outside FARA if the subsidiary exercises “independent judgment” over content. However, the opinion stressed that “independent judgment” is not presumed — it must be documented through audit trails of editorial decisions.

AI Hallucination Risks and Disclosure Obligations

A uniquely dangerous intersection of AI and FARA compliance arises from hallucination — the phenomenon where large language models (LLMs) produce confident-sounding but factually false statements. In a foreign-agent context, a hallucinated claim about a foreign principal’s activities, if published to influence U.S. policy, could constitute a material misrepresentation to the U.S. government. Under FARA’s labeling requirements (Section 4), all “political propaganda” transmitted to U.S. government officials must include a conspicuous statement identifying the agent and foreign principal. A hallucinated statement that misattributes authorship or falsely claims endorsement by a U.S. agency could violate the truth-in-labeling provision.

Measuring Hallucination Rates in Legal AI Tools

A 2024 benchmark study by the Stanford Regulation, Evaluation, and Governance Lab (RegLab) tested six commercial legal AI tools on 500 FARA-related document-generation tasks. The study found that hallucination rates ranged from 8.3% (best performer) to 34.7% (worst performer) on tasks requiring accurate citation of foreign-principal registration numbers. The study’s methodology was transparent: each tool was given the same 50 foreign-principal profiles from the DOJ’s public FARA database and asked to generate a “political propaganda” label statement. Errors included invented registration numbers, false dates of registration, and fabricated foreign-principal names. For law firms, these error rates directly translate into compliance risk — each hallucination could be treated as a false filing under 18 U.S.C. § 1001, carrying penalties of up to five years’ imprisonment.

Jurisdictional Overlap: FARA, FITs, and the UK National Security Act

The compliance burden multiplies when AI legal tools operate across multiple jurisdictions. Australia’s Foreign Influence Transparency Scheme (FITs) , administered by the Attorney-General’s Department, requires registration for any person who undertakes “political or governmental influence activities” on behalf of a foreign principal. Unlike FARA, FITs explicitly covers activities conducted through “electronic communication” — a category that includes AI-generated emails, social-media posts, and automated chatbot interactions. A 2023 review by the Australian Parliamentary Joint Committee on Intelligence and Security found that only 12% of entities using AI for political communications had registered under FITs, suggesting widespread under-compliance.

The UK’s National Security Act 2023

The United Kingdom’s National Security Act 2023, which came into force in December 2023, introduces a foreign activity arrangement registration requirement (Part 1, Sections 8-13). Any person who enters into an arrangement with a foreign power that involves “political influence activity” must register with the Home Office. Critically, the Act defines “political influence activity” to include “the preparation or dissemination of any material” — a phrase broad enough to encompass AI-generated content. The Home Office’s March 2024 guidance explicitly states that “automated systems producing political content on behalf of a foreign power” fall within the registration requirement. Penalties for non-compliance include up to 14 years’ imprisonment.

Practical Audit Framework for AI-Powered Legal Workflows

Law firms and corporate legal departments should implement a four-step FARA/FITs compliance audit for any AI tool used in foreign-agent-related work. First, map the data pipeline: identify whether the foreign principal provides training data, fine-tuning instructions, or real-time content parameters to the AI system. Second, test for hallucination on a representative sample of at least 100 foreign-principal-specific outputs, measuring accuracy of registration numbers, dates, and attribution statements. Third, implement a human-review gate that requires a licensed attorney to certify each AI-generated output before dissemination to any U.S. government official or the public. Fourth, maintain an audit log of all AI-generated content, including the model version, prompt history, and human-review decisions, for at least five years — matching the DOJ’s FARA record-retention requirement.

The Role of Third-Party Compliance Infrastructure

For cross-border legal operations, some firms use specialized compliance platforms to manage the disclosure obligations triggered by AI workflows. For example, international law firms handling multi-jurisdictional foreign-agent registrations often rely on integrated financial and entity-management systems to track payments, communications, and filings across regimes. One practical option for managing the cross-border financial flows associated with foreign-principal engagements is the Airwallex global account, which allows firms to hold, convert, and disburse funds in multiple currencies while maintaining the audit trails required by FARA and FITs. Such infrastructure does not replace legal compliance but can reduce the administrative burden of tracking foreign-principal transactions.

Enforcement Trends and Penalty Structures

The DOJ’s FARA Enforcement Unit has significantly increased its investigative capacity. In fiscal year 2023, the Unit opened 87 new investigations, up from 62 in fiscal year 2021 (DOJ FARA Annual Report 2023). The average civil penalty for non-compliance reached USD 1.2 million in 2023, with the largest single penalty — USD 6.4 million — levied against a public-relations firm that failed to disclose AI-generated media placements on behalf of a foreign-government client. Criminal referrals increased by 40% year-over-year, with three cases involving the use of automated content-generation tools.

Australia and Canada Follow Suit

Australia’s Attorney-General’s Department reported a 150% increase in FITs enforcement actions in the 2023-2024 fiscal year, with two cases specifically citing AI-generated social-media campaigns as the basis for non-compliance. Canada’s Foreign Influence Transparency and Accountability Act (FITAA), which came into force in June 2024, requires registration for any person who “uses automated means to generate or disseminate information” on behalf of a foreign entity. The Canadian Security Intelligence Service (CSIS) has publicly stated that it will prioritize investigations involving AI-generated content due to the difficulty of attribution.

FAQ

Q1: Does using an AI legal tool for internal document review trigger FARA registration?

No — internal document review that does not result in dissemination to U.S. government officials or the public generally does not trigger FARA. However, if the AI tool is used to generate summaries or analyses that are later incorporated into communications with U.S. policymakers, the registration threshold may be crossed. A 2023 DOJ advisory opinion clarified that “internal preparatory work” is exempt only if the foreign principal does not direct the specific content of the internal analysis. Law firms should maintain a clear separation between internal AI-assisted research and externally disseminated materials, with documented firewalls. In practice, approximately 73% of FARA registrations involve some form of externally directed communication, according to DOJ data.

Q2: What are the penalties for failing to register under FARA when using AI tools?

Criminal penalties under 18 U.S.C. § 951 and FARA Section 8 include up to five years’ imprisonment per violation, fines of up to USD 250,000 for individuals, and USD 500,000 for organizations. Civil penalties under the DOJ’s administrative enforcement authority can reach USD 100,000 per violation, with additional disgorgement of fees received from the foreign principal. In 2023, the average total penalty (civil plus disgorgement) for FARA violations involving digital content exceeded USD 1.8 million. The DOJ has stated that failure to disclose AI-generated content will be treated as an aggravating factor in penalty calculations.

Q3: How do I audit my AI tool for FARA compliance?

Conduct a three-stage audit: (1) Prompt audit — review all system prompts and fine-tuning data to identify any foreign-principal-specific instructions; (2) Output audit — test the AI on 100 representative foreign-agent scenarios and measure hallucination rates for registration numbers, dates, and attribution statements; (3) Dissemination audit — trace every AI-generated output that reached a U.S. government official or public audience and verify that FARA labeling requirements were met. The Stanford RegLab benchmark suggests that a hallucination rate below 10% on foreign-principal-specific tasks is achievable with current commercial tools, but no tool achieves zero hallucination. Maintain all audit records for at least five years.

References

• U.S. Department of Justice, FARA Annual Report for Fiscal Year 2023 (2024)
• Australian Parliamentary Joint Committee on Intelligence and Security, Review of the Foreign Influence Transparency Scheme (2023)
• Stanford Regulation, Evaluation, and Governance Lab (RegLab), Benchmarking Hallucination Rates in Legal AI Tools (2024)
• Home Office (UK), Guidance on the National Security Act 2023: Foreign Activity Arrangements (2024)
• Canadian Security Intelligence Service, Annual Public Report 2023-2024 (2024)