Skip to content

Menu

Network by SubjectChannelsBlogsHomeAboutContact
AI Legal Journal logo
Subscribe
Search
Close
PublishersBlogsNetwork by SubjectChannels
Subscribe

GSA Issues Final Rule on Large Language Model Procurements

By Sharmistha Das, Matthew Ferraro, Kate Growley & Rajeev Raghavan on October 6, 2026
Email this postTweet this postLike this postShare this post on LinkedIn

Background

On September 28, 2026, the General Services Administration (GSA) issued a new clause for incorporation into government contracts of artificial intelligence systems, 552.239-7001, Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems, as a GSA Regulation Deviation (the Final Clause). The Final Clause will become effective on October 19, 2026.

GSA’s efforts to regulate contractor handling of government information processed by large language models (LLM) have moved through three phases. An initial draft released in March 2026 drew substantial industry criticism for its breadth, including an outright prohibition on AI components “manufactured, developed, or controlled by non-U.S. entities.” GSA responded with a revised version published for public comment on June 17, 2026 (the June Proposed Clause), which removed the non-U.S. component prohibition, introduced role-based supply-chain definitions, and restructured IP protections.

The Final Clause contains several notable changes over the June Proposed Clause, which are summarized below. Contractors should assess the impact of these provisions on their systems in anticipation of the incorporation of the clause into GSA LLM procurements. The clause is effective October 19; however, GSA contracting officers could exercise discretion to incorporate the clause in advance of that date. 

Narrowed Applicability Threshold

The June Proposed Clause applied whenever “Government Data will be processed by a Large Language Model Artificial Intelligence System,” potentially capturing contracts where a contractor’s systems merely incidentally handled Government Data using an LLM. 

The Final Clause instead imposes a two-part test for applicability: (1) the Government must be specifically procuring a system where LLM functionality is a “material feature,” and (2) Government Data must be “submitted directly to or produced by the LLM.” The Final Clause does not apply (and is self-deleting) in the event of two conditions: 1

  • LLM use is limited to the contractor’s “internal business, back-office, operational, or performance-support systems” that the Government is not buying directly, and thus are not “accessed by the Government,” even if the internal use means the LLMs may “incidentally” process Government Data.
  • LLM functionality is “incidental or ancillary” to the commercial products or services the Government is procuring. 

Corresponding definitional changes further narrow the applicability of the Final Clause. The Final Clause clarifies and, in some cases, limits the definitions of “Data Inputs” and “Data Outputs,” which are the two types of data that comprise Government Data subject to the clause. The definition of “Data Inputs” is narrowed to cover data submitted “by or on behalf of” the Government rather than the June Proposed Clause’s applicability to data “created for” the Government. Metadata and logs are excluded from the definition of “Data Outputs.” Collectively, these changes materially reduce the scope of “Government Data” and thus the activities and LLM uses regulated by the clause.

Flow-down Framework Restructured

The June Proposed Clause had proposed a mandatory, role-based flow-down structure built around four defined supply-chain roles with obligations flowing through four companion clauses attached to each role: LLM Developer, LLM System Operator, LLM System Integrator, and LLM Service Provider.

The Final Clause replaces this taxonomy with a principles-based approach tied to the National Institute of Standards and Technology (NIST) AI Risk Management Framework 1.0, Appendix A task categories (AI Design, AI Development, AI Deployment, and Operation and Monitoring). The Final Clause now requires flow-down to any subcontractor performing those tasks to the extent it “collects, processes, stores, retains, trains on, fine-tunes on, or otherwise handles Government Data.” Notably, the Final Clause does not require contractors to flow down clause provisions that “impose requirements concerning the origin, ownership, jurisdiction, or foreign control” to a “Fully Open Model” or “Open-LLM Component” (both of which are newly defined in the Final Clause).

Despite pivoting to the NIST-based taxonomy noted above, the Final Clause still references the four supply chain roles from the June Proposed Clause, but the Final Clause no longer defines those roles or includes their associated companion clauses. It is unclear whether GSA will issue revised companion clauses.

Order of Precedence: Final Clause Still Prevails Over Commercial Terms

A major industry concern with the June Proposed Clause was its potential incompatibility with the commercial acquisition framework under FAR Part 12, which seeks to align government procurement with standard commercial terms and practices. By placing its requirements in the schedule of supplies/services tier — the highest position in the FAR 52.212-4 order of precedence — and overriding all contractor commercial agreements, the June Proposed Clause created potential procurement roadblocks for LLM suppliers, many of whom were expecting to rely heavily on their commercial terms in contracting under GSA vehicles. 

The Final Clause clarifies that the clause “supplements, and does not replace or modify, FAR 52.212-4 or GSAR 552.212-4” and adds a savings clause for “applicable law, regulation, or mandatory Government contract clauses.” However, the Final Clause leaves unchanged the operative rule that the clause overrides all contractor commercial agreements, and industry concerns regarding the order of precedence will likely persist.

Expanded IP Protections for Contractors

The Final Clause makes three targeted changes to IP-related provisions that should increase IP protections for contractors subject to the clause, on balance.

  1. New preexisting materials acknowledgment. The Final Clause addresses concerns about contractors’ pre-existing IP by adding a new provision expressly acknowledging that the Government does not acquire ownership in the contractor’s preexisting commercial products, proprietary technology, or materials developed for use across multiple customers when such items are used or enhanced in performance of the contract. The Final Clause specifically lists “software, tools, configurations, workflows, documentation, knowledge base articles, models, scripts, service-generated data, derivative data, analytics, technical methods, know-how, and other materials” as the type of materials that could be used or enhanced without transfer of rights. 
  2. IP assignment carve-out for general capability gains. The June Proposed Clause broadly required assignment to the Government of any IP rights in “improvements, enhancements, feedback, or derivative works” derived from Government Data. The Final Clause adds an express carveout: the assignment does not apply to “general capability gains or generalized improvements that do not incorporate, derive from, reveal, or disclose Government Data.”
  3. Expanded Background Data ownership. The June Proposed Clause allowed contractors to retain ownership of Background Data “in its original form.” The Final Clause deletes the phrase “in its original form” from its text, meaning contractors should retain ownership of Background Data even if modified or enhanced during performance. The Final Clause also expands the Background Data definition to cover data licensed by the contractor (adding to “owned or controlled”) and expressly includes information used to “develop or improve” the LLM.

Incident Reporting Narrowed; Duplicative Reporting Reduced

The June Proposed Clause proposed mandatory notification to the Government within 72 hours of “any incident … affecting any contractors, including Third-party entities, handling Government Data,” potentially requiring notice of supply-chain incidents with no direct impact on Government Data. The Final Clause narrows the trigger to incidents “affecting an LLM  used in the performance of the contract … that may affect the confidentiality, integrity, or availability of Government Data” (emphasis added).

Additionally, for FedRAMP-authorized systems, incident reports, continuous monitoring submissions, and related reports submitted pursuant to FedRAMP, CISA, or other applicable federal cybersecurity reporting requirements now satisfy the clause’s notification obligation to the extent they contain substantially the same required information and are concurrently provided to the contracting officer. This safe harbor was entirely absent from the June Proposed Clause.

“Unbiased AI Principles” Standard Substantially Relaxed

The June Proposed Clause proposed detailed mandatory obligations requiring contractors to ensure the LLM was “truthful,” was a “neutral, nonpartisan tool” free of “ideological dogmas,” and contained an express prohibition on introducing partisan or ideological bias through training data, fine-tuning, retrieval augmented generation references, system prompts, or other configuration methods.

The Final Clause replaces this entire framework with a single obligation: Contractors must “use reasonable efforts to design, train, and configure the LLM to respond to user prompts seeking factual information or analysis, prioritize accuracy, scientific inquiry, and objectivity.” The Final Clause removes the express prohibition on embedding partisan or ideological judgments and the continuous improvement/monitoring requirement.

Non-Compliance Remedies: Suspension Right Expanded, Liability Cap Fixed

The Final Clause gives the Government the right to suspend use of the LLM “at any time” without qualification, whereas the June Proposed Clause would have allowed suspension “until performance issues are satisfactorily addressed.”

The June Proposed Clause proposed to make the contractor responsible for decommissioning costs if the contractor failed to comply with the Unbiased AI Principles. The Final Clause broadens this decommissioning liability to make the contractor liable for decommissioning costs based on termination for any notice of noncompliance with the clause. However, the Final Clause caps contractor liability for decommissioning costs at 25% of the value of the affected task or delivery order, with re-procurement costs and replacement-system development expressly excluded from that calculation.

Key Considerations for Contractors

In light of the Final Clause, contractors with existing or anticipated GSA LLM procurements should consider whether:

  • Current LLM deployments meet the narrowed applicability threshold, and whether the self-deletion provisions apply to any tools currently within scope.
  • Existing Background Data definitions or equivalent definitions in commercial agreements align with the Final Clause’s expanded definition, including the “licensed by” formulation and the coverage of materials used to develop or improve the LLM.
  • Existing commercial and government-facing LLM terms and procurement strategy can accommodate the Final Clause’s order of precedence directive.
  • Internal governance frameworks require updating to address the practical implications of the relaxed Unbiased AI Principles and reflect a reasonable-efforts rather than zero-tolerance approach.
  • Incident reporting workflows, particularly for FedRAMP-authorized systems, should be updated to take advantage of the new FedRAMP/CISA safe harbor and eliminate duplicative reporting.

[1] The Final Clause does not separate these conditions with an “and” or an “or,” leaving it unclear whether they both must be satisfied for the clause to be self-deleting or if either condition, if met, suffices to treat the clause as self-deleting.

Photo of Sharmistha Das Sharmistha Das

Sharmi Das’ experience at the Department of Homeland Security (DHS), Department of Justice (DOJ), the White House, the U.S. Senate, and private practice positions her to guide clients through regulatory challenges, government-facing issues, and scrutiny from Congress and other oversight bodies. Sharmi has

…

Sharmi Das’ experience at the Department of Homeland Security (DHS), Department of Justice (DOJ), the White House, the U.S. Senate, and private practice positions her to guide clients through regulatory challenges, government-facing issues, and scrutiny from Congress and other oversight bodies. Sharmi has handled dozens of congressional inquiries and managed a program that developed hundreds of regulatory actions relating to homeland security matters, including technology, cybersecurity, contracts and grants, intelligence, health, and immigration. She participated in hundreds of policy discussions at the White House and DHS on high-profile issues that were often in the headlines, including domestic and international crises and emergencies.

Sharmi brings over a decade of experience analyzing statutory and regulatory text to both challenge and defend agency actions in litigation. She uses her knowledge of the Administrative Procedure Act (APA) and federal rulemaking process to help clients shape regulatory authorities, comply with them, and challenge them. In both the executive and legislative branches, Sharmi crafted strategies to resolve inquiries from the Hill, other federal oversight bodies, and the public, often under immense public scrutiny.

Sharmi was awarded the Distinguished Service Medal by the Secretary of Homeland Security for exceptional service, the Department’s highest civilian honor. She was also recognized as an honorary Judge Advocate General by the U.S. Coast Guard. Sharmi values public service and maintains a diverse pro bono practice, including her time seconded to the Legal Aid Society of the District of Columbia.

Read more about Sharmistha Das
Show more Show less
Photo of Matthew Ferraro Matthew Ferraro

Matthew F. Ferraro is a partner in Crowell & Moring’s Privacy and Cybersecurity Group, where he helps clients address complex regulatory matters at the intersection of advanced technology, national security, and crisis management. He advises leading organizations on high-impact matters related to artificial

…

Matthew F. Ferraro is a partner in Crowell & Moring’s Privacy and Cybersecurity Group, where he helps clients address complex regulatory matters at the intersection of advanced technology, national security, and crisis management. He advises leading organizations on high-impact matters related to artificial intelligence (AI) and other emerging technologies, cyberattacks, domestic and international privacy compliance, internal investigations, foreign direct investment reviews, and high-stakes crises.

Before joining the firm, Matthew served as the Senior Counselor for Cybersecurity and Emerging Technology to the Secretary of Homeland Security. As a principal advisor to the Secretary and a member of the U.S. Department of Homeland Security’s leadership team, he served at the heart of U.S. government policymaking around AI and cybersecurity. He assisted in the development and drafting of key AI, cyber, and technology policies and regulations; advised on the deployment of AI to fulfill the department’s missions; and counseled on cyber-incident responses and investigations. Matthew also helped establish and served as the Executive Director of the Artificial Intelligence Safety and Security Board, a flagship public-private advisory committee focused on AI’s use in critical infrastructure and chaired by the Secretary and composed of industry, nonprofit, and government luminaries.

Read more about Matthew Ferraro
Show more Show less
Photo of Kate Growley Kate Growley

Kate M. Growley (CIPP/US, CIPP/G) is a director with Crowell & Moring International and based in Hong Kong. Drawing from over a decade of experience as a practicing attorney in the United States, Kate helps her clients understand, navigate, and shape the policy…

Kate M. Growley (CIPP/US, CIPP/G) is a director with Crowell & Moring International and based in Hong Kong. Drawing from over a decade of experience as a practicing attorney in the United States, Kate helps her clients understand, navigate, and shape the policy and regulatory environment for some of the most complex data issues facing multinational companies, including cybersecurity, privacy, and digital transformation. Kate has worked with clients across every major sector, with particular experience in technology, health care, manufacturing, and aerospace and defense. Kate is a Certified Information Privacy Professional (CIPP) in both the U.S. private and government sectors by the International Association of Privacy Professionals (IAPP). She is also a Registered Practitioner with the U.S. Cybersecurity Maturity Model Certification (CMMC) Cyber Accreditation Body (AB).

Read more about Kate Growley
Show more Show less
Photo of Rajeev Raghavan Rajeev Raghavan

Rajeev Raghavan is a former special counsel to the Director of the Federal Bureau of Investigation (FBI) and federal prosecutor who guides clients through high-stakes cybersecurity and privacy incidents, government investigations and enforcement actions, sensitive national security matters, and litigation.

Read more about Rajeev Raghavan
  • Posted in:
    Administrative, Corporate Compliance
  • Blog:
    Government Contracts Legal Forum
  • Organization:
    Crowell & Moring LLP
  • Article: View Original Source

LexBlog logo
Copyright © 2026, LexBlog. All Rights Reserved.
Legal content Portal by LexBlog LexBlog Logo