GSA Issues a Revised Proposed Clause Governing AI Data-Safeguarding Requirements
On June 17, 2026, the General Services Administration (GSA) published a revised proposed contract clause, 552.239-7001, “Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems (LLMs),” which would impose substantial obligations on contractors processing government data through LLMs. This alert follows up on our prior coverage of GSA’s developing rules on the use of AI in GSA contracts.
The rule revises a draft issued on March 6, 2026, in connection with GSA Multiple Award Schedule (MAS) Refresh 31. That earlier draft, titled “Basic Safeguarding of Artificial Intelligence Systems,” drew immediate and significant industry criticism for its broad scope, its prohibition on foreign AI components (including open-source data and research), the unlimited government license it granted, and GSA’s reliance on informal rulemaking. In response, GSA extended the comment period to April 3, 2026, deferred the clause to MAS Refresh 32, and has now published this substantially revised version in the Federal Register. Comments to the proposed rule are due on August 3, 2026.
We begin with what has changed.
A Narrower Scope, But Broader Contractual Reach
The March draft applied whenever AI capabilities were provided to or used by a contractor in performing a GSA contract—a trigger broad enough to sweep in contractors using AI tools in virtually any context. The June version narrows that reach substantially. It now applies only when government data, defined as “Data Inputs” (e.g., user prompts, source data) and "Data Outputs" (e.g., system responses, analyses, metadata and synthetic data), is processed by an LLM, and it expressly excludes (1) LLMs embedded in common commercial products such as word processors or map-navigation systems, and (2) LLMs whose functionality is incidental to the contract’s primary purpose.
At the same time, the clause now reaches further. The March draft, introduced through MAS Refresh 31, was tethered to GSA MAS contracts. The June version amends the GSAR generally so it will apply across all GSA-administered contracts and solicitations, not just MAS contracts, wherever the LLM trigger is met. Any contractor holding a GSA contract should assess whether its use of LLM tools in performance could satisfy the trigger, even where AI is not the subject matter of the contract.
For example, a contractor providing IT, professional, or program-management services might use a third-party LLM internally to prepare deliverables. If government data flows through that tool, the clause may apply even though the contract is not for AI services, and a government employee’s prompt to a contractor-deployed chatbot likely qualifies as government data under the clause’s definitions.
Key Changes from the March Draft
License to Government: “Any Lawful Purpose” Provision Removed
The March draft granted the government an irrevocable license to use the AI system “for any lawful Government purpose,” and it barred the system from refusing to produce outputs based on the contractor’s or service provider’s discretionary policies. Our March alert flagged both as among the most significant practical concerns for contractors using commercial AI platforms with built-in content moderation or safety guardrails. Both are gone in the June version. The government’s license is now limited to the purposes and scope of work defined in the contract or task/delivery order and reaches only the commercially available features and functionality needed to fulfill the contract. 552.239-7001(e)(2).
American AI: Hard Prohibition Softened
The March draft flatly prohibited using foreign AI systems in performance, including any AI components manufactured, developed, or controlled by non-US entities, a bar that reached every component of the system, including foreign-origin open-source model components. The June version instead adopts a “maximize the use of” standard, requiring contractors to maximize their use of LLMs that meet US jurisdiction-and-control criteria. 552.239-7001(f)(2). A new risk-based carve-out permits incidental foreign-developed components, such as open-source components or published research, provided they introduce no foreign-control risk.
Supply Chain Framework: Four Roles Replace Single “Service Provider” Definition
The March draft made a contractor responsible for the compliance of any AI service provider anywhere in the supply chain. It defined “Service Provider” broadly as any entity that directly or indirectly provides, operates, or licenses an AI system, whether or not a subcontractor, and held prime contractors responsible for that entity’s adherence to the entire clause.
The June version replaces that single definition with a four-role framework (LLM Developer, LLM System Operator, LLM System Integrator, and LLM Service Provider) and gives each role its own flowdown supplemental clause (552.239-7001-1 through -4), with obligations calibrated to that role’s function in the LLM supply chain. 552.239-7001(a)(2).
Even with these role-specific flowdowns, the June draft does not fully resolve the burden it places on GSA contractors to ensure compliance by third-party AI providers, including providers that may have no privity with the government and may not even know they are supporting a federal contract.
Unbiased AI: Obligation Strengthened
The Unbiased AI provisions are the one area where the June version is stricter. The March draft required contractors to “make commercial efforts to ensure” the AI system complied with the Unbiased AI Principles; the June version drops that qualifier, so contractors must simply “ensure” compliance.
The substance of the AI Principles is unchanged: the AI system must be truthful, must prioritize historical accuracy and scientific inquiry, and must not manipulate responses in favor of ideological positions. The March draft named Diversity, Equity, and Inclusion as an example of prohibited ideological dogma; the June version drops that specific reference and instead prohibits manipulation in favor of “ideological dogma” generally. 552.239-7001(j)(i). It also enumerates the methods by which such content might be introduced, e.g., training-data selection, fine-tuning, retrieval-augmented generation (RAG) references, system prompts, and other configuration choices, making the prohibition harder to evade. 552.239-7001(j).
The government still retains the right to run unannounced automated assessments at any time using its own benchmarks. 552.239-7001(j)(2). It need not disclose those benchmarks or methodologies to the contractor in advance, and must reveal them only when it cites biased AI as the basis for an adverse action. Contractors also remain potentially liable for reasonable decommissioning costs if terminated for cause for Unbiased AI noncompliance. 552.239-7001(j)(2)(iii)-(iv), (j)(3)(ii). The June version does, however, require the GSA contracting officer to set a predetermined cap on those termination costs, expressed as a percentage of contract value.
These provisions carry significant practical implications for contractors whose AI service providers apply built-in content-moderation or safety policies. With the “commercial efforts” qualifier gone, a contractor can no longer point to a provider’s commercial practices to excuse noncompliance. Contractors should evaluate how their AI systems handle politically and socially sensitive queries and raise the compliance implications with their providers before the clause is finalized.
Contractor Background Data Strengthened
The June version defines “Background Data” to include pre-existing proprietary content, reference materials, knowledge bases, and IP owned or controlled by the contractor that may be incorporated into LLM processing, and it expressly lets the contractor retain ownership of that data in its original form—a protection absent from the March draft. 552.239-7001(b).
Order-Specific Flexibility
The March draft let the parties bilaterally supplement or revise key paragraphs at the order level. The June version contains no equivalent; any adjustment now requires the contracting officer’s written authorization.
For completeness, we summarize the core requirements that did not change from the March draft.
- The government retains ownership of all government data, and the contractor must assign to the government any custom developments or contractor-created intellectual property immediately upon creation. 552.239-7001(e)(1)(iv).
- Prohibited uses of government data include training or improving contractor-provided AI models and informing the contractor’s business operations, including marketing, sales, or strategy. 552.239-7001(e)(3).
- Strict government data handling obligations remain, including limits on human access, logical segregation from non-government customer data, and secure deletion on contract completion. The 72-hour incident-reporting requirement to the Cybersecurity and Infrastructure Security Agency (CISA) and the contracting officer, including daily status updates and 90-day log preservation, is also unchanged. 552.239-7001(e)(4), (f)(5).
- Human-oversight requirements remain, giving the government a means to oversee and trace AI processes. The LLM must be able to summarize its reasoning from input to output and make that record auditable, including where it uses Retrieval-Augmented Generation (RAG) and agentic processes. 552.239-7001(f)(4), (f)(6).
- Contractor notice requirements remain, including:
- Notice to the government of: (1) any modification or configuration of the LLM to comply with non-US federal government statutes, regulations, or policies, due within 30 days after award, 552.239-7001(f)(3); and (2) any planned material change, including adding, replacing, or materially changing any LLMs or service providers, due at least 30 days before the change. 552.239-7001(i)(1); and
- Notice of any change that may materially increase output bias, reduce truthfulness, or raise safety concerns, due within seven days of identification, or as soon as practicable if the change is unplanned or results from a security emergency. 552.239-7001(i)(3), (i)(4).
What Now?
The August 3, 2026, comment deadline is an opportunity to shape the clause before it becomes binding. GSA has shown itself responsive to industry input: the most significant concerns raised in the March comment period are visibly addressed in the June version.
Regardless of whether contractors submit comments, the following steps are advisable now:
- Inventory the LLMs used—or likely to be used—in GSA contract performance, including tools used internally on non-AI contracts, to determine whether the trigger is satisfied.
- Review existing agreements with commercial AI service providers for conflicts with the clause’s data-handling, prohibited-use, and compliance requirements—conflicts the clause resolves in the government’s favor.
- Assess AI supply chains for foreign-origin components and evaluate whether they qualify for the risk-based carve-out.
- Engage AI service providers on Unbiased AI compliance, particularly around built-in content-moderation policies and safety guardrails.
- Document instances where commercially deployed LLMs decline government-relevant queries. Such a record will help in formulating comments and, if compliance questions arise later, will serve as contemporaneous evidence of the technical constraints contractors faced under the clause as drafted.
- Monitor for MAS Refresh 32, which is expected to roll the clause into existing MAS contracts by mass modification.
Conclusion
GSA’s June 17 proposed rule reflects meaningful responsiveness to industry input, but the Unbiased AI obligations, the reach-through to third-party service providers, and the loss of bilateral order-level flexibility remain open issues with real compliance stakes. The August 3 comment deadline is the appropriate forum to address them.
Comments may be submitted at regulations.gov under “Notice-MVAC-2026-01.”
A Narrower Scope, But Broader Contractual Reach
The March draft applied whenever AI capabilities were provided to or used by a contractor in performing a GSA contract—a trigger broad enough to sweep in contractors using AI tools in virtually any context. The June version narrows that reach substantially. It now applies only when government data, defined as “Data Inputs” (e.g., user prompts, source data) and "Data Outputs" (e.g., system responses, analyses, metadata and synthetic data), is processed by an LLM, and it expressly excludes (1) LLMs embedded in common commercial products such as word processors or map-navigation systems, and (2) LLMs whose functionality is incidental to the contract’s primary purpose.
At the same time, the clause now reaches further. The March draft, introduced through MAS Refresh 31, was tethered to GSA MAS contracts. The June version amends the GSAR generally so it will apply across all GSA-administered contracts and solicitations, not just MAS contracts, wherever the LLM trigger is met. Any contractor holding a GSA contract should assess whether its use of LLM tools in performance could satisfy the trigger, even where AI is not the subject matter of the contract.
For example, a contractor providing IT, professional, or program-management services might use a third-party LLM internally to prepare deliverables. If government data flows through that tool, the clause may apply even though the contract is not for AI services, and a government employee’s prompt to a contractor-deployed chatbot likely qualifies as government data under the clause’s definitions.
Key Changes from the March Draft
License to Government: “Any Lawful Purpose” Provision Removed
The March draft granted the government an irrevocable license to use the AI system “for any lawful Government purpose,” and it barred the system from refusing to produce outputs based on the contractor’s or service provider’s discretionary policies. Our March alert flagged both as among the most significant practical concerns for contractors using commercial AI platforms with built-in content moderation or safety guardrails. Both are gone in the June version. The government’s license is now limited to the purposes and scope of work defined in the contract or task/delivery order and reaches only the commercially available features and functionality needed to fulfill the contract. 552.239-7001(e)(2).
American AI: Hard Prohibition Softened
The March draft flatly prohibited using foreign AI systems in performance, including any AI components manufactured, developed, or controlled by non-US entities, a bar that reached every component of the system, including foreign-origin open-source model components. The June version instead adopts a “maximize the use of” standard, requiring contractors to maximize their use of LLMs that meet US jurisdiction-and-control criteria. 552.239-7001(f)(2). A new risk-based carve-out permits incidental foreign-developed components, such as open-source components or published research, provided they introduce no foreign-control risk.
Supply Chain Framework: Four Roles Replace Single “Service Provider” Definition
The March draft made a contractor responsible for the compliance of any AI service provider anywhere in the supply chain. It defined “Service Provider” broadly as any entity that directly or indirectly provides, operates, or licenses an AI system, whether or not a subcontractor, and held prime contractors responsible for that entity’s adherence to the entire clause.
The June version replaces that single definition with a four-role framework (LLM Developer, LLM System Operator, LLM System Integrator, and LLM Service Provider) and gives each role its own flowdown supplemental clause (552.239-7001-1 through -4), with obligations calibrated to that role’s function in the LLM supply chain. 552.239-7001(a)(2).
Even with these role-specific flowdowns, the June draft does not fully resolve the burden it places on GSA contractors to ensure compliance by third-party AI providers, including providers that may have no privity with the government and may not even know they are supporting a federal contract.
Unbiased AI: Obligation Strengthened
The Unbiased AI provisions are the one area where the June version is stricter. The March draft required contractors to “make commercial efforts to ensure” the AI system complied with the Unbiased AI Principles; the June version drops that qualifier, so contractors must simply “ensure” compliance.
The substance of the AI Principles is unchanged: the AI system must be truthful, must prioritize historical accuracy and scientific inquiry, and must not manipulate responses in favor of ideological positions. The March draft named Diversity, Equity, and Inclusion as an example of prohibited ideological dogma; the June version drops that specific reference and instead prohibits manipulation in favor of “ideological dogma” generally. 552.239-7001(j)(i). It also enumerates the methods by which such content might be introduced, e.g., training-data selection, fine-tuning, retrieval-augmented generation (RAG) references, system prompts, and other configuration choices, making the prohibition harder to evade. 552.239-7001(j).
The government still retains the right to run unannounced automated assessments at any time using its own benchmarks. 552.239-7001(j)(2). It need not disclose those benchmarks or methodologies to the contractor in advance, and must reveal them only when it cites biased AI as the basis for an adverse action. Contractors also remain potentially liable for reasonable decommissioning costs if terminated for cause for Unbiased AI noncompliance. 552.239-7001(j)(2)(iii)-(iv), (j)(3)(ii). The June version does, however, require the GSA contracting officer to set a predetermined cap on those termination costs, expressed as a percentage of contract value.
These provisions carry significant practical implications for contractors whose AI service providers apply built-in content-moderation or safety policies. With the “commercial efforts” qualifier gone, a contractor can no longer point to a provider’s commercial practices to excuse noncompliance. Contractors should evaluate how their AI systems handle politically and socially sensitive queries and raise the compliance implications with their providers before the clause is finalized.
Contractor Background Data Strengthened
The June version defines “Background Data” to include pre-existing proprietary content, reference materials, knowledge bases, and IP owned or controlled by the contractor that may be incorporated into LLM processing, and it expressly lets the contractor retain ownership of that data in its original form—a protection absent from the March draft. 552.239-7001(b).
Order-Specific Flexibility
The March draft let the parties bilaterally supplement or revise key paragraphs at the order level. The June version contains no equivalent; any adjustment now requires the contracting officer’s written authorization.
For completeness, we summarize the core requirements that did not change from the March draft.
- The government retains ownership of all government data, and the contractor must assign to the government any custom developments or contractor-created intellectual property immediately upon creation. 552.239-7001(e)(1)(iv).
- Prohibited uses of government data include training or improving contractor-provided AI models and informing the contractor’s business operations, including marketing, sales, or strategy. 552.239-7001(e)(3).
- Strict government data handling obligations remain, including limits on human access, logical segregation from non-government customer data, and secure deletion on contract completion. The 72-hour incident-reporting requirement to the Cybersecurity and Infrastructure Security Agency (CISA) and the contracting officer, including daily status updates and 90-day log preservation, is also unchanged. 552.239-7001(e)(4), (f)(5).
- Human-oversight requirements remain, giving the government a means to oversee and trace AI processes. The LLM must be able to summarize its reasoning from input to output and make that record auditable, including where it uses Retrieval-Augmented Generation (RAG) and agentic processes. 552.239-7001(f)(4), (f)(6).
- Contractor notice requirements remain, including:
- Notice to the government of: (1) any modification or configuration of the LLM to comply with non-US federal government statutes, regulations, or policies, due within 30 days after award, 552.239-7001(f)(3); and (2) any planned material change, including adding, replacing, or materially changing any LLMs or service providers, due at least 30 days before the change. 552.239-7001(i)(1); and
- Notice of any change that may materially increase output bias, reduce truthfulness, or raise safety concerns, due within seven days of identification, or as soon as practicable if the change is unplanned or results from a security emergency. 552.239-7001(i)(3), (i)(4).
What Now?
The August 3, 2026, comment deadline is an opportunity to shape the clause before it becomes binding. GSA has shown itself responsive to industry input: the most significant concerns raised in the March comment period are visibly addressed in the June version.
Regardless of whether contractors submit comments, the following steps are advisable now:
- Inventory the LLMs used—or likely to be used—in GSA contract performance, including tools used internally on non-AI contracts, to determine whether the trigger is satisfied.
- Review existing agreements with commercial AI service providers for conflicts with the clause’s data-handling, prohibited-use, and compliance requirements—conflicts the clause resolves in the government’s favor.
- Assess AI supply chains for foreign-origin components and evaluate whether they qualify for the risk-based carve-out.
- Engage AI service providers on Unbiased AI compliance, particularly around built-in content-moderation policies and safety guardrails.
- Document instances where commercially deployed LLMs decline government-relevant queries. Such a record will help in formulating comments and, if compliance questions arise later, will serve as contemporaneous evidence of the technical constraints contractors faced under the clause as drafted.
- Monitor for MAS Refresh 32, which is expected to roll the clause into existing MAS contracts by mass modification.
Conclusion
GSA’s June 17 proposed rule reflects meaningful responsiveness to industry input, but the Unbiased AI obligations, the reach-through to third-party service providers, and the loss of bilateral order-level flexibility remain open issues with real compliance stakes. The August 3 comment deadline is the appropriate forum to address them.
Comments may be submitted at regulations.gov under “Notice-MVAC-2026-01.”
Related Attorneys
Related Capabilities
© 2026 Jenner & Block LLP. Attorney Advertising. Jenner & Block LLP is an Illinois Limited Liability Partnership including professional corporations. This publication, presentation, or event is not intended to provide legal advice but to provide information on legal matters and/or firm news of interest to our clients and colleagues. Readers or attendees should seek specific legal advice before taking any action with respect to matters mentioned in this publication or at this event. The attorney responsible for this communication is Brent E. Kidwell, Jenner & Block LLP, 353 N. Clark Street, Chicago, IL 60654-3456. Prior results do not guarantee a similar outcome. Jenner & Block London LLP, an affiliate of Jenner & Block LLP, is a limited liability partnership established under the laws of the State of Delaware, USA and is authorised and regulated by the Solicitors Regulation Authority with SRA number 615729. Information regarding the data we collect and the rights you have over your data can be found in our Privacy Notice. For further inquiries, please contact dataprotection@jenner.com.
On June 17, 2026, the General Services Administration (GSA) published a revised proposed contract clause, 552.239-7001, “Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems (LLMs),” which would impose substantial obligations on contractors processing government data through LLMs. This alert follows up on our prior coverage of GSA’s developing rules on the use of AI in GSA contracts.
The rule revises a draft issued on March 6, 2026, in connection with GSA Multiple Award Schedule (MAS) Refresh 31. That earlier draft, titled “Basic Safeguarding of Artificial Intelligence Systems,” drew immediate and significant industry criticism for its broad scope, its prohibition on foreign AI components (including open-source data and research), the unlimited government license it granted, and GSA’s reliance on informal rulemaking. In response, GSA extended the comment period to April 3, 2026, deferred the clause to MAS Refresh 32, and has now published this substantially revised version in the Federal Register. Comments to the proposed rule are due on August 3, 2026.
We begin with what has changed.
A Narrower Scope, But Broader Contractual Reach
The March draft applied whenever AI capabilities were provided to or used by a contractor in performing a GSA contract—a trigger broad enough to sweep in contractors using AI tools in virtually any context. The June version narrows that reach substantially. It now applies only when government data, defined as “Data Inputs” (e.g., user prompts, source data) and "Data Outputs" (e.g., system responses, analyses, metadata and synthetic data), is processed by an LLM, and it expressly excludes (1) LLMs embedded in common commercial products such as word processors or map-navigation systems, and (2) LLMs whose functionality is incidental to the contract’s primary purpose.
At the same time, the clause now reaches further. The March draft, introduced through MAS Refresh 31, was tethered to GSA MAS contracts. The June version amends the GSAR generally so it will apply across all GSA-administered contracts and solicitations, not just MAS contracts, wherever the LLM trigger is met. Any contractor holding a GSA contract should assess whether its use of LLM tools in performance could satisfy the trigger, even where AI is not the subject matter of the contract.
For example, a contractor providing IT, professional, or program-management services might use a third-party LLM internally to prepare deliverables. If government data flows through that tool, the clause may apply even though the contract is not for AI services, and a government employee’s prompt to a contractor-deployed chatbot likely qualifies as government data under the clause’s definitions.
Key Changes from the March Draft
License to Government: “Any Lawful Purpose” Provision Removed
The March draft granted the government an irrevocable license to use the AI system “for any lawful Government purpose,” and it barred the system from refusing to produce outputs based on the contractor’s or service provider’s discretionary policies. Our March alert flagged both as among the most significant practical concerns for contractors using commercial AI platforms with built-in content moderation or safety guardrails. Both are gone in the June version. The government’s license is now limited to the purposes and scope of work defined in the contract or task/delivery order and reaches only the commercially available features and functionality needed to fulfill the contract. 552.239-7001(e)(2).
American AI: Hard Prohibition Softened
The March draft flatly prohibited using foreign AI systems in performance, including any AI components manufactured, developed, or controlled by non-US entities, a bar that reached every component of the system, including foreign-origin open-source model components. The June version instead adopts a “maximize the use of” standard, requiring contractors to maximize their use of LLMs that meet US jurisdiction-and-control criteria. 552.239-7001(f)(2). A new risk-based carve-out permits incidental foreign-developed components, such as open-source components or published research, provided they introduce no foreign-control risk.
Supply Chain Framework: Four Roles Replace Single “Service Provider” Definition
The March draft made a contractor responsible for the compliance of any AI service provider anywhere in the supply chain. It defined “Service Provider” broadly as any entity that directly or indirectly provides, operates, or licenses an AI system, whether or not a subcontractor, and held prime contractors responsible for that entity’s adherence to the entire clause.
The June version replaces that single definition with a four-role framework (LLM Developer, LLM System Operator, LLM System Integrator, and LLM Service Provider) and gives each role its own flowdown supplemental clause (552.239-7001-1 through -4), with obligations calibrated to that role’s function in the LLM supply chain. 552.239-7001(a)(2).
Even with these role-specific flowdowns, the June draft does not fully resolve the burden it places on GSA contractors to ensure compliance by third-party AI providers, including providers that may have no privity with the government and may not even know they are supporting a federal contract.
Unbiased AI: Obligation Strengthened
The Unbiased AI provisions are the one area where the June version is stricter. The March draft required contractors to “make commercial efforts to ensure” the AI system complied with the Unbiased AI Principles; the June version drops that qualifier, so contractors must simply “ensure” compliance.
The substance of the AI Principles is unchanged: the AI system must be truthful, must prioritize historical accuracy and scientific inquiry, and must not manipulate responses in favor of ideological positions. The March draft named Diversity, Equity, and Inclusion as an example of prohibited ideological dogma; the June version drops that specific reference and instead prohibits manipulation in favor of “ideological dogma” generally. 552.239-7001(j)(i). It also enumerates the methods by which such content might be introduced, e.g., training-data selection, fine-tuning, retrieval-augmented generation (RAG) references, system prompts, and other configuration choices, making the prohibition harder to evade. 552.239-7001(j).
The government still retains the right to run unannounced automated assessments at any time using its own benchmarks. 552.239-7001(j)(2). It need not disclose those benchmarks or methodologies to the contractor in advance, and must reveal them only when it cites biased AI as the basis for an adverse action. Contractors also remain potentially liable for reasonable decommissioning costs if terminated for cause for Unbiased AI noncompliance. 552.239-7001(j)(2)(iii)-(iv), (j)(3)(ii). The June version does, however, require the GSA contracting officer to set a predetermined cap on those termination costs, expressed as a percentage of contract value.
These provisions carry significant practical implications for contractors whose AI service providers apply built-in content-moderation or safety policies. With the “commercial efforts” qualifier gone, a contractor can no longer point to a provider’s commercial practices to excuse noncompliance. Contractors should evaluate how their AI systems handle politically and socially sensitive queries and raise the compliance implications with their providers before the clause is finalized.
Contractor Background Data Strengthened
The June version defines “Background Data” to include pre-existing proprietary content, reference materials, knowledge bases, and IP owned or controlled by the contractor that may be incorporated into LLM processing, and it expressly lets the contractor retain ownership of that data in its original form—a protection absent from the March draft. 552.239-7001(b).
Order-Specific Flexibility
The March draft let the parties bilaterally supplement or revise key paragraphs at the order level. The June version contains no equivalent; any adjustment now requires the contracting officer’s written authorization.
For completeness, we summarize the core requirements that did not change from the March draft.
- The government retains ownership of all government data, and the contractor must assign to the government any custom developments or contractor-created intellectual property immediately upon creation. 552.239-7001(e)(1)(iv).
- Prohibited uses of government data include training or improving contractor-provided AI models and informing the contractor’s business operations, including marketing, sales, or strategy. 552.239-7001(e)(3).
- Strict government data handling obligations remain, including limits on human access, logical segregation from non-government customer data, and secure deletion on contract completion. The 72-hour incident-reporting requirement to the Cybersecurity and Infrastructure Security Agency (CISA) and the contracting officer, including daily status updates and 90-day log preservation, is also unchanged. 552.239-7001(e)(4), (f)(5).
- Human-oversight requirements remain, giving the government a means to oversee and trace AI processes. The LLM must be able to summarize its reasoning from input to output and make that record auditable, including where it uses Retrieval-Augmented Generation (RAG) and agentic processes. 552.239-7001(f)(4), (f)(6).
- Contractor notice requirements remain, including:
- Notice to the government of: (1) any modification or configuration of the LLM to comply with non-US federal government statutes, regulations, or policies, due within 30 days after award, 552.239-7001(f)(3); and (2) any planned material change, including adding, replacing, or materially changing any LLMs or service providers, due at least 30 days before the change. 552.239-7001(i)(1); and
- Notice of any change that may materially increase output bias, reduce truthfulness, or raise safety concerns, due within seven days of identification, or as soon as practicable if the change is unplanned or results from a security emergency. 552.239-7001(i)(3), (i)(4).
What Now?
The August 3, 2026, comment deadline is an opportunity to shape the clause before it becomes binding. GSA has shown itself responsive to industry input: the most significant concerns raised in the March comment period are visibly addressed in the June version.
Regardless of whether contractors submit comments, the following steps are advisable now:
- Inventory the LLMs used—or likely to be used—in GSA contract performance, including tools used internally on non-AI contracts, to determine whether the trigger is satisfied.
- Review existing agreements with commercial AI service providers for conflicts with the clause’s data-handling, prohibited-use, and compliance requirements—conflicts the clause resolves in the government’s favor.
- Assess AI supply chains for foreign-origin components and evaluate whether they qualify for the risk-based carve-out.
- Engage AI service providers on Unbiased AI compliance, particularly around built-in content-moderation policies and safety guardrails.
- Document instances where commercially deployed LLMs decline government-relevant queries. Such a record will help in formulating comments and, if compliance questions arise later, will serve as contemporaneous evidence of the technical constraints contractors faced under the clause as drafted.
- Monitor for MAS Refresh 32, which is expected to roll the clause into existing MAS contracts by mass modification.
Conclusion
GSA’s June 17 proposed rule reflects meaningful responsiveness to industry input, but the Unbiased AI obligations, the reach-through to third-party service providers, and the loss of bilateral order-level flexibility remain open issues with real compliance stakes. The August 3 comment deadline is the appropriate forum to address them.
Comments may be submitted at regulations.gov under “Notice-MVAC-2026-01.”
A Narrower Scope, But Broader Contractual Reach
The March draft applied whenever AI capabilities were provided to or used by a contractor in performing a GSA contract—a trigger broad enough to sweep in contractors using AI tools in virtually any context. The June version narrows that reach substantially. It now applies only when government data, defined as “Data Inputs” (e.g., user prompts, source data) and "Data Outputs" (e.g., system responses, analyses, metadata and synthetic data), is processed by an LLM, and it expressly excludes (1) LLMs embedded in common commercial products such as word processors or map-navigation systems, and (2) LLMs whose functionality is incidental to the contract’s primary purpose.
At the same time, the clause now reaches further. The March draft, introduced through MAS Refresh 31, was tethered to GSA MAS contracts. The June version amends the GSAR generally so it will apply across all GSA-administered contracts and solicitations, not just MAS contracts, wherever the LLM trigger is met. Any contractor holding a GSA contract should assess whether its use of LLM tools in performance could satisfy the trigger, even where AI is not the subject matter of the contract.
For example, a contractor providing IT, professional, or program-management services might use a third-party LLM internally to prepare deliverables. If government data flows through that tool, the clause may apply even though the contract is not for AI services, and a government employee’s prompt to a contractor-deployed chatbot likely qualifies as government data under the clause’s definitions.
Key Changes from the March Draft
License to Government: “Any Lawful Purpose” Provision Removed
The March draft granted the government an irrevocable license to use the AI system “for any lawful Government purpose,” and it barred the system from refusing to produce outputs based on the contractor’s or service provider’s discretionary policies. Our March alert flagged both as among the most significant practical concerns for contractors using commercial AI platforms with built-in content moderation or safety guardrails. Both are gone in the June version. The government’s license is now limited to the purposes and scope of work defined in the contract or task/delivery order and reaches only the commercially available features and functionality needed to fulfill the contract. 552.239-7001(e)(2).
American AI: Hard Prohibition Softened
The March draft flatly prohibited using foreign AI systems in performance, including any AI components manufactured, developed, or controlled by non-US entities, a bar that reached every component of the system, including foreign-origin open-source model components. The June version instead adopts a “maximize the use of” standard, requiring contractors to maximize their use of LLMs that meet US jurisdiction-and-control criteria. 552.239-7001(f)(2). A new risk-based carve-out permits incidental foreign-developed components, such as open-source components or published research, provided they introduce no foreign-control risk.
Supply Chain Framework: Four Roles Replace Single “Service Provider” Definition
The March draft made a contractor responsible for the compliance of any AI service provider anywhere in the supply chain. It defined “Service Provider” broadly as any entity that directly or indirectly provides, operates, or licenses an AI system, whether or not a subcontractor, and held prime contractors responsible for that entity’s adherence to the entire clause.
The June version replaces that single definition with a four-role framework (LLM Developer, LLM System Operator, LLM System Integrator, and LLM Service Provider) and gives each role its own flowdown supplemental clause (552.239-7001-1 through -4), with obligations calibrated to that role’s function in the LLM supply chain. 552.239-7001(a)(2).
Even with these role-specific flowdowns, the June draft does not fully resolve the burden it places on GSA contractors to ensure compliance by third-party AI providers, including providers that may have no privity with the government and may not even know they are supporting a federal contract.
Unbiased AI: Obligation Strengthened
The Unbiased AI provisions are the one area where the June version is stricter. The March draft required contractors to “make commercial efforts to ensure” the AI system complied with the Unbiased AI Principles; the June version drops that qualifier, so contractors must simply “ensure” compliance.
The substance of the AI Principles is unchanged: the AI system must be truthful, must prioritize historical accuracy and scientific inquiry, and must not manipulate responses in favor of ideological positions. The March draft named Diversity, Equity, and Inclusion as an example of prohibited ideological dogma; the June version drops that specific reference and instead prohibits manipulation in favor of “ideological dogma” generally. 552.239-7001(j)(i). It also enumerates the methods by which such content might be introduced, e.g., training-data selection, fine-tuning, retrieval-augmented generation (RAG) references, system prompts, and other configuration choices, making the prohibition harder to evade. 552.239-7001(j).
The government still retains the right to run unannounced automated assessments at any time using its own benchmarks. 552.239-7001(j)(2). It need not disclose those benchmarks or methodologies to the contractor in advance, and must reveal them only when it cites biased AI as the basis for an adverse action. Contractors also remain potentially liable for reasonable decommissioning costs if terminated for cause for Unbiased AI noncompliance. 552.239-7001(j)(2)(iii)-(iv), (j)(3)(ii). The June version does, however, require the GSA contracting officer to set a predetermined cap on those termination costs, expressed as a percentage of contract value.
These provisions carry significant practical implications for contractors whose AI service providers apply built-in content-moderation or safety policies. With the “commercial efforts” qualifier gone, a contractor can no longer point to a provider’s commercial practices to excuse noncompliance. Contractors should evaluate how their AI systems handle politically and socially sensitive queries and raise the compliance implications with their providers before the clause is finalized.
Contractor Background Data Strengthened
The June version defines “Background Data” to include pre-existing proprietary content, reference materials, knowledge bases, and IP owned or controlled by the contractor that may be incorporated into LLM processing, and it expressly lets the contractor retain ownership of that data in its original form—a protection absent from the March draft. 552.239-7001(b).
Order-Specific Flexibility
The March draft let the parties bilaterally supplement or revise key paragraphs at the order level. The June version contains no equivalent; any adjustment now requires the contracting officer’s written authorization.
For completeness, we summarize the core requirements that did not change from the March draft.
- The government retains ownership of all government data, and the contractor must assign to the government any custom developments or contractor-created intellectual property immediately upon creation. 552.239-7001(e)(1)(iv).
- Prohibited uses of government data include training or improving contractor-provided AI models and informing the contractor’s business operations, including marketing, sales, or strategy. 552.239-7001(e)(3).
- Strict government data handling obligations remain, including limits on human access, logical segregation from non-government customer data, and secure deletion on contract completion. The 72-hour incident-reporting requirement to the Cybersecurity and Infrastructure Security Agency (CISA) and the contracting officer, including daily status updates and 90-day log preservation, is also unchanged. 552.239-7001(e)(4), (f)(5).
- Human-oversight requirements remain, giving the government a means to oversee and trace AI processes. The LLM must be able to summarize its reasoning from input to output and make that record auditable, including where it uses Retrieval-Augmented Generation (RAG) and agentic processes. 552.239-7001(f)(4), (f)(6).
- Contractor notice requirements remain, including:
- Notice to the government of: (1) any modification or configuration of the LLM to comply with non-US federal government statutes, regulations, or policies, due within 30 days after award, 552.239-7001(f)(3); and (2) any planned material change, including adding, replacing, or materially changing any LLMs or service providers, due at least 30 days before the change. 552.239-7001(i)(1); and
- Notice of any change that may materially increase output bias, reduce truthfulness, or raise safety concerns, due within seven days of identification, or as soon as practicable if the change is unplanned or results from a security emergency. 552.239-7001(i)(3), (i)(4).
What Now?
The August 3, 2026, comment deadline is an opportunity to shape the clause before it becomes binding. GSA has shown itself responsive to industry input: the most significant concerns raised in the March comment period are visibly addressed in the June version.
Regardless of whether contractors submit comments, the following steps are advisable now:
- Inventory the LLMs used—or likely to be used—in GSA contract performance, including tools used internally on non-AI contracts, to determine whether the trigger is satisfied.
- Review existing agreements with commercial AI service providers for conflicts with the clause’s data-handling, prohibited-use, and compliance requirements—conflicts the clause resolves in the government’s favor.
- Assess AI supply chains for foreign-origin components and evaluate whether they qualify for the risk-based carve-out.
- Engage AI service providers on Unbiased AI compliance, particularly around built-in content-moderation policies and safety guardrails.
- Document instances where commercially deployed LLMs decline government-relevant queries. Such a record will help in formulating comments and, if compliance questions arise later, will serve as contemporaneous evidence of the technical constraints contractors faced under the clause as drafted.
- Monitor for MAS Refresh 32, which is expected to roll the clause into existing MAS contracts by mass modification.
Conclusion
GSA’s June 17 proposed rule reflects meaningful responsiveness to industry input, but the Unbiased AI obligations, the reach-through to third-party service providers, and the loss of bilateral order-level flexibility remain open issues with real compliance stakes. The August 3 comment deadline is the appropriate forum to address them.
Comments may be submitted at regulations.gov under “Notice-MVAC-2026-01.”
Related Attorneys
Related Capabilities
© 2026 Jenner & Block LLP. Attorney Advertising. Jenner & Block LLP is an Illinois Limited Liability Partnership including professional corporations. This publication, presentation, or event is not intended to provide legal advice but to provide information on legal matters and/or firm news of interest to our clients and colleagues. Readers or attendees should seek specific legal advice before taking any action with respect to matters mentioned in this publication or at this event. The attorney responsible for this communication is Brent E. Kidwell, Jenner & Block LLP, 353 N. Clark Street, Chicago, IL 60654-3456. Prior results do not guarantee a similar outcome. Jenner & Block London LLP, an affiliate of Jenner & Block LLP, is a limited liability partnership established under the laws of the State of Delaware, USA and is authorised and regulated by the Solicitors Regulation Authority with SRA number 615729. Information regarding the data we collect and the rights you have over your data can be found in our Privacy Notice. For further inquiries, please contact dataprotection@jenner.com.
News and Insights
Podcasts
Partner Laurel Loomis Rimon Discusses Fintech Enforcement, Debanking, and Regulatory Risk on Fintech Layer Cake Podcast
Partner Laurel Loomis Rimon was featured on the Fintech Layer Cake podcast, where she discussed how fintech enforcement and prosecution actually work in practice, and what exposes fintechs and banks to regulatory risk.
July 15, 2026
Event
Partner Suedeen Kelly to Speak at ABA SEER's 34th Fall Conference
On Thursday, October 15, Partner Suedeen Kelly will speak on a panel at the American Bar Association's Section of Environment, Energy, and Resources (ABA SEER) 34th Fall Conference in Albuquerque, New Mexico.
October 15, 2026
Publications
Supreme Court Clarifies Scope of Private Rights of Action Under the Investment Company Act, Private Equity Law Report
Partners Charles Riely, Todd C. Toral, and Martin Glass authored a guest article for Private Equity Law Report examining the US Supreme Court's June 11, 2026, ruling on the scope of private rights of action under the Investment Company Act of 1940.
July 14, 2026
Publications
Emily Loeb Discusses Congressional Oversight Preparedness in Bloomberg Law
Partner Emily Loeb, co-chair of Jenner & Block's Congressional Investigations Practice, spoke with Bloomberg Law article about how companies can prepare for potential oversight exposure ahead of this fall's midterm elections.
July 7, 2026
Publications
In New York Law Journal, The True Lender Doctrine and the OppFi Decision
Partners Jeremy Creelan, Michael Ross, Megan Poetzel, and Laurel Loomis Rimon, and Associate Molly Oberstein-Allen authored an article for the New York Law Journal examining the "True Lender" doctrine in light of a May 2026 California decision that provides the most detailed judicial framework to date for evaluating bank-nonbank lending partnerships.
July 1, 2026