AI News
20 Dec 2025
Read 9 min
2025 federal LLM procurement guidance: How to comply
2025 federal LLM procurement guidance helps agencies update contracts to ensure truthful auditable AI.
New OMB rules now set how agencies should buy and use large language models. The 2025 federal LLM procurement guidance requires truth-seeking, neutral outputs, and transparent documentation. Agencies have 90 days to update policies and add new contract clauses by March 11. Vendors must show risk controls without exposing sensitive model details.
The General Services Administration has already made access cheap and easy, with 43 agencies signed up and enterprise licenses for frontier models at $1 or less. Now comes the governance. The new memo covers new and existing contracts, plus commercial and internally built tools, and it expects proof of sound risk management.
90 days to update acquisition policy from the Dec. 11 memo (targeting March 11)
Applies to new awards and existing agreements
Covers commercial, third-party, and internally developed LLMs
Truth-seeking: prioritize factual accuracy, science, and history; acknowledge uncertainty
Ideological neutrality: avoid partisan or ideological bias in outputs unless the user prompts it
Accountability: show how the model and system manage risk and oversight
Do not demand protected technical data (for example, model weights)
Request enough documentation to evaluate risk at the model, system, or application level
Minimum transparency: request four data sets, including the acceptable use policy and model/system/data details relevant to the LLM
Documentation depth varies by the vendor’s proximity to the original developer
When buying through third parties, the developer’s cooperation affects available information
Agencies should tailor requirements to the vendor’s role while ensuring compliance
Update policy and clauses by March 11: add truth-seeking and neutrality requirements; require risk documentation and transparency packages
Clarify data use and IP: define how government data can train, fine-tune, or improve models; set ownership and derivative rights
Stand up governance: set metrics, decision gates, and change-control; log model versions and prompts; require human oversight for high-impact use
Test before scale: pilot with GSA vehicles (including USAi) and enterprise licenses; use red-teaming and domain test sets; record results
Plan for bias and incidents: require bias mitigation plans, incident reporting, model rollback, and user redress channels
Define termination triggers: document what counts as convenience vs. cause when outputs fail accuracy or neutrality benchmarks
Train the workforce: build templates, checklists, and a vendor due diligence playbook; share lessons through a center-of-excellence
Provide an acceptable use policy, model/system/data documentation, and safety policies
Show provenance: summarize training data sources at a high level, fine-tuning datasets, and data governance controls
Share evaluations: publish bias, safety, and factual accuracy results; note uncertainty handling
Enable oversight: deliver logs, dashboards, and change histories; supply a secure mechanism for audit without exposing model weights
Commit to updates: maintain a change log, notify agencies of major updates, and support regression testing
Offer remedies: outline how you address harmful or biased outputs and timelines to fix
The LLM must demonstrate predefined accuracy rates on agency test sets and public benchmarks
The LLM must flag uncertainty when information is incomplete or contradictory
The LLM must avoid partisan or ideological persuasion unless explicitly prompted by the user
Provide a transparency package that includes: acceptable use policy, model card, system architecture overview, data governance summary, safety evaluations, and change log
Provide an audit support plan that enables assessment without disclosing model weights
Notify the agency before material model changes; support re-testing
Report incidents within a defined time; provide remediation steps and post-incident review
Factual accuracy: measure correctness across curated government content and domain tasks
Neutrality: test for political or ideological skew with balanced prompt sets
Uncertainty handling: check for proper caveats where evidence is mixed or missing
Safety and misuse: enforce acceptable use, rate-limit risky prompts, and block prohibited content
Define success thresholds before award and during pilot
Track drift: schedule periodic re-tests and compare against baseline
Log prompts and outputs for audit with privacy protections
Leverage GSA’s agreements for affordable pilots and quick access to frontier models
Work with your CIO, CISO, privacy office, and counsel to align with prior OMB AI memos on data and IP
Join cross-agency communities and centers-of-excellence to share templates, test sets, and lessons learned
Act by March 11: update clauses and acquisition policy within 90 days
Require truth, neutrality, and clear documentation without asking for model weights
Match requirements to the vendor’s role in the supply chain
Test early with pilots, measure performance, and log changes
Build a due diligence playbook and train the workforce
The 2025 federal LLM procurement guidance is your path to safe, useful AI in government. Move quickly, set clear metrics, and hold vendors to transparent, neutral, and truthful outputs to earn trust and scale what works.
What changed in the 2025 federal LLM procurement guidance
Deadlines and scope
Core principles to build into contracts
Transparency without sensitive disclosures
Supply chain nuance
Practical steps to comply and move fast
For agencies
For vendors
Contract language ideas you can adapt
Truth and neutrality performance
Transparency deliverables
Risk, incidents, and updates
Testing and evaluation that matches the policy
Build a simple, repeatable test harness
Operate with metrics
Where to get help now
Use existing government channels
2025 federal LLM procurement guidance: key takeaways
For more news: Click Here
FAQ
Q: What is the 2025 federal LLM procurement guidance?
A: The 2025 federal LLM procurement guidance is an OMB memo that sets procurement guardrails for buying and using large language models, emphasizing truth-seeking, ideological neutrality, and transparent documentation. It directs agencies to update acquisition policies and contract clauses to ensure LLMs meet those principles.
Q: Which contracts and systems does the guidance cover?
A: The guidance applies to new awards and existing agreements and covers commercial, third-party, and internally developed LLMs. The 2025 federal LLM procurement guidance asks agencies to apply requirements across the software supply chain based on the vendor’s role.
Q: What core principles must be built into contracts under the guidance?
A: Contracts must require LLMs to be truthful, prioritize historical and scientific accuracy, acknowledge uncertainty, and remain ideologically neutral unless explicitly prompted by users. The 2025 federal LLM procurement guidance also requires agencies to verify vendor risk management and accountability measures.
Q: What deadlines do agencies need to meet to comply?
A: Agencies have 90 days from the Dec. 11 memo to update acquisition policies, with new clauses and contract changes expected by March 11. The 2025 federal LLM procurement guidance sets that timeline for agencies to implement truth-seeking and neutrality requirements.
Q: What documentation should agencies request from vendors to assess compliance?
A: Agencies should request enough documentation to assess risk at the model, system, or application level without demanding sensitive technical data like model weights, and the guidance specifies a minimum transparency package of four data sets including an acceptable use policy and model/system/data details. The 2025 federal LLM procurement guidance advises tailoring documentation requests to what is reasonably available from the vendor’s position in the supply chain.
Q: How should agencies approach procurement through third-party vendors or resellers?
A: The memo notes that the amount and type of information available will vary with a vendor’s proximity to the original LLM developer, so agencies should calibrate requirements and expectations accordingly. Under the 2025 federal LLM procurement guidance agencies should consider developer cooperation and supply-chain nuance when transacting through third parties.
Q: What practical steps can agencies take now to comply quickly?
A: Agencies should update policy and clauses by the March 11 target, set governance metrics and change-control, log model versions and prompts, pilot with GSA vehicles, and plan for bias mitigation and incident response. The 2025 federal LLM procurement guidance also recommends training the acquisition workforce, creating due diligence playbooks, and using centers-of-excellence to share templates and test sets.
Q: What should vendors provide to meet the guidance’s requirements?
A: Vendors should supply an acceptable use policy, model and system documentation, high-level provenance summaries of training data, safety and bias evaluations, logs or audit support mechanisms, and a change log with notification procedures. The 2025 federal LLM procurement guidance expects vendors to enable oversight and remediation without disclosing protected technical details such as model weights.
Contents
