If you are a government contractor offering government agencies products utilizing Large Language Models (LLM), your disclosure requirements just increased.
Per a new memo from the Office of the Management of the Budget (OMB), when procuring LLM’s, government agencies must require vendors to provide sufficient information for the agencies to be able to determine that the product complies with Unbiased A Principles and that it is not in violation of Executive Order 13960, “Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government”.
The goal of the memo is to ensure that purchased models do not run afoul of two Unbiased AI Principles, as per Executive Order 14319, “Preventing Woke Al in the Federal Government”:
(1) Truth-seeking: LLMs must be truthful in responding to user prompts seeking factual information or analysis, and must prioritize historical accuracy, scientific inquiry, and objectivity, and shall acknowledge uncertainty where reliable information is incomplete or contradictory.
(2) Ideological Neutrality: LLMs must be neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas. Developers shall not intentionally encode partisan or ideological judgments into an LLM’ s outputs unless those judgments are prompted by or otherwise readily accessible to the end user.
At minimum, agencies are requires to request from vendors:
- Acceptable Use Policy. A document typically drafted by the original LLM developer to characterize and differentiate appropriate and inappropriate use of their product offering.
- Model, System, and/or Data Cards. These materials from the LLM developer outline all essential information about the model, system, and/or data as it relates to the product offering, including summaries of the training process, identified risks and mitigations, and model evaluation scores on LLM benchmarks.
- End User Resources. Such resources may include product tutorials, developer guides, or other best tools to help customers ensure proper use of the LLM and maximize utility.
- Mechanism For End User Feedback. This can be a general inbox, specific product point of contact, or similar mechanism for providing feedback to the vendor on outputs that violate the Unbiased AI Principles.
Depending on its planned use of a particular LLM, an agency may also require:
- Pre-Training and Post-Training Activities. This can include: (1) Actions undertaken that would impact the factuality and grounding of LLM outputs, (2) System-level prompts that provide natural language instructions to the model specifying guidelines on responding to user-generated queries, (3) The type of outputs restricted via content moderation and safety filters, (3) Use of red teaming, (4) any training done outside the US or (5) Any modifications or configurations undertaken to comply with any regulation from a government other than the U.S. Federal Government.
- Model Evaluations. This can be: results of bias evaluations or benchmark scores for vendor evaluations to measure a model’s bias, helpfulness, honesty, or accuracy
- Enterprise-Level Controls.
- Third Party Modifications. Disclosure of additional controls to modify an LLMs output applied by the vendor if the vendor is not a direct developer of the LLM.
What should government contractors do now:
In gearing up for government procurement involving LLMs, government contractors should consider:
- Improving their own diligence process when acquiring LLMs from the LLM developers and building into the contract the same documentation, transparency and quality assurance requirements that they would now be required to provide to the government.
- Improve their own internal processes for assessing, documenting, monitoring and remediating LLMs, especially in connection with bias.
- Put together, or improve, the process for assessing bias in their LLMs.
With new AI transparency and accountability laws in California, Utah, Maine, and now, a newly signed law in New York, developers and deployers of AI should consider a greater level of transparency, accountability and documentation whether or not they sell to the Federal Government.