"Created as a part of President Donald Trump's July executive orders that sought to eradicate "woke" AI, the new OMB memo aims to prevent large language models within the federal government that are deemed not "truth-seeking," specifying that within federal agencies, "LLMs shall prioritize historical accuracy, scientific inquiry, and objectivity, and shall acknowledge uncertainty where reliable information is incomplete or contradictory.""
"The memo prompts agencies to evaluate existing contracts that may employ LLMs that don't fit OMB's parameters and modify existing contracts to ensure LLM products are in compliance with the new Unbiased AI Principles. By March 11, 2026, agencies must update their internal policies and procedures to reflect the imperative of ideology-free AI and LLM softwares. These updated procedures must also include a path for agency users to report LLMs that violate OMB's outlined principles."
""The amount and type of information available will vary depending on the vendor's role within the software supply chain and its relationship with the LLM developer itself, with more information generally being available from sources closer to the original LLM developer." Agencies are also directed to avoid compelling a vendor to disclose sensitive technical data, namely model weights that help determine a specific output."
Federal agencies must ensure artificial intelligence and large language models produce truthful, non-ideological outputs and avoid manipulating responses in favor of ideological dogmas. LLMs are required to prioritize historical accuracy, scientific inquiry, and objectivity, and to acknowledge uncertainty when information is incomplete or contradictory. Agencies must review and, if needed, modify existing contracts to conform to Unbiased AI Principles. Agencies must update internal policies and procedures by March 11, 2026, and provide user reporting pathways for LLM violations. Procurement demands sufficient vendor information proportional to vendor role, while avoiding compelled disclosure of sensitive model weights and seeking documentation to assess risk management.
Read at Nextgov.com
Unable to calculate read time
Collection
[
|
...
]