In December, the Office of Management and Budget (OMB), released a memorandum outlining requirements for implementing two core Unbiased Artificial Intelligence (AI) Principles derived from the executive order (EO), Preventing Woke AI in the Federal Government. The EO, released last July, mandates federal agencies to prioritize truthfulness and ideological neutrality in large language models (LLMs). It also prohibits government agencies from purchasing any LLMs that emphasize ideological agendas—such as diversity, equity, and inclusion (DEI)—over federal accuracy.
[The White House has released a fact sheet for the EO, which can be found here. ]
According to the EO, DEI ideology in LLMs have altered historical figures’ races, refused to celebrate white achievement, and prioritized DEI beliefs over national security safety. To address the alleged suppression, manipulation, and discrimination from ideological AI, the executive order directs the OMB Director to issue guidance to agencies on implementing the following core Unbiased AI Principles:
- Truth-seeking: LLMs shall prioritize truthfulness, historical accuracy, and scientific inquiry, acknowledging uncertainty when necessary.
- Ideological Neutrality: LLMs shall remain neutral and nonpartisan, avoiding ideological bias.
According to the memo, federal agencies procuring LLMs must request specific information from their AI vendor to ensure compliance with Unbiased AI Principles. Agencies should seek transparency regarding a vendor’s acceptable use policies, model systems, and pre- and post-training evaluations. They must refrain from requesting sensitive technical information such as model weights, which are critical in learning patterns from data, making predictions, and determining outputs. By March 11, 2026, all agencies are required to update their procurement policies to ensure LLM contracts adhere to the principles of truth-seeking and ideological neutrality.
The EO states, “Americans will require reliable outputs from AI, but when ideological biases or social agendas are built into AI models, they can distort the quality and accuracy of the output.” However, research has established that implicit biases are an unavoidable reality and, therefore, FABBS questions whether LLMs will unintentionally mirror these biases from developer design choices and training data. FABBS cautions that success in reaching ideological neutrality will depend on scientific rigor and collaboration with behavioral and brain scientists to manage inherent bias, rather than the targeting of ‘woke’ technology and programming.
