EPIC, Coalition Urge GSA to Remove ‘Unbiased AI’ Requirements From Federal AI Contracts

On August 3, EPIC, the Center for Democracy & Technology, the Electronic Frontier Foundation, and Upturn submitted comments to the General Services Administration urging the agency to remove ill-defined “unbiased AI principles” requirements from its updated draft terms and conditions for federal contracts involving large language models. 

GSA has substantially improved the draft since releasing the initial version, about which the coalition also submitted comments. Improvements include the removal of a provision that would have required AI systems to produce outputs for “all lawful purposes” and the addition of important safeguards for government data, privacy, confidentiality, and interoperability. 

However, the latest draft retains an “unbiased AI principles” provision that the coalition warns is technically infeasible, poorly defined, and vulnerable to political misuse. The provision invokes concepts including “historical accuracy,” “scientific inquiry,” and “objectivity” without explaining how developers should implement or measure those requirements. EPIC and its coalition partners argue that requiring AI systems to adhere to undefined standards of “ideological neutrality” could undermine the accuracy, reliability, and safety of models procured by the federal government. 

The draft would also allow the government to evaluate AI systems using its own benchmarks to determine compliance with “unbiased AI principles.” Because the underlying requirements are undefined, this could give the government significant discretion to determine whether a model is compliant. The coalition warns that this creates opportunities for political weaponization and could pressure developers to modify their models in response to the political objectives of a particular administration. Noncompliance could also serve as grounds for terminating a contract, leaving contractors responsible for decommissioning costs. 

EPIC and its coalition partners urge GSA to preserve the draft’s strong privacy and security protections while removing the “unbiased AI principles” provision.