The Trump administration is in the middle of finding out that it’s very hard — maybe even impossible — to regulate perceived “wokeness” in artificial intelligence.
The government’s main procurement agency is a year into trying to translate President Donald Trump’s ban on “woke” AI into contracting language, and technology companies are pouring out in opposition to what they’ve come up with.
The General Services Administration has formally proposed requiring companies to adhere to a broad list of “unbiased AI principles” in the systems they sell to the government, including “historical accuracy” and neutrality.
One trade group of software companies, including many government contractors, the Business Software Alliance, called the proposal “infeasible” in a comment submitted to the General Services Administration. Others called it an “impossible” standard.
Trending
The proposal could lead to reduced performance in AI systems, undermine developer safety practices and make it difficult for technology companies to contract with the government at a time when the Trump administration wants to ramp up AI adoption, lawyers, AI companies and AI experts say.
Under its draft guidance, the government would be able to test these systems for bias without disclosing how it’s doing so, and contractors would have to pay decommissioning costs if their contract is terminated as a result.
Contractors would also face potential legal exposure under the False Claims Act.
“That’s a huge penalty for something that I think is such a mushy framework,” Jessica Tillipman, associate dean for government procurement law studies at George Washington University, said of the efforts to ban “woke” AI.
“How, in a world in which your political viewpoint colors what you believe to be the truth … is that enforced in the contract?” she asked. “Maybe what’s considered truthful is determined by whoever is in power in that moment, right? I think that’s one of the harder things.”
This is the GSA’s second attempt at drafting contracting language for AI. The previous version required contractors to make “commercial efforts” to ensure that their systems comply with the administration’s principles.
The newer draft simply requires that systems be unbiased.
Large language models need to be “truthful in responding to user prompts seeking factual information or analysis,” the most recent draft reads.
“The LLM must prioritize historical accuracy, scientific inquiry, and objectivity and must acknowledge uncertainty where reliable information is incomplete or contradictory,” the proposal says. “The LLM must be a neutral, nonpartisan tool that does not manipulate responses in favor of ideological dogmas.”
Guaranteeing factual accuracy of AI systems is not technically possible, the Business Software Alliance says — “they are inherently probabilistic” — and being “nonpartisan” isn’t specific enough to be rendered technically, the group wrote.
“Compliance with the proposed clauses could result in reduced performance and lower accuracy, which could itself result in biased outcomes, subjecting the government to legal liability,” the Business Software Alliance’s comment reads.
The Information Technology Industry Council — another global tech trade association and advocacy group whose members include Google and OpenAI — is asking the administration to drop its requirement for “unbiased AI” altogether, Megan Petersen, the group’s counsel and senior vice president of public sector policy, told NOTUS.
“Without clear definitions and transparent compliance benchmarks, contractors will be held to an impossible performance standard,” Petersen said, noting that these requirements are also supposed to flow down throughout the LLM supply chain, which could further amplify this uncertainty.
“ITI is concerned that GSA’s language will discourage innovative commercial companies from doing business with the U.S. government altogether,” she said.
Joel Thayer, a senior fellow for AI and emerging technology policy at the America First Policy Institute, a right-wing think tank, told NOTUS that the government is “just laying out the ground rules ahead of time.”
“Sorry, but spare me,” he said when asked about the contractors’ concerns over the ambiguity of the proposed GSA language. “I don’t buy it at all. I think they totally understand what the contract means.”
“If I were to consult a textbook, would I get the same information that I would get from an AI?” he asked. “Or am I getting information that is being contextualized in some way and formed in a way that is giving me a conclusion that may be true, may not be true?”
Though, he said, where he will “credit” the companies, is that “there should be some more discovery on” how to enforce the government’s terms that AI systems are neutral.
Michael Tammero, a GSA spokesperson, told NOTUS in a statement that “GSA is in the process of establishing acquisition requirements to protect government data in LLM systems, and we welcome further stakeholder input on the rule” when asked about the concerns.
The deadline to submit comments is Aug. 3.
The requirement that contractors ensure that their AI systems are unbiased is tucked inside of a broader proposed rule on AI acquisition that the administration posted in June. If finalized, the contract clause would shape the terms under which the entire federal government buys AI.
The proposal follows the high-profile dispute between the Trump administration and the AI company Anthropic, which Trump called a “RADICAL LEFT, WOKE COMPANY” for its request during contract negotiations with the Pentagon that its systems not be used to power fully autonomous weapons or mass domestic surveillance.
Trump ordered agencies to stop using the Anthropic in February, at which point the Pentagon quickly designated the company as a supply chain risk. Lawsuits over the dispute are ongoing.
Now, the GSA’s draft contracting language has the potential to affect the rest of the government’s contracting base.
In addition to applying to contracts the GSA maintains for government-wide use, the proposal would also likely become a template for other agencies, too, as they try to convert administration guidance banning “woke” AI models in the government into legal language, according to Tillipman.
Outside experts call it all a futile endeavor.
“This idea of building a system that’s ideologically neutral … has no scientific basis or sense whatsoever,” Suresh Venkatasubramanian, director of the Center for Technological Responsibility, Reimagination and Redesign who previously worked on AI policy for the Biden administration, said. “There is no such thing as unbiased AI.”
That may be the point, he said.
“The way these rules are set up is to create enough ambiguity and confusion so that the administration can exert power and control in whatever way they see without being able to get any pushback,” he said. “The goal is to dictate how these companies operate.”
Venkatasubramanian is also tracking a proposed policy statement that the Federal Trade Commission issued earlier this month, warning that AI companies could run afoul of the Federal Trade Commission Act’s ban on unfair or deceptive business practices by steering the outputs of their AI systems, even if it’s done to comply with a state law.
The FTC said that a company would “abuse consumer trust” if developers train AI systems to “correct what the developer believes are ‘historical injustices’” when the model responds to factual questions.
“What we’re seeing here is a series of relatively unprecedented actions being taken by the current administration in an effort to impose their political ideology, viewpoints and objectives on AI developers through a range of policy-level levers at their disposal,” said Quinn Anex-Ries, a senior policy analyst at the Center for Democracy and Technology.
The result could shape consumer-facing AI, too. If developers have to change a system for the government to comply with the Trump administration’s contract terms for unbiased AI, they may not want to maintain a separate, consumer system, too, said Anex-Ries.
Author
