President Trump’s Battle Against “Woke AI” Poses a Threat to Civil Liberties
The recently-launched “AI Action Plan” from the White House takes aim at what it labels “woke AI,” targeting large language models (LLMs) that present information at odds with the administration’s stances on climate change, gender, and other matters. It also addresses initiatives aimed at reducing the generation of biased content related to race and gender, alongside hate speech. The replication of such biases presents a serious challenge that AI developers have faced for more than ten years.
A new executive order titled “Preventing Woke AI in the Federal Government,” issued alongside the AI Action Plan, aims to coerce AI companies into adjusting their models to align with the Trump Administration’s ideological objectives.
This executive order mandates that AI firms receiving federal contracts demonstrate that their LLMs are devoid of alleged “ideological biases” relating to “diversity, equity, and inclusion.” This heavy-handed approach to censorship will not enhance the accuracy or “trustworthiness” of the models, as the Trump Administration asserts, but rather represents a clear effort to stifle the evolution of LLMs and limit them as instruments of expression and information access. While the First Amendment allows the government to choose to procure only services that align with its viewpoints, it cannot wield that authority to manipulate the services and information accessible to the public. Profitable federal contracts could compel private companies to embed features (or biases) they otherwise wouldn’t, which ultimately affects the end-users. This would impact the 60 percent of Americans relying on LLMs for information, and it would force developers to reverse efforts to reduce bias—resulting in less accurate models likely to cause harm, particularly when deployed by the government.
Reduced Accuracy, Increased Bias and Discrimination
It is well-known that AI models—including generative AI—often discriminate against racial and gender minorities. AI models utilize machine learning to identify and replicate patterns in the data on which they are trained. If the training data contains biases against racial, ethnic, and gender minorities—which it frequently does—then the AI model is likely to “learn” to discriminate against those groups. In simpler terms, garbage in, garbage out. Moreover, these models often reflect the biases of the individuals who train, test, and evaluate them.
This issue spans various AI applications. For instance, “predictive policing” tools, which rely on arrest data reflecting over-policing in black neighborhoods, often recommend increased policing in those areas, frequently based on flawed predictions of crime. Generative AI models share similar concerns. LLMs already suggest harsher sentences, higher rates of criminal convictions, and lower-status job opportunities for individuals from marginalized communities. Even though people of color comprise less than half of the U.S. prison population, 80 percent of images generated by Stable Diffusion depict inmates with darker skin. Similarly, over 90 percent of AI-generated images of judges are male, in contrast to the reality of 34 percent being female judges.
These models not only exhibit bias but are fundamentally flawed. Race and gender should not serve as objective criteria for hiring or sentencing decisions. The discriminatory outcomes stem from patterns in the training data, which may reflect bias or randomness—not any “objective” truth. Setting aside fairness, biased models are simply inferior: they yield more frequent errors. Alleviating bias-related mistakes would ultimately enhance the models’ accuracy, not degrade it.
Biased LLMs Pose Significant Risks—Especially When Used by the Government
However, inaccuracy is not the only concern. When government entities employ biased AI for decision-making, it adversely affects real lives. Government officials frequently make choices that influence individuals’ freedoms, financial access, healthcare, housing, and more. The White House’s AI Action Plan advocates for a substantial uptick in the use of LLMs and other AI by agencies—while practically mandating the use of biased models that perpetuate systemic and historical injustices. Utilizing AI merely to reinforce existing systems squanders the potential of this emerging technology.
We necessitate robust safeguards to prevent government agencies from acquiring biased and harmful AI tools. Through a series of executive orders and the AI Action Plan, the Trump Administration has dismantled already inadequate safeguards from the Biden era. This raises the likelihood of AI-fueled civil rights violations, jeopardizing everyone’s rights.
Furthermore, the Administration could exploit the new regulations to pressure companies into diluting publicly available models. Businesses such as healthcare providers and landlords increasingly rely on AI for significant decisions affecting individuals, meaning that growing bias in commercial models could also result in harm.
We have advocated against utilizing machine learning for predictive policing decisions and other punitive judgments for these very reasons and remain committed to safeguarding your right to avoid biased government determinations shaped by machine learning.
