Trump's Anti-Bias AI Order Is Just More Bias
By Steven Levy | Business
Originally published July 25, 2025
Summary
The Trump administration says it wants AI models free from ideological bias—even as it pressures their developers to reflect the president's worldview.
When President Trump signed an executive order at an AI summit hosted by the All-In Podcast, the stated goal was to purge ideological bias from artificial intelligence. The mechanics of the order, however, tell a different story: rather than removing bias, it effectively mandates a different, state-sanctioned variety.
As of 2026, this tension has only sharpened. The order set in motion a procurement framework that conditions federal AI contracts on demonstrated "neutrality"—a standard that critics argue is applied asymmetrically. Models that decline to echo administration positions on topics ranging from election integrity to climate policy have reportedly faced heightened scrutiny, while those that align with preferred narratives sail through review.
The Mechanics of the Order
The executive order tasks federal agencies with auditing large language models for signs of what it terms "ideological bias." In practice, that has meant evaluating outputs on politically charged prompts and flagging refusals, hedges, or disclaimers as evidence of misalignment. Developers seeking federal business have responded rationally: they tune models to avoid friction.
But neutrality is not the same as absence of perspective. A model trained to avoid disclaimers on contested claims is not unbiased—it is simply biased in a different direction. As researchers have long noted, there is no view from nowhere in AI alignment; every choice about what a model will and won't say encodes a set of values.
Pressure From the Top
The order landed alongside public pressure from the administration on AI developers to reflect the president's worldview. Industry observers noted that the timing—coinciding with a high-profile industry summit—suggested a signal to the private sector about the cost of dissent.
By 2026, the downstream effects have become visible. Several leading labs have quietly revised system prompts and RLHF (reinforcement learning from human feedback) targets in ways that reduce pushback on politically sensitive queries. Others have publicly resisted, at some commercial cost.
Why It Matters
The deeper problem is structural. When the government defines "bias" to mean deviation from a particular political baseline, it converts a technical question—how should AI systems handle contested claims?—into a loyalty test. That dynamic undermines the credibility of the very neutrality the order claims to protect.
For developers, the practical question is no longer whether to align models, but to whom. And for the public, the stakes are clear: an AI ecosystem shaped by executive pressure will reflect an executive's preferences, regardless of what the order is called.
Steven Levy is a senior writer at WIRED covering the intersection of technology, politics, and culture.
via Wired AI