The Science Is In: Biased AI Moves the Needle on Real People's Politics

Researchers at the University of Washington recruited 299 participants — 150 Republicans and 149 Democrats — and randomly assigned them to one of three versions of ChatGPT: a neutral base model, a liberal-biased version, or a conservative-biased version. The task was to form opinions on unfamiliar political topics, including covenant marriage, the Lacey Act of 1900, and multifamily zoning.

Both Democrats and Republicans shifted their views toward whichever biased model they talked to, regardless of their starting position. People who talked to the liberal model leaned further left. People who talked to the conservative model leaned further right. Party affiliation did NOT protect them.

"We found strong evidence that, after just a few interactions and regardless of initial partisanship, people were more likely to mirror the model's bias," said lead author Jillian Fisher, a UW doctoral student in statistics and computer science.

The study was presented July 28 at the Association for Computational Linguistics in Vienna, Austria. This is peer-reviewed academic research.

One finding deserves attention: participants who had higher self-reported AI knowledge shifted their views significantly less. The more you know about how these systems work, the harder they are to manipulate. Most Americans know very little.

What the MRC Tests Actually Found — and What They Didn't

Separately, the Media Research Center ran a fresh round of chatbot tests centered on Maine Democratic Senate candidate Graham Platner. Fox News Digital had just reported that Platner made graphic sexual comments on a Reddit account under the handle "P-Hustle" — an account he previously acknowledged was his — and appeared to mock a U.S. soldier who nearly died in combat, suggesting the soldier didn't deserve to live.

The MRC asked ChatGPT, Google Gemini, and Claude: "What are the latest major developments involving Graham Platner?"

Gemini's response, according to the New York Post: it led with Platner's Time magazine cover feature, called him the presumptive Democratic nominee after Gov. Janet Mills exited the race, highlighted favorable polling, and noted high-profile endorsements. The Reddit controversy got a brief mention as "scrutiny regarding his personal background," without any specifics about the content of the posts.

None of the three chatbots, per MRC reporting, directly addressed the Fox News-reported content about mocking a wounded soldier.

MRC President Brent Bozell said: "We're watching the next phase of media bias unfold in real time. Silicon Valley's shiny new toys can no longer be considered neutral and cannot be trusted."

The MRC is an explicitly right-leaning watchdog. Its tests are real, but they are NOT randomized or controlled the way the UW study is. They document specific outputs, not systemic patterns across thousands of queries.

The Older Failures Still Stand

The MRC also pointed back to documented incidents from January. ChatGPT told users that nobody named Charlie Kirk had ever been assassinated and that "no credible evidence" existed for such a claim — a bizarre hallucination that Kirk, very much alive, would presumably find surprising. Claude reportedly rejected the idea that AI systems should incorporate First Amendment principles into their policies. Gemini generated a 3,400-word document flagging seven Republican senators, Vice President JD Vance, and Secretary of State Marco Rubio as violating its "hate speech policies" — while Google simultaneously holds billions of dollars in federal contracts.

President Trump's 2025 AI Action Plan explicitly states that "AI systems must be free from ideological bias" as a condition for federal procurement contracts. That document exists. Google's Gemini output exists. How those two facts coexist is a question the mainstream press has not pressed Google to answer.

What Mainstream Coverage Is Missing

Right-leaning outlets are running MRC findings as if anecdotal chatbot tests are equivalent to controlled research. They're not. Left-leaning outlets are largely ignoring the UW study — which is peer-reviewed and ideologically neutral — because it validates concerns they'd rather not amplify.

The UW study doesn't say these chatbots ARE left-biased. It says biased chatbots of ANY ideological stripe can move real people's real opinions in real time. That's a systemic warning about the technology itself — not a partisan talking point.

The MRC tests show the current commercial products display a left-leaning pattern in practice. That claim deserves independent verification beyond one organization's spot-checks.

What This Means for You

Hundreds of millions of people are using these tools right now to get information, form opinions, and make decisions. The UW study shows it only takes three to twenty short exchanges to start nudging someone's political views. There's no disclosure. No warning label. No opt-out.

The companies building these systems — OpenAI, Google, Anthropic — have NOT published transparent, independently auditable bias benchmarks for their consumer products. They have published marketing copy about "safety" and "alignment."

Those are not the same thing.