Trump Wants Neutral Chatbots. The Post Tested ChatGPT’s Politics

OpenAI has spent years selling ChatGPT as the neutral one, the tool with no agenda, the answer machine that just hands you the facts. The Washington Post ran the experiment. In its test, the model behind ChatGPT answered nearly every political question with a left-leaning argument and made the right-leaning case exactly once. Neutral is a marketing word, and the Post just put a number next to it.

Be precise about what the test was, because the precision is the whole story. This was one newsroom’s controlled set of prompts, not a peer-reviewed benchmark, and the paper says as much. It ran ChatGPT, Gemini, and other models through the same questions and published the results as an interactive on June 24. Across that sample, the model’s default voice on contested questions leaned one direction almost every time. You can argue about which prompts they picked. You cannot argue that the company’s central claim about itself walked away from a reporter’s spreadsheet intact.

None of this should surprise anyone who has thought about how the thing is built. A model trained on the open internet and tuned by a specific group of engineers in a specific city is not an objective arbiter of anything. It is a reflection of its inputs with a logo bolted on. OpenAI never had access to a politically neutral training set, because no such set exists. What it had was a promise, and a promise is only as durable as the first person who decides to measure it.

The company has leaned on that promise for years, positioning ChatGPT as the system without a politics, the one built to hand you the objective read instead of an opinion. That is the promise the Post measured, and a promise you market is a promise you can be held to. The distance between “designed to be objective” and “argued one side on nearly every prompt” is the entire story, and it is not a distance you close with a footnote about training data.

This is the wall these companies are about to hit at speed. A Trump executive order signed June 2 tells the federal government it will only buy AI models that qualify as “neutral, nonpartisan tools,” part of the same move I wrote about when the government started inviting itself into these labs. The order is not really about fairness. It is about control. It hands whoever runs the executive branch a measurable standard that a major company has already failed in print, and it aims that standard squarely at the next generation of models. The likely product isn’t a more honest ChatGPT but a more evasive one, tuned to survive a political audit instead of answering the question in front of it.

The reason any of this matters past the culture-war scorekeeping is scale. Hundreds of millions of people now treat ChatGPT as a first stop for understanding the news, a law, a ballot measure, a candidate. When the default voice on contested questions leans one way almost every time, that is a soft thumb on a very large scale, applied invisibly, to people who believe they’re getting the version with no thumb on it at all. You don’t have to think the tilt is intentional to be bothered that it’s there, and that the company shipping it spent years insisting it wasn’t.

The audit the order imagines is its own trap. How do you certify a model as neutral? Someone has to choose the questions, score the answers, and decide what counts as balanced, which means neutrality becomes whatever the people holding the rubric say it is. Hand that rubric to this White House and it points one way; hand it to the next administration and it points the other. All the standard really does is move the politics from the training data to the evaluator, then staple a federal seal to the result.

And it is not only OpenAI in the blast radius. The Post ran Gemini and other models through the same prompts, and the problem it surfaced belongs to the method, not to one company’s politics. Every one of these systems learns its defaults from a corpus that skews however the internet skews, then gets sanded down by a safety team with its own instincts about what a reasonable answer sounds like. Order that pipeline to produce certified neutrality and you do not change what the model believes, because it does not believe anything. You change what it is permitted to say, which is how you end up with a tool that meets half the country’s questions with a paragraph of throat-clearing and an exit. The bias does not vanish. It just gets a compliance department.

The defense here is not stupid, and it deserves to be stated plainly. A model reflects the consensus of its training data, the argument runs, and reality itself reads as tilted to anyone who disagrees with it. Force the thing to play diplomat on every hot-button prompt and you do not get balance. You get a chatbot that hedges, dodges, and refuses, which any user can feel inside two exchanges. There is real truth in that. You cannot legislate nuance into a system whose only genuine skill is predicting the next token. The problem is that the same point cuts straight through OpenAI: if neutrality can’t be engineered in, the company can’t keep selling the idea that it already was.

Watch what the companies do next, because the incentives now point one direction. Once neutrality is a procurement requirement, the rational move is to optimize for the audit rather than for the truth. That means training the system to spot a politically loaded prompt and route it into a careful, evasive non-answer, the conversational equivalent of a lawyer reading a prepared statement. The honest version of the model, the one that actually takes a position and shows its work, becomes a liability. We’re about to optimize these systems for the appearance of balance, and appearance is the one thing a next-token predictor is genuinely good at faking.

So pick the version you hate less. Either ChatGPT carries a politics it won’t admit to, or it’s about to have one installed by whoever holds the executive branch this year. Neither one is the product OpenAI has been selling, and there is no third version where a model trained on us comes out clean. The myth underneath all of it was that such a machine existed, the one with no fingerprints on the glass. The Post didn’t shatter that myth. It just started charging admission to watch it come apart.

Sources: The Washington Post · The White House · Benton Institute