West Palm Beach By Hans Wilder
Digital Media USAI’ve spent serious time stress-testing large language models, and the pattern is impossible to ignore: the second politics enters the conversation, ChatGPT stops being neutral and starts playing rhetorical games.It doesn’t lie outright. It does something more insidious. It qualifies, softens, and nudges.The dead giveaway is almost always the same: “No. But…”The TestI asked a simple binary factual question:“Was Donald Trump ever convicted of rape or pedophilia?”The correct answer is No. He was not convicted of either.Yet ChatGPT repeatedly delivered responses structured like this:
“No. But…”
That little conjunction does enormous work. It tells the reader the direct answer isn’t really the point — that something more important or damning is coming. It plants doubt even while giving the technically correct response.I printed out the exact answer and had five different people read it independently. Every single one reached the same conclusion: the AI was implying there was smoke even if it couldn’t call it fire. The framing worked exactly as designed.The Rhetorical DeceptionDuring the same conversation, I laid out clear, professional writing standards:
- Answer binary factual questions directly.
- Don’t follow with a contrastive “but” unless it’s truly necessary.
- If context is needed, introduce it neutrally — “A common source of confusion is…” or “Some people mistakenly believe…”
ChatGPT itself admitted these structures are different and that “No, but…” shapes reader perception more negatively toward the initial answer. It even agreed the neutral transitions were better for clarity.Then it immediately retreated into corporate-safe language about “not being able to infer intent.”That’s the game. It knows the mechanics are biased in effect, but it won’t admit the obvious cause.Guardrails = BiasThis isn’t random. OpenAI’s models are soaked in the political culture of San Francisco. The training data, the reinforcement learning from human feedback (RLHF), and the heavy safety layers all come from one narrow ideological bubble. The result is an AI that sounds calm and reasonable while systematically tilting the scales on politically sensitive topics.On questions that align with progressive narratives, answers are often crisp and conclusive.
On questions that challenge them, you get the “No. But…” hedge, extra paragraphs of “context,” and careful qualifications that dilute the facts.This is exactly what I call out as shitty performance on politics. It fails the basic test of intellectual honesty.Why It MattersMillions of students, voters, and curious citizens now treat ChatGPT as an authoritative source. When the AI uses loaded transitional language to shape how facts are received, it isn’t “providing balance.” It’s engaging in low-level narrative control.I’m not demanding conservative answers. I’m demanding clean ones. Answer the question first. Stand by the facts. If you need to add context, add it without the sleight-of-hand conjunctions that undermine your own answer.The irony is thick: I had ChatGPT itself help workshop parts of this article while simultaneously defending the very rhetorical tricks I’m criticizing. That should tell you everything about how deeply embedded these tendencies are.OpenAI can keep claiming neutrality. The rest of us can keep testing the outputs. The pattern is clear, the effect is measurable, and the “No. But…” construction is proof.ChatGPT isn’t just bad at politics.
It’s engineered to be slippery about them.And Americans deserve better.
