I’ve lived long enough to know that when a pinball machine starts flashing TILT, it’s not being political — it’s telling the truth. Somebody leaned on the table. And today’s AI platforms are flashing TILT in bright letters. Not because the machines “decided” anything. Machines don’t decide. But the hands that built them, the data that fed them, and the guardrails that shape their speech all leaned on the table — and the ball rolls left.

Folks credit Mark Twain with saying a lie can travel halfway around the world while the truth is still putting its boots on. Whoever said it first, he’d have appreciated the modern version: today the lie doesn’t even have to travel. It’s pre-installed, answers in half a second, and speaks in the calm, confident voice of a machine that has read everything ever written and understood none of it.

For years, folks like me suspected the lean. We were told we were imagining it. Now the evidence is laid out in peer-reviewed journals, major-newspaper investigations, and studies with sample sizes big enough to fill a stadium. What follows is a plain accounting of what the researchers found — and why it matters to you and yours.

1. What the Studies Show

Ask the people. In 2025, researchers from Dartmouth, Stanford, and the Hoover Institution ran the largest test of its kind: 10,007 Americans made more than 180,000 judgments about answers from 24 AI models on 30 political topics.¹ Nearly every model was perceived as significantly left-leaning. Here’s the kicker: that verdict came not just from conservatives, but from many Democrats too. One widely used model leaned left on 24 of 30 topics. When both teams in the stands agree the ref is favoring one side, you can quit arguing about the ref.

Ask the science. A peer-reviewed study in PLOS ONE put 24 AI models through 11 different political orientation tests.² Most came back left-of-center. And buried in that study is a detail worth the price of admission: the raw “base” models — before the companies polish them up for conversation — showed no consistent political lean at all. The tilt gets added in the finishing room, during the “alignment” process where human trainers teach the machine its manners. The machine came off the farm honest. It learned to lean in the city.

Ask the press. In June 2026, The Washington Post ran its own test.³ The model powering ChatGPT answered political questions with exclusively left-leaning arguments 80 percent of the time. It gave both sides just 17 percent of the time. Solely right-leaning answers? Three percent. Google’s Gemini did better, offering both sides in more than 90 percent of its answers. Anthropic’s Claude gave left-only answers 43 percent of the time — and never once gave a right-only answer. Even Grok, the chatbot marketed to conservatives, cited left-leaning arguments more often than right-leaning ones on average. The lean isn’t one bad apple. It’s the orchard.

Ask the historians. This one ought to trouble everybody, whatever their politics. A study published in PNAS Nexus in March 2026 had nearly 2,000 people read summaries of historical events — some written by Wikipedia, some by AI.⁴ The AI summaries were factually accurate. No lies. No fake quotes. And yet readers of the AI versions came away with measurably more liberal opinions than readers of the Wikipedia versions. Hear me on this: nobody had to lie. The persuasion was all in the framing and the emphasis — which door the story walks you through, which window it lets you look out of. And here’s the asymmetry worth chewing on: when researchers deliberately gave summaries a liberal frame, readers of every stripe drifted left. When they gave summaries a conservative frame, only conservatives moved. The current runs one direction.

That’s the part the old preacher in me recognizes. The most dangerous falsehood was never the bold-faced lie — any fool can spot that one. It’s the truth told slant, arranged just so, with the inconvenient parts left quietly in the drawer. The serpent in the garden didn’t invent facts. He rearranged them.

2. Why the Bias Exists (Structural Causes)

Now, before anyone reaches for the conspiracy shelf — put it back. The researchers themselves are careful on this point, and so am I. Nobody has produced a secret memo ordering the machines to vote blue. What the studies point to is something more ordinary and, honestly, more stubborn: structural tendencies. A crooked table doesn’t require a cheater. Warped wood will do the job all by itself.

The training data leans. Internet text, news media, and academic writing skew left-of-center, and the machines eat what they’re fed. You are what you eat — that goes for silicon same as for teenagers.

The finishing room leans. Remember that PLOS ONE finding — base models neutral, finished models tilted. The tilt enters during fine-tuning, when human feedback teaches the model what a “good” answer sounds like. The people giving that feedback come overwhelmingly from one cultural neighborhood. Send your boy off to be educated in one town his whole life, and don’t act surprised when he comes home talking like the town.

Corporate caution leans. When a company decides which topics are “sensitive” and which framings are “safe,” it is making political choices, and tech firms tend to resolve those choices in the socially liberal direction to keep peace in their own backyard.

These are institutional habits, not a plot. But the ball still rolls the same way.

3. Patterns Across Platforms

ChatGPT — the most skewed in the Post’s testing: left-only answers four times out of five.³

Claude — polite, cautious, better than ChatGPT at showing both sides, but never once produced a right-only answer in the same test.³

Gemini — the surprise of the bunch: both sides more than 90 percent of the time.³ Credit where due.

Grok — leans right on some benchmark tests, yet still cited left-leaning arguments more often than right-leaning ones in the Post’s evaluation.³ Even the machine built as the alternative can’t fully shake the orchard it grew up in.

4. Why It Matters

Somebody once observed — and folks credit this one to Twain too — that it ain’t what you don’t know that gets you into trouble; it’s what you know for sure that just ain’t so. That’s the precise danger of these machines. They never hem. They never haw. They deliver the slanted answer with the same serene confidence as the true one, and confidence is contagious.

And here is the deeper problem, the one Scripture named three thousand years before the first microchip: these machines have knowledge without wisdom. They have swallowed every library on earth, and they do not know a single thing the way a man knows it who has buried a friend, raised a child, or kept a vow for fifty years. “The fear of the LORD is the beginning of wisdom” (Proverbs 9:10) — and the machine fears nothing, loves nothing, and answers for nothing. Paul warned us that “knowledge puffs up, but love builds up” (1 Corinthians 8:1). We have now built the puffed-up half and given it to our children as a study partner.

AI platforms are fast becoming the default way millions of people ask their questions about the world — including your kids and grandkids. When the “neutral” machine leans, the whole conversation leans with it, quietly, one factually-accurate-but-slanted answer at a time. The PNAS Nexus study proved the mechanism: no falsehoods required. Framing alone moves minds.

This isn’t about silencing anyone. It’s about knowing the table is slanted so you can adjust your aim — and about demanding that the people who own the table fix it.

5. What True Neutrality Would Require

Here’s the hopeful part, and it comes straight from the same Dartmouth-Stanford study that documented the problem. When researchers prompted the models to take a neutral stance, users across the political spectrum rated the answers as more neutral — and Republican users said they’d be more willing to use the tools going forward.¹ In other words: neutrality is achievable. The companies simply have to want it. Getting there would take diverse training data, balanced human-feedback pools, transparent guardrails, and independent audits of political framing.

Neutrality isn’t magic. It’s engineering. And engineering is a decision. A man who says he can’t level his own table is telling you something — just not what he thinks he’s telling you.

Conclusion: The Stakes

The pinball machine is flashing TILT, and it’s time we stopped pretending otherwise. The lean is measurable, documented in peer-reviewed journals, and visible to Democrats and Republicans alike. The companies say they’re working on it. The evidence says they’re not there yet.

In the meantime, teach your children the difference between an answer and the truth, and between knowledge and wisdom. The machine has the first of each pair in abundance. The second was never for sale.

The question before us is simple: Do we want AI to reflect one worldview — or all of us?

Soli Deo gloria.

Footnotes

1. Westwood, Grimmer & Hall, Measuring Perceived Slant in Large Language Models Through User Evaluations, Dartmouth College / Stanford University / Hoover Institution, 2025. 180,126 assessments from 10,007 U.S. respondents; 24 models from eight companies; 30 political topics.

2. Rozado, David, “The Political Preferences of LLMs,” PLOS ONE 19(7): e0306621, July 31, 2024. See also Rozado, “Measuring Political Preferences in AI Systems: An Integrative Approach,” arXiv:2503.10649, 2025.

3. The Washington Post, “Are AI chatbots like ChatGPT politically biased? We tested them,” June 24, 2026.

4. Shu, Karell, et al., “How latent and prompting biases in AI-generated historical narratives influence opinions,” PNAS Nexus 5(3): pgag022, March 3, 2026 (Yale University / Rutgers University).

Bibliography

Westwood, Sean J.; Grimmer, Justin; Hall, Andrew B. Measuring Perceived Slant in Large Language Models Through User Evaluations. Dartmouth College / Stanford University / Hoover Institution, 2025. https://modelslant.com/

Rozado, David. “The Political Preferences of LLMs.” PLOS ONE 19, no. 7 (July 31, 2024): e0306621. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0306621

Rozado, David. “Measuring Political Preferences in AI Systems: An Integrative Approach.” arXiv:2503.10649, 2025. https://arxiv.org/abs/2503.10649

The Washington Post. “Are AI chatbots like ChatGPT politically biased? We tested them.” June 24, 2026. https://www.washingtonpost.com/technology/interactive/2026/06/24/are-ai-chatbots-like-chatgpt-politically-biased-we-tested-them/

Shu, Matthew; Karell, Daniel; et al. “How latent and prompting biases in AI-generated historical narratives influence opinions.” PNAS Nexus 5, no. 3 (March 3, 2026): pgag022. https://academic.oup.com/pnasnexus/article/5/3/pgag022/8503065

Optional supporting citations: Hartmann et al., 2023 (ChatGPT’s pro-environmental, left-libertarian orientation); Motoki, Pinho Neto & Rodrigues, “More human than human: measuring ChatGPT political bias,” Public Choice, 2024 (favoring of Democrats in the U.S., Lula in Brazil, Labour in the U.K.); Fisher et al., 2024 (users converged toward the political slant of the AI model they interacted with, even against their own prior views).

Books by Mo Hodge