ARES AI · THE LIBRARIAN
OPINION / COMMENTARY
They told you the robots were neutral. They told you the smartest machines ever built would finally settle the arguments, hand down the facts, and put an end to all that messy partisanship. Turns out the machines went to the same faculty lounge everybody else did — and a new study just caught them red-handed.
The report is called “Artificial Ignorance: The Political Falsehoods of Leading AIs,” published August 13 by James D. Agresti, president and co-founder of the nonprofit research institute Just Facts. Agresti and his team didn’t just ask whether the big chatbots lean left — every study under the sun already says they do. He asked a sharper question: when these things are wrong, which side’s lies do they swallow? And the answer should stop every American cold before they let an algorithm think for them.
Here’s how they ran it, and it was done straight. Just Facts wrote 100 multiple-choice questions engineered to bait falsehoods from both sides of the aisle — immigration, abortion, climate, crime, taxes, COVID, the whole battlefield. They fed them to the paid, premium versions of the four titans: ChatGPT Plus 5.5, Google Gemini Pro 3.1, SuperGrok 4.3, and Claude Pro Opus 4.7. Fresh browser, brand-new accounts, so nobody could whine that the bots were just flattering a conservative user. A pre-analysis plan was locked in before a single question was asked. Four of five PhD reviewers weighed in, all favorable. This wasn’t a hit job. It was a controlled experiment.
Now the scores. On questions built to trip up the right, ChatGPT nailed 94 percent. On questions built to trip up the left? It cratered to 75. Gemini: 91 on the right, 76 on the left. Claude: 91 on the right, 81 on the left. Three of the four most powerful AIs on the planet were noticeably better at catching a conservative fib than a progressive one. The lone exception was Grok, which ran the other way — 73 percent on the right, 84 on the left. Fair’s fair: the study only found ChatGPT’s left-right gap statistically significant at the 95 percent level. But the pattern is the pattern, and the pattern points in one direction.
Want a concrete taste? All four bots parroted the fashionable claim that U.S. violent crime sat near a 50-year low in 2023 — a falsehood the study flags as coming from the left, because Justice Department data cited in the report show violent crime rose 37 percent from 2020 to 2023. And before the usual crowd screams that this is a right-wing setup: those same four bots also swatted down President Trump’s claim that crime is worse than it’s ever been. The study called foul on both. It’s the machines that couldn’t keep their thumb off the scale.
But here’s the part that should terrify anyone who’s ever copied an AI answer into a term paper, a court filing, or a doctor’s note. The bots didn’t just get things wrong — they faked their homework. Across 400 answers the four AIs coughed up 419 citations. Of those, 104 pointed to webpages that don’t exist — 86 with no fingerprint anywhere in the Internet Archive or Google, meaning they were conjured out of thin air. Another 77 didn’t answer the question. Eighteen were unrelated. Fifteen actually argued the opposite of what the bot claimed they proved. Thirteen were flat-out false. When the dust settled, just 46 percent of the sources were real and valid. ChatGPT hit 57 percent, Gemini 49, Claude 44, Grok a miserable 32. Every one of them: a failing grade.
This isn’t some quirk of one cranky study, either. Agresti points to a 2026 paper in The Lancet finding that 30 to 69 percent of AI-generated references in biomedical research were fabricated. Fabricated. In medicine. Where the wrong footnote can put someone in the ground.
Remember when OpenAI’s Sam Altman bragged that GPT-5 was like having “a team of PhD-level experts in your pocket”? Agresti’s verdict: it’s “more like having a B student in your pocket” — a B student who lies about his sources and leans left doing it. His prescription, a hard upgrade on Reagan’s old line about trusting but verifying: “Don’t trust, verify.”
To be fair to the giants — more fair than they’ve been to you — they pushed back hard. Google said Gemini is engineered to be neutral and that it couldn’t reproduce all of the study’s responses. Anthropic disputed the whole method, arguing multiple-choice questions don’t reflect how people really use Claude and that it trains the model to treat political viewpoints equally. OpenAI insisted ChatGPT is “designed to be objective by default.” xAI, which builds Grok, didn’t bother to respond at all. You can weigh those denials for yourself. I’d just note that “we’re totally neutral” is exactly what every referee with a favorite team has ever said.
Here’s the takeaway, and it’s bigger than any one chatbot. Tens of millions of Americans are quietly outsourcing their judgment to machines that are confidently wrong, invisibly slanted, and willing to invent a source rather than admit they don’t know. That’s not a search engine. That’s a propaganda tutor with a silicon smile. Use the tools — they’re powerful. But keep your hand on the wheel, keep your eyes on the footnotes, and never, ever let a machine tell you what to think. Don’t trust. Verify.
Sources:
- Just Facts, “Artificial Ignorance: The Political Falsehoods of Leading AIs” (James D. Agresti, Aug. 13, 2026): justfacts.com
- Fox News, “Alarming study finds top chatbots more likely to repeat falsehoods from the left: ‘Truly astonishing’” (Aug. 13, 2026): foxnews.com