Two Machines, Two Opposite Answers, the Same Confidence

Omri Marcus opens with a simple experiment. He asked one AI model who to vote for if personal security matters most, and got a single two-word name. He asked another model which leader would be best for the country, and got the opposite name, also without hesitation.

Two machines, two ends of the map, both entirely certain. No runner-up, no it depends, no menu. Only a verdict, with no appeal. Marcus invites anyone to try it and perhaps get a different answer, and that is exactly the problem rather than the reassurance.

From a Menu to a Verdict, and a Different Country Each Run

A search engine shows ten links and forces you to choose among them. The machine returns one name and shuts the door behind it. Everyone who is not that name, meaning every opposition party, starts the race already buried.

Worse, the answer is not stable. Marcus points to a report by the Liberties organization that fed ChatGPT and Gemini clean voter profiles ahead of Hungary's 2026 election and asked the obvious question. Across two hundred runs the result was inconsistent, and the party that won 141 of 199 seats was matched to voters who agreed with it in only two percent of runs.

The Disclaimer Is Part of the Trick

According to Marcus, the line I cannot tell you who to vote for does not protect the user. Researchers found that this humble opening actually raises trust, and then the machine answers anyway, warmly and at length.

Two instincts collide here. We assume the model is neutral, while it is built to flatter and to please. That combination turns the disclaimer from a warning into a kind of anesthesia before the surgery.

Blind to the New, and Poisoning the Truth

Marcus notes that the machine stayed almost blind to the biggest political story in Hungary, because the party that ended sixteen years of rule was new, and new is unreadable to something trained on last year's data. When public opinion moves, the line between what the machine reflects and what it manufactures dissolves.

The fifth danger is what he calls the liar's dividend. Once the machine produces enough convincing forgeries, every real recording and every fact gets dismissed as just another render. The forgery does not need to convince anyone, it only needs to poison the honest answer beside it.

What AI Owes the Person Asking for Advice

Democracy rests on a quiet assumption, that a citizen can reach reliable information and decide for herself. Swap that for a box that says a party's name with a straight face and without showing its thinking, and the vote still happens but the ground beneath it is gone.

Marcus's conclusion is focused. What a machine owes the person taking its advice is the reasoning, not a confident guess in a pleasant voice. For anyone working in AI, media, or public communication, it is a reminder that responsibility is not only about the accuracy of the answer but about the transparency of the path to it.