An AI “Predicted” the Next President. The Odds It Quoted Came From a Betting Market.

Two headlines on UNILAD Tech this week promised something big. On October 15, “AI reveals how Trump could become first president to serve a third term since 1941.” A day later, “AI predicts next US President as Trump teases unconstitutional third term.” Both stories shared a premise: someone typed a question about 2028 into a chatbot and wrote up what came back.

Look at what the AI actually produced, though, and the “prediction” gets thin. The chatbot summarized constitutional law that has been settled text since 1951. The one hard number in the story, a 21% chance that JD Vance wins in 2028, didn’t come from an AI at all. It came from Polymarket, a betting exchange. That gap between the headline and the source is a useful way into a bigger question: what is an AI election prediction actually worth?

What the AI actually said

The chain starts with a LADbible article published October 15. The writer asked an unnamed AI whether Trump could run as Vance’s vice president and then take over if Vance resigned. The answer, quoted in full by LADbible, was cautious: “Donald Trump running as JD Vance’s vice president, then becoming president if Vance resigned, is a constitutional loophole that has been debated by legal scholars. But there are significant legal obstacles, and it would almost certainly face a Supreme Court challenge.”

UNILAD Tech’s October 15 story names ChatGPT and lays out three “routes” it offered: the Vance swap, Trump becoming Speaker of the House (the Constitution doesn’t say the Speaker must be a member of Congress), and amending the Constitution outright. The October 16 follow-up turns that into a prediction that Trump “likely won’t be on the Republican ticket in either position,” then adds Polymarket’s numbers.

None of the pieces say which ChatGPT model answered, whether it browsed the web, or how the question was worded. Those details matter a lot. The same chatbot can give different answers to the same question on different runs, and rewording a prompt can flip its tone from “possible loophole” to “almost certainly not.”

The part the chatbot got right

On the law, the AI’s summary holds up, mostly because the law is well documented. The 22nd Amendment, ratified in 1951, says plainly: “No person shall be elected to the office of the President more than twice.” Congress passed it after Franklin D. Roosevelt won four presidential elections, in 1932, 1936, 1940 and 1944. His third inauguration, in January 1941, is the date behind the “since 1941” headline. He died in April 1945, a few months into his fourth term, and remains the only president to serve more than two.

The vice president workaround runs into the last sentence of the 12th Amendment: “no person constitutionally ineligible to the office of President shall be eligible to that of Vice-President of the United States.” Scholars disagree on whether being barred from election makes someone “ineligible to the office,” and no court has ruled on it. The Constitution Annotated, the Library of Congress’s official legal commentary, acknowledges the open question. Changing the 22nd Amendment would take a two-thirds vote in both chambers of Congress plus ratification by 38 states. Rep. Andy Ogles of Tennessee introduced such an amendment in January 2025, and it went to committee.

Trump himself has gone back and forth. In October 2025 he called the vice president route “too cute” and said “the people wouldn’t like that.” At an Arizona rally on October 14, 2026, with the crowd chanting “Four more years,” he said, “I’d love to,” then added, as reported by the Associated Press: “We have a clause that is sort of hard to break.”

Where the numbers really came from

The 21% figure for Vance, the 13% for Ocasio-Cortez and the 64% chance of a Democratic win all come from Polymarket. When we checked the exchange’s public data on October 18, Vance was at 20.5%, Ocasio-Cortez at 13.2%, Jon Ossoff at 10.2%, Gavin Newsom at 9.0% and Marco Rubio at 8.2%. The party market had Democrats at 63.5% and Republicans at 34.5%. Trump’s own contract was trading at 1.25 cents on the dollar, on more than $10 million in trading. The whole winner market has seen about $717 million in volume.

Those numbers look precise. They have also moved a lot in a single year.

A year ago Newsom was a close second at 22.5%. Today he’s fourth. Ossoff, now third, didn’t have a contract until February 2026. Ocasio-Cortez has doubled. None of that means the market is broken. It means the market is doing what markets do: repricing on news, polls and attention. Read that way, a 20% favorite is a candidate who loses four times out of five.

How good is anyone at this?

Election forecasting is hard even with the best tools. The American Association for Public Opinion Research found that 2020 polling error was the highest in 40 years for the national popular vote, and the highest in at least 20 years for state-level races. Betting markets have their own misses. Most of them favored Hillary Clinton on the eve of the 2016 election. Polymarket had Trump as the favorite in the final weeks of 2024 and was right. Those are calls made days out, though. Calls made more than two years out, before a single primary ballot, have much weaker records.

AI forecasting is improving, and researchers measure it. The Forecasting Research Institute runs ForecastBench, which scores predictions on real questions that resolve later, with lower Brier scores meaning better accuracy. In January 2026 it reported that human superforecasters still held first place, ahead of the best AI systems by 0.017 Brier points, or about a year of AI progress at the recent pace. The top AI entries weren’t plain chatbots. xAI averaged eight Grok forecasts, each made with search tools. Another team, Cassi, ran a pipeline that ensembled several models and nudged its answers toward market prices. A one-off question to a consumer chatbot is nothing like that.

Chatbots have a rocky record on election questions

The bigger worry with asking a chatbot about elections isn’t the long-range guess. It’s the confident wrong answer to a simple factual question. Independent testers have documented that repeatedly.

The AlgorithmWatch and AI Forensics study found Bing Chat inventing scandals about real candidates, often citing a source that had reported correctly. In the AI Democracy Projects test, none of the five models could correctly say that Texas bans campaign apparel such as a MAGA hat at polling places. “The chatbots are not ready for prime time when it comes to giving important nuanced information about elections,” said Seth Bluestein, a Republican city commissioner in Philadelphia who took part in the testing.

The companies’ answer has been guardrails: send voting questions to authoritative sites and refuse some election topics outright. Anthropic’s 2024 announcement said its model “is not trained frequently enough to provide real-time information about specific elections” and can hallucinate incorrect information. Those guardrails mostly target voting logistics, like where and when to vote. Questions about loopholes and hypothetical third terms usually get through, which is how stories like this week’s get written.

Why a chatbot “prediction” is mostly an echo

A language model doesn’t run a forecast. It produces likely text based on its training data and, if browsing is on, whatever pages it retrieves. Ask about 2028 and you get a blend of legal commentary, cable news and, quite possibly, the same Polymarket numbers the headline cited. That makes an AI answer a decent summary of what the internet already says and a poor independent signal. It also makes these answers fragile. Research on AI summaries shows that one wrong detail can stick in a reader’s memory, and an election claim screenshotted from a chatbot travels fast.

What you can do

• For voting logistics, go straight to official sources: CanIVote.org, Vote.gov or your state election office.
• When a story says “AI predicts,” look for the model, the prompt and whether web search was on. If they’re missing, treat the answer as an anecdote.
• Read a market price as a probability, not a verdict, and check how much it has moved over time.
• Ask the chatbot for its sources, then open them. Constitutional text and the Constitution Annotated are free online.

The chatbot in this week’s stories did a reasonable job explaining a 75-year-old amendment. It didn’t predict the next president, and it never ran anything resembling a forecast. With more than two years and a full primary season to go, the honest forecast for 2028 is the one prediction markets are already pricing in, with no candidate above about one chance in five: nobody knows yet.

Sources and further reading

• UNILAD Tech: AI predicts next US President as Trump teases unconstitutional third term
• UNILAD Tech: AI reveals how Trump could become first president to serve a third term since 1941
• LADbible: AI weighs in on legal loophole that some say could lead Trump to office in 2029
• Polymarket: Presidential Election Winner 2028
• Proof News / AI Democracy Projects: Seeking Reliable Election Information? Don’t Trust AI
• AI Forensics: An Analysis of Microsoft’s Copilot in Swiss and Bavarian Elections
• Anthropic: Preparing for global elections in 2024
• Forecasting Research Institute: LLMs Are Closing the Gap on Human Superforecasters

About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.