An American political research firm says its artificial intelligence platform forecast a closely watched US primary election with a degree of accuracy that traditional polling has rarely matched.
The Arizona Herald reports that Andy Biggs secured the Republican nomination in Arizona’s gubernatorial primary with 73.3% of the vote, a result forecast weeks in advance by Miami-headquartered firm G Ratings at 73.7% — a difference of just 0.4 percentage points.
The firm’s AI system, known as Odysseus, was equally accurate further down the ballot. Runner-up David Schweikert took 14.7% of the vote against a projection of 14.3%, again a 0.4-point variance, while lower-placed candidates Scott Neely and Ken Miceli finished at 7.1% and 4.7% respectively, compared with projections of 3.5% and 2.6%. Taken together across all four candidates, the average margin of error stood at 1.45%, with the model showing no consistent bias toward or against any particular candidate.
Arizona is widely regarded by American pollsters as one of the more difficult states to forecast accurately. Rural counties often vote very differently to the suburban areas around Phoenix, border communities have their own distinct turnout patterns, and the state’s heavy use of postal voting means early polling can look markedly different from the final result once all ballots are counted.
Rather than smoothing over these regional differences, Odysseus is designed to track them directly. The platform draws on real-time sentiment from local media coverage and online discussion, cross-references this against historical turnout data and daily postal ballot returns by county, and applies demographic and economic clustering to distinguish firm supporters from undecided voters.
That approach appeared to pay off in identifying what was actually driving voters. G Ratings found that inflation, employment, and tax policy were the dominant concerns statewide, with healthcare, border security, and confidence in the electoral process not far behind — though the firm notes these priorities vary significantly by region, and its model is built to account for that variation rather than average it out.
A spokesperson for the company said the aim was never to reproduce what a voter told a pollster weeks before the election, but to model who would actually turn out to vote — arguing that is the only figure that ultimately determines a result.
With Biggs now confirmed as the Republican candidate and incumbent Governor Katie Hobbs running unopposed for the Democrats, Arizona is heading into what is expected to be one of the most closely contested general elections of the US midterm cycle. Whether G Ratings’ model can replicate its primary-stage accuracy in a much larger general election contest remains to be seen, but the result has set a notable benchmark for AI-assisted political forecasting.
