When Four AI Models Agree — And All Four Get It Wrong
Four AI models predicted George Russell would win the 2026 Formula 1 World Championship. Instead, his Mercedes teammate Kimi Antonelli made history. What went wrong, and what can we learn about trusting AI predictions?
Back in March 2026, Mercia AI asked four different artificial intelligence systems to predict the Formula 1 World Drivers' Champion. ChatGPT, Gemini, Perplexity and Grok were each given the challenge of assessing the new season, which featured some of the biggest technical regulation changes in the sport's history.
All four reached the same conclusion: George Russell would become the 2026 World Champion.
Well, that predication didn't age particularly well.
Fast-forward eight months to the Mexican Grand Prix on 1 November, and we now know how that prediction turned out. Kimi Antonelli, Russell's 20-year-old Mercedes teammate, secured the World Championship with four races remaining, becoming the youngest champion in Formula 1 history.
However, rather than quietly forgetting our original predictions, this is an opportunity to examine what happened. After all, the purpose of the experiment wasn't simply to see whether AI could correctly name a champion. It was also to explore how different models assess uncertainty, interpret evidence and arrive at conclusions about events that haven't happened yet.
And as it turns out, four AI models agreeing with one another doesn't necessarily mean they're right.
Looking Back at the Original Predictions
Our original article, F1 2026: When Four AI Models Agree in a Reset Year, was published on 1 March 2026, ahead of the new season.
Each model considered the impact of the new technical regulations, including changes to power units, aerodynamics and car design. With teams facing considerable uncertainty about their relative competitiveness, the 2026 season presented an interesting forecasting challenge.
Despite using different AI systems, the predictions were remarkably consistent.
AI model | Predicted champion | Probability | Confidence |
ChatGPT | George Russell | 22% | Low |
Gemini | George Russell | 38% | Medium |
Perplexity (Opus 4.6) | George Russell | 28% | Low |
Grok 4.20 Beta | George Russell | 35% | Medium |
Mercedes featured prominently in the reasoning behind these predictions. The team had considerable experience developing competitive power units, and the new regulations appeared to offer an opportunity to return to the front of the grid.
Russell also seemed a sensible choice. He was the more experienced Mercedes driver, had demonstrated his abilities against established competitors and was entering the season with several years of Formula 1 experience behind him. Antonelli, meanwhile, was approaching only his second season in the championship.
The collective conclusion was understandable, particularly given how much uncertainty surrounded the new cars. But as the season unfolded, it became clear that the models had underestimated one particularly important factor: the potential of the young Italian sitting in the other Mercedes.
Mercedes Was the Right Team, but the Wrong Driver Won
One of the more interesting aspects of this experiment is that the models weren't entirely wrong about the competitive landscape.
Mercedes did emerge as a championship-winning team under the new regulations, suggesting that the underlying assessment of its technical capabilities had some merit. The mistake was identifying which Mercedes driver would ultimately take advantage of that performance.
Antonelli began the season strongly, finishing second in Australia before claiming victories in China and Japan. As the championship progressed, he established himself as a serious title contender, demonstrating the consistency and racecraft needed to compete at the front.
By the time Formula 1 arrived in Mexico, his advantage was so substantial that he needed only a ninth-place finish to secure the championship. He finished third, becoming World Champion with four Grands Prix still remaining.
For an experiment that had placed considerable emphasis on experience, the outcome was particularly revealing. Russell's established reputation had made him the logical choice, but Antonelli's development proved more significant than the models appeared to anticipate.
This doesn't mean experience was an unreasonable consideration. In Formula 1, it can be enormously valuable, particularly when teams are adapting to unfamiliar technical regulations. However, relying too heavily on previous performance can create blind spots when a younger competitor improves rapidly.
The models had identified a plausible championship-winning environment. What they hadn't adequately anticipated was how quickly Antonelli could become the driver best positioned to exploit it.
Why Did All Four AI Models Agree?
Perhaps the most interesting question isn't why the models selected Russell, but why every one of them reached the same conclusion.
When several AI systems provide similar answers, it's natural to feel more confident about the result. We tend to assume that agreement between different sources strengthens the evidence, particularly when those sources appear to be working independently.
However, AI models don't necessarily approach a question with completely independent information or assumptions. Different systems may draw upon overlapping publicly available material, including historical results, technical analysis, media reporting and established opinions about drivers and teams.
They can also favour similar patterns of reasoning. In our experiment, Mercedes' expected competitiveness and Russell's experience were recurring considerations. Once those factors were given considerable weight, selecting Russell became a logical conclusion.
The problem is that this agreement may have reflected shared assumptions rather than genuinely independent assessments.
In other words, four models making the same prediction doesn't automatically provide four separate pieces of supporting evidence. If their reasoning depends on similar information and expectations, they may also share the same blind spots.
We cannot establish precisely why each model favoured Russell without examining its underlying processes, and the original responses don't prove that the systems influenced one another. Nevertheless, their agreement demonstrates why AI consensus should be treated as something worth investigating rather than as a guarantee of accuracy.
This is a useful lesson well beyond Formula 1, particularly as more people begin comparing responses from multiple AI tools before making decisions.
Were the Models Actually That Confident?
There's another detail in the original predictions that deserves closer examination: the probabilities.
Although all four models selected Russell, none gave him a probability greater than 38%. ChatGPT estimated his chances at just 22%, while Perplexity gave him 28%. Even Gemini, which produced the highest estimate, assigned a 62% combined probability to other drivers winning the championship.
That distinction matters because selecting the most likely individual outcome is not the same as saying that outcome is certain, or even more likely than all alternatives combined.
Consider ChatGPT's 22% estimate. If that probability were properly calibrated, we would expect an event assigned a 22% chance to occur roughly 22 times out of 100 comparable situations. The remaining 78 outcomes would go another way.
Russell failing to win the championship therefore doesn't, by itself, prove that the probability estimate was unreasonable.
To evaluate the accuracy of these forecasts properly, we would need many predictions covering different races or seasons, along with their assigned probabilities and actual outcomes. We could then examine whether events given higher probabilities occurred more frequently than those assigned lower probabilities.
What we can conclude from this experiment is simpler: all four models selected the wrong championship winner, despite differing levels of confidence.
That is an important result, but it shouldn't be exaggerated into a claim that AI forecasting is inherently unreliable. One unsuccessful forecast cannot establish that, just as one successful forecast cannot prove that a model is consistently accurate.
What Can We Learn Beyond Formula 1?
Although Formula 1 provides an entertaining example, the underlying lessons have practical applications for individuals and organisations using artificial intelligence.
Imagine a small business asking several AI systems to forecast next year's sales, identify the most promising new market or recommend whether to invest in a particular product. If every system produces a similar recommendation, that apparent consensus might provide reassurance.
But what if those systems are relying on similar historical data? What if customer behaviour is changing, an emerging competitor hasn't been considered or the business is entering a market unlike anything in its previous experience?
The agreement might still be useful, but it doesn't remove the need to examine the evidence and understand the assumptions behind the recommendations.
This is where AI works best as a decision-support tool rather than an unquestioned authority. Instead of simply asking which outcome is most likely, users can ask what evidence supports the conclusion, which assumptions matter most and what circumstances might cause the forecast to change.
For smaller organisations, particularly those with limited time and resources, this can make a substantial difference. AI can help identify patterns, examine alternatives and communicate findings, but the quality of those outputs still depends on the available information and how carefully the results are interpreted.
Good forecasting isn't simply about producing a confident answer. It's about understanding uncertainty and recognising when new evidence should cause us to reconsider an earlier conclusion.
From Getting It Right in 2025 to Getting It Wrong in 2026
There's a particularly interesting contrast between Mercia AI's two Formula 1 championship prediction experiments.
In 2025, four AI models correctly selected Lando Norris as the eventual World Champion. That was an encouraging result, especially given the uncertainty surrounding a closely contested championship.
For 2026, the same broad approach produced another unanimous prediction. This time, however, all four models backed George Russell, only for his teammate to secure the title before the season had even finished.
The contrasting results illustrate why we should be cautious about drawing sweeping conclusions from individual successes or failures.
Correctly predicting one championship doesn't establish that an AI model has developed a reliable forecasting method. Equally, getting the following championship wrong doesn't necessarily mean its analysis was without value.
In 2026, the models identified Mercedes as a strong contender but underestimated Antonelli's potential. Their forecasts contained elements of useful reasoning, even though their final selections proved incorrect.
Perhaps the most important lesson is that the reasoning behind a prediction can be just as valuable to examine as the prediction itself. Understanding why an AI system reached a conclusion helps us recognise both its strengths and its limitations.
AI Isn't a Crystal Ball
In the original March article, we made an observation that now seems particularly appropriate:
"It does not make the prediction correct."
That was our conclusion about four AI models independently agreeing on the same championship winner. Eight months later, Antonelli has provided a memorable demonstration of why that caution matters.
Artificial intelligence can be useful for analysing complex information, comparing possibilities and helping people make decisions. However, agreement between models should encourage us to examine their reasoning, not abandon our own judgement.
The best approach is to treat AI forecasts as informed possibilities rather than guaranteed outcomes, particularly in environments where circumstances can change rapidly.
As for our Formula 1 predictions, congratulations to Kimi Antonelli on an extraordinary achievement. Becoming the youngest World Champion in the sport's history is remarkable, regardless of what four AI systems thought might happen.
And for Mercia AI, there's always the 2027 season.
Perhaps next time we'll pay a little more attention to the other side of the Mercedes garage.
At Mercia AI, we explore how artificial intelligence can help people and small organisations understand information, assess uncertainty and make better-informed decisions. The lessons from this Formula 1 experiment apply just as readily to everyday AI use and business data analysis.