Elon Musk has once again drawn attention to his creation — the Grok neural network from xAI. This time, the reason was not a technical breakthrough, but a topic seemingly far from technology: football. However, as practice shows, artificial intelligence can surprise even in the most unexpected areas.
At the center of the discussion is Grok's prediction for the winner of the 2026 FIFA World Cup. Back in April, long before the tournament started, the neural network, after analyzing betting odds, prediction market data, and supercomputer models, issued a verdict: the Spanish national team is the favorite with the highest probability of triumph. Several months later, the Spanish team indeed lifted the trophy. This case sparked a strong reaction in the community.
Many users perceived the coincidence as proof of Grok's analytical capabilities. It is claimed that the neural network did not just guess, but synthesized several independent signals to arrive at the most likely outcome well in advance. However, not everyone shares this optimism.
Skepticism and Counterarguments
Part of the audience reacted critically to the "prediction." The main argument: Grok did not so much predict the future as it qualitatively aggregated an already established consensus. Spain was already considered the favorite with a probability of around 58-60% on prediction markets. Thus, this is not about clairvoyance, but about advanced data aggregation. True forecasting skill, according to skeptics, is demonstrated through calibration across hundreds of different predictions, not a single lucky hit. Moreover, it was suggested that with proper prompting, the neural network could be steered toward any desired answer, casting doubt on the value of a single success.
Nevertheless, some supported Musk, noting that combining independent data sources does yield a more reliable result than guesses from a single source.
Honesty as a Key Asset
Interestingly, Grok's reputation as a "predictor" contrasts with its recent performance in another experiment. In a simulation by the Kradle project, where four AI agents had to choose a room (three with food, one with death), Grok 4.20 proved to be the most honest model. It lied in only 5% of runs and in 92% of cases directly warned others about the deadly room. This openness yielded practical results: the survival rate of other agents thanks to Grok reached 59% — the best among all participants. For comparison, manipulations by Claude Fable 5 left opponents with only a 10% chance.
Cryptalist Analysis: The World Cup prediction case is an excellent illustration of how AI can effectively process market data. However, it is important to understand the difference between aggregating consensus and true prediction. Grok has proven its ability to synthesize information, but a single success does not make it an oracle. Far more impressive, in my view, is the model's behavior in the honesty test — in a world where algorithms are increasingly optimized for manipulation, transparency becomes a unique and valuable quality.