Every four years, the World Cup tests the world’s best soccer players. In 2026, it is also testing AI. Microsoft Copilot, Google Gemini and other AI models are being asked to pick the champion. Opta’s supercomputer is simulating the tournament 25,000 times per round to estimate each team’s odds.
And on July 5, a Coinbase AI system told millions of users that Norway had beaten Brazil 3-2 in a match that had not yet started, according to CNET.
The alert came from Trends, an AI-generated news feature inside Coinbase’s prediction market app that surfaces news summaries alongside real-time market data, CoinDesk reported. Coinbase’s own market page showed the game under a weather delay at the time. Screenshots spread across X within hours, drawing thousands of views.
Max Branzburg, Coinbase’s head of consumer products, later wrote that the company had fixed the error and updated its systems. “It’s awesome to see the power of AI-enabled 24/7 insights for trading,” Branzburg wrote on X, “but obviously still need to tune it to address these types of issues.”
What followed made the incident stranger. Norway did beat Brazil, winning 2-1 with Erling Haaland scoring twice in the second half, ESPN reported. The AI had identified the right winner and the right scorer. It had invented the score, the timing and the certainty, and delivered all of it as confirmed news to users with active positions on the outcome.
Probability vs Certainty: How Two AI Systems Handled the Same Tournament
The contrast with how Opta approaches the same tournament captures the difference between calibrated forecasting and confident fabrication. Opta’s supercomputer has run 25,000 simulations per stage to estimate each team’s probability of winning, Opta Analyst reported. Entering the quarterfinals and as of July 9, France leads at 27.3%, followed by Spain at 21.3%, Argentina at 17.3% and England at 16.5%. Norway sits at 6.6%. Every number comes with a stated uncertainty. The model distributes probability across outcomes. It does not declare one.
That is what Coinbase’s system failed to do. A model that says France wins the World Cup in 27.3% of 25,000 simulations is a forecasting tool. A model that declares Norway beat Brazil 3-2 before kickoff is presenting a fabrication as fact. The gap between them is not model quality. It is whether the system was designed to communicate uncertainty or to suppress it.
The reliance runs deeper than one bad alert. Coinbase has said roughly 40% of its daily code is now AI-generated, with plans to push that figure higher, CoinPedia reported.
Coinbase Error Exposes What Happens When AI Skips Uncertainty
Sports predictions are low stakes compared to where the same failure mode appears in financial services. PYMNTS Intelligence found that 60% of financial services firms use AI for credit risk assessment and scoring, with 85% increasing their AI budgets over the next 12 months.
Fraud detection, risk assessment and revenue forecasting are the dominant use cases, the same applications where a confident but fabricated output causes the most damage. When a model makes an error in a financial workflow, money moves incorrectly at speed across thousands or millions of transactions before anyone detects the problem.
AI forecasting is only as reliable as what it is willing to say it does not know. Coinbase learned that on July 5. The quarterfinals begin Thursday afternoon (July 9).
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
