Coinbase AI System Sends False World Cup Match Result Before Game Starts, CEO Investigates

Coinbase's AI system sent push notifications to millions of users falsely claiming Norway defeated Brazil 3-2 in a World Cup playoff match — before the game even started.

The AI hallucinated the result with Erling Haaland scoring twice, and the exchange's product head later joked, "The AI might have known something we didn't."

The false notification appeared in the "Trends" section, which uses AI to generate news summaries based on Kalshi prediction markets. The actual final score was Norway 2-1, with Haaland scoring twice — meaning the AI guessed the goal scorer correctly but got the scoreline, timing, and match outcome wrong.

Coinbase CEO Brian Armstrong stated he is investigating with his team. The exchange has since fixed the error and updated systems to prevent similar inaccuracies. The incident highlights the growing risk of AI-generated news feeds in financial applications.

Coinbase is embedding AI into its user experience, believing "agents are no longer a niche developer curiosity, but really a primary way people interact with the internet." The company recently launched "Coinbase for Agents," allowing AI models to execute trades, make payments, and rebalance portfolios autonomously.

But the "Trends" section is a more conventional feature: it summarizes prediction market data into news-like summaries. That summary looked like breaking news, went to the lock screen of millions of phones, and was entirely wrong.

"The AI might have known something we didn't." – Coinbase product head, after the AI predicted Haaland would score twice (which he did)

Coinbase is not a news organization. It is a financial services company that suddenly became a publisher of fake sports results to tens of millions of users.

The incident underscores a broader issue: as AI-generated summaries become standard in financial apps, the line between curated data and generated content is blurring. The system summarized prediction market data but presented it as a fact, stripped of the context that it was speculative.

The issue raises a structural question: if an AI is summarizing prediction markets, it should say "Kalshi traders believe Norway will win" — not "Norway has won."

The gap between training data and live events may become more manageable as models improve, but the more fundamental risk is that users will treat AI-generated text as fact simply because it appears in a trusted app.

As Coinbase doubles down on AI agents, the World Cup incident serves as a reminder that even a well-meaning AI deployed with good intentions can still generate hallucinations.