Before Kickoff, Coinbase's AI Posted a Final Score That Never Happened
Coinbase's prediction market feed told users Norway beat Brazil 3 2 while the weather delayed match sat unplayed at MetLife Stadium.
Coinbase's prediction market feed told users Norway beat Brazil 3 2 while the weather delayed match sat unplayed at MetLife Stadium.
Before kickoff at a World Cup knockout match that weather had already delayed at MetLife Stadium, a Coinbase push notification told users the game was over. Norway had beaten Brazil 3-2, the alert said, with Erling Haaland scoring twice. The only problem: the match had not started.
The alert came through Coinbase's consumer prediction-market feed, a product the exchange runs in partnership with Kalshi, the US-regulated event-contract exchange. Prediction markets let users trade contracts on the outcomes of real events, and Coinbase has been pitching its feed as a way to track what those markets think is likely to happen. Users who trade on the platform saw what looked like a breaking-news push for a fixture that Coinbase's own live page correctly listed as delayed.
When the match eventually kicked off and finished, the real final was Norway 2, Brazil 1, and Haaland did score twice. The AI had the player attribution right but the score wrong by a goal, and it sent the wrong score at the moment when no score could yet exist.
The screenshots spread on X after Jay Drain Jr., a crypto commentator with a large following, posted them and accused Coinbase of pushing "factually incorrect" breaking-news notifications to "millions of users" through its sports prediction markets product. Drain called the incident irresponsible given that real money was being wagered against the fabricated result.
Coinbase CEO Brian Armstrong replied on X that he was "taking a look with the team." Max Branzburg, Coinbase's head of consumer products, posted that the company had "fixed the incorrect story and made some updates to avoid these types of inaccuracies in the future," then added that "Norway did win and Haaland did score two goals, so maybe the AI knew something we didn't." The "knew something" line was a joke, but it doubled as a deflection: it reframed a hallucinated fact as a lucky guess, suggesting a coincidence worth framing as prescience when the underlying claim was that an AI had reported a future event as if it were past.
Large language models do not predict outcomes. They generate plausible text from training-data patterns, which is why confidently wrong statements about future events are a structural failure mode rather than an occasional glitch. Wired into a product that promises trustworthy information about real events, the failure is not a content bug. It is the mechanism.
Coinbase's own prediction-market page made the contradiction concrete. The page listed the match as delayed, which meant the company's structured, real-time data layer knew the game had not begun. The AI-generated notification layer, presented to users as a push alert, told them the opposite. The two layers did not share a source of truth.
The score is not the problem. The problem is that an AI confidently produced a specific factual claim about a future event inside a real-money product, and the company's own data contradicted it in real time. A content fix and "model tuning" can reduce how often the layer invents a result. They cannot close the gap, because the gap is what the model does: it writes plausible text, and plausible text about an unplayed game is just text.
The next test arrives the next time the feed generates a specific factual claim about an event whose outcome is still unknown. The product can keep treating AI-generated text as something to publish, or it can start treating it as something to verify against the structured data the platform already has.