Some World Cup bettors were in for a rude surprise on Sunday, when an AI-generated notification from the crypto exchange Coinbase shared false breaking news about a 3-2 Norway victory in a knockout match against Brazil. The AI said Norwegian player Erling Haaland scored two goals to clinch the win.
The kicker? The match hadn't even started yet. It had been delayed due to inclement weather conditions at the New Jersey-based MetLife Stadium. At the time, Coinbase's official prediction market page showed that the game had been delayed.
Hours later, once the game actually played out, the result was slightly different. Norway won, and Haaland scored two goals, but the final tally was just 2-1. For some bettors, that's enough of a difference to tank a sizable sum of cash -- at a time when gambling addiction rates are surging among young people, thanks to the widespread proliferation of online prediction markets and sportsbooks.
Read more: Stream Every World Cup Elimination Game for Free With a VPN Coinbase is the largest US-based public market for trading cryptocurrencies like bitcoin and ethereum, but has expanded its ventures to include stock and derivative trading. Recently, the company partnered with Kalshi to start offering prediction markets.
Screenshots of the inaccurate Coinbase AI news flooded social media as some critics pointed out how propagating false results is "dangerous and irresponsible." In response to one criticism of the Coinbase AI mistake shared on X, the company's CEO, Brian Armstrong, claimed that he was "taking a look with the team."
When reached for comment, Coinbase referred to an X post written by the company's head of consumer products, Max Branzburg.
Branzburg's reply to the X post criticizing the AI model's mistake indicated that the company fixed its incorrect World Cup story and "made some updates to avoid these types of inaccuracies in the future."
"And hey -- it turns out Norway did win and Haaland did score two goals, so maybe the AI knew something we didn't," Branzburg wrote in the post.
The AI didn't "know" anything. Large language models -- the technology that most commonly gets referred to as artificial intelligence -- can't actually know or predict the outcome of future events. These machines have no more precognitive ability than you or I do.
LLMs are trained on aggregations of large quantities of data and attempt to generate relevant information or identify patterns based on presented queries. This technology frequently leads to fabrications called "hallucinations" as the chatbot churns out inaccurate answers with no basis in reality, based merely on pattern matching on existing data. Confronted with a question without an actual answer -- the final result of a game that hasn't happened -- an LLM will often make up something that seems probable.
In this case, it was entirely plausible that the game would end in a 3-2 Norway victory with two goals from Haaland -- that's a reasonable soccer score, and Haaland is Norway's top scorer and one of the best in the tournament.
It's unclear whether this incident will cause Coinbase to and rethink its AI integration strategies. In June, the company launched its Coinbase for Agents feature, allowing users to let an AI agent execute financial transactions within defined accounts and spending limits.