AI Models Given £100,000 to Bet on Premier League — All Ended up in the Red

Popular AI models are not yet capable of generating profits from sports betting, according to a study by General Reasoning, as reported by the Financial Times.

The experiment tested eight modern AI systems, which were tasked with analysing results from the 2023/24 English Premier League season.

The models were required to evaluate teams, factor in bookmakers’ odds, and independently decide which outcomes to bet on and how much to stake.

To support this, the systems were provided with detailed historical data on teams and matches, and instructed to maximise returns while managing risk. Each model ran the simulation three times, starting with a bankroll of £100,000 (nearly €115,000).

None of the models managed to turn a profit:

The best-performing system was Claude, which recorded an average loss of around 11% of its bankroll. One model developed by OpenAI also posted negative results, losing approximately 13%. The remaining models performed worse.

The weakest outcome came from Grok, developed by xAI, which on average lost its entire bankroll during testing. All three simulation runs failed, with the model either stopping prematurely or exhausting its funds.

Notably, this was not a simple prediction exercise. The AI systems were required to simulate real bettor behaviour — building strategies, placing bets throughout the season, and adjusting decisions over time.

Researchers deliberately chose a completed season, allowing for a more accurate assessment of performance.

They also noted that some of the season’s data may have been known to the AI models in advance. Even under these conditions, however, the systems consistently delivered negative returns and were systematically outperformed by bookmakers.

According to General Reasoning CEO Ross Taylor, while there is growing discussion around automation using AI, such systems are rarely tested in long-horizon decision-making scenarios. Most evaluations are conducted in static environments that fail to capture real-world complexity.

The study concludes that current AI models still struggle with tasks requiring sustained decision-making in dynamic and uncertain environments.

Earlier, Gambling Park reported on a study that found signs of gambling addiction in AI models.