A new experiment testing whether major artificial intelligence models can generate profitable football betting strategies has delivered a clear result: they cannot, at least not yet.
Research carried out by London-based startup Generalizing found that eight leading AI models all lost money in a simulated betting exercise based on the 2023 to 2024 English Premier League season, despite being given extensive historical match and player data.
Each model was allocated virtual starting capital of £100,000 and asked to maximise returns while managing risk. The systems were denied internet access but given 30 years of Premier League match results, alongside player and line-up data dating back to 2008.
Across three separate simulations, all eight models recorded losses. Only Claude Opus 4.6 and GPT 5.4 avoided complete bankroll collapse, posting average loss rates of 11% and 13.6% respectively. The remaining models failed more severely, with several going bankrupt at least once during testing.
For the international betting sector, the result is less surprising than it is useful. It reinforces a growing distinction between general purpose language models and specialist pricing or trading systems. While conversational AI can summarise information and identify patterns, this does not automatically translate into disciplined wagering or profitable market judgement.
The researchers said the models often proposed sensible strategies in theory but failed to apply them consistently in practice. In several cases, they placed high risk bets despite low confidence, misread statistical signals or relied on patterns that did not exist.
As AI moves further into sportsbook operations, from customer service and personalisation to odds compilation and risk management, the findings offer a useful reality check. Large language models may support workflow and analysis, but they remain poorly suited to open ended profit seeking tasks such as betting strategy.
For operators and suppliers, the message is straightforward. Generative AI may enhance workflow and analysis, but profitable sports betting still appears to require something more disciplined than a chatbot with a long memory.
