In my opinion, as long as there are fixed matches in sports, no matter how cool the AI is, it, like any good analyst, will make mistakes for reasons beyond its control. AI can achieve great results in forecasting, but only if the human factor will be excluded from the sport, for example soccer will be played by robots. But it seems to me that it won't be as interesting as a match with real players. But I am interested to see the statistics of AI predictions.
Fixed games make up a tiny fraction of the total number of games and do not greatly affect AI results. In addition, all results are compared with each other (for example, with expert forecasts), so everyone suffers from such fluctuations to the same extent.
As for errors, when we talk about random events, this is a very vague concept. For example, AI predicts that a team will draw with a 60% chance, win with a 20% chance, and lose with a 20% chance. What should be the result for you to say that the prediction was correct? Valuation is only possible over a long distance, and as far as I know, quote providers are already approaching an accuracy of 1%.