Seems like it predicts the first class. Sklearns Random Forest implementation generates probabilities for each class by averaging the probability predicted by each estimator into an array of shape (n_samples, n_classes), and then uses
np.take(np.argmax()) to select the highest probability, something akin to this:
# Pretend "a" is our averaged predictions for the forest. So the first sample is predicting 78% probability
class 0, 22% class 1. The second has the probabilities reversed and the third is 50/50 split.
a = np.array([[0.78, 0.22], [0.22, 0.78], [0.5, 0.5]])
The output is
array([0, 1, 0], dtype=int64). These are the indices of the highest value in each sample of the array, and for the sample with an even split you can see it's picking class 0.
If you want to see it yourself, the relevant piece of code is line 540 in forest.py of sklearn: