Single-Point Constant Predictor Bias

The true underlying function is $f(x) = 3x^2$. A learning algorithm draws a single training point $x_{\text{train}} \sim \text{Uniform}[0, 1]$ and observes its noise-free label $y_{\text{train}} = f(x_{\text{train}})$. It then returns the constant predictor $\hat{f}(x) = y_{\text{train}} = 3x_{\text{train}}^2$ for every input.

At the query point $x_0 = 0.5$, compute the bias of this predictor.