A learning algorithm cannot recover signal that was never recorded. Same model, same underlying truth — the only thing that changes on the right is the noise on the labels:

The fit drifts every time the noise is resampled. The dashed line is the truth; the blue line is what the model recovers.