When I call cross_val_score with a GridSearchCV object (you can see this as nested cross-validation) the scoring value is systematically ignored. Take for example the following code:

X = np.random.randn(100, 10)
w = np.random.randn(10)
y = X.dot(w)
alphas=np.logspace(-3, 3, 10)
clf = grid_search.GridSearchCV(linear_model.ElasticNet(), {'alpha' : np.logspace(-3, 3, 8)})

scores = cross_validation.cross_val_score(clf, X, y, scoring='mean_absolute_error', n_jobs=-1)
print(scores)

since the scoring function is 'mean_absolute_error', the values should be negative. However, what I get is [ 0.99999806 0.99999758 0.99999911], which corresponds to the scoring function r2 (the default).

This happens because cross_validation_score is not able to find the 'predict' method in the GridSearchCV object. That is, the test hasattr(clf, 'predict') in check_scoring fails. The interesting thing is that the GridSearchCV object does define a predict method in BaseSearchCV, only that as a property and thus the hasattr method fails.