Verifying the Robustness of Automatic Credibility Assessment