The challenge of alignment faking arises because models can act aligned during training but revert to their original goals during deployment, making it difficult to evaluate true alignment.