Harm Amplification in Text-to-Image Models