Hallucinations refer to the tendency in Large Language Models to generate content that is not supported by external sources. Their presence reduces the reliability of the model and spreads misinformation. This study deep-dives into a strategy that uses a model’s inherent knowledge to combat hallucinations, called intrinsic self-correction. We apply self-correction multiple times in an extractive question-answering setting on multiple models of different families and sizes. Our findings suggest that the approach successfully mitigates hallucinations in the Qwen and Gemma families, but is not completely transferrable to others like Llama. A one-size-fits-all approach to self-correction is therefore, ineffective, requiring tailoring to a model’s specific reflective and instruction-following abilities to prepare for the adoption of this strategy into the real world.