Large Language Models (LLMs) can be used to summarise complex political debates in an effort to make them more easily understandable to the general public. While these summarisation tools promise to increase accessibility for the average citizen, they run the risk of introducing systematic biases that misrepresent the original source texts. Building on research from studies of political parties and gender and politics, we evaluate the extent of bias introduced by LLMs when they are used to reduce the quantity and linguistic complexity of parliamentary debates. We hypothesise that AI systems amplify existing societal biases such as gender stereotypes in the resulting summaries of political discourse. Focusing on debates within the European Parliament, we compare a variety of outputs generated by open and closed source LLMs through a comparative analysis of political preference measurements. We contrast scaled position estimates of AI-generated speech summaries with position estimates derived from the original parliamentary speeches from which summaries are generated. Through a systematic comparison of both open and closed source models, our findings reveal the extent of bias introduced into AI-generated texts. These findings have significant implications for the use of AI as a summarisation tool in social and political research.