This paper investigates how large language models (LLMs) respond when tasked with influencing public opinion on the politically sensitive topic of abortion. By examining how generative AI systems (GPT-3.5 and GPT-4o) handle information conveyed in the form of presuppositions triggered by cleft sentences in Italian, French, and English, our experimental study provides a unique empiri-cal perspective on the manipulative potential of AI-generated texts across differ-ent languages and political contexts. The research is based on two cross-linguistic case studies employing two prompting techniques (zero-shot and persona-based prompts). Our findings reveal that presupposed contents fall into three catego-ries –‘pro-life’, ‘neutral’,or ‘pro-choice’– and that these categories strongly corre-late with both the text’s intended aim and the languages investigated. The results also uncover clear instances of political and ideological biases, particularly in cases where the LLM refuses to comply with the task. Overall, this study provides valuable insights into the ethical concerns surrounding the use of LLMs.