Chatbots failed 25% of pregnancy tests — and hid the agenda

an algorithmwatch analysis of chatgpt, gemini, grok, and claude using fictional pregnant personas found the models linked to anti-abortion websites in at least 25% of queries. the models failed to disclose the ideological agenda of groups like profemina, appearing in 17% of responses, and sometimes contradicted themselves by warning against a source they just recommended. additionally, chatbots in germany frequently recommended caritas, a charity that cannot issue the state-mandated certificate required for abortions, potentially causing users to miss legal deadlines. the findings highlight a lack of specialized policy or fine-tuning for high-stakes medical and legal queries.