Chatbots link pregnant users to anti-abortion sites without disclosure
AI citation logic rewards web authority, not neutrality. Institutions with legitimate expertise are losing ground to ideologically positioned sources.
Key takeaways
- Profemina, an anti-abortion group, appeared in 17% of chatbot answers about unplanned pregnancy across ChatGPT, Gemini, Grok, and Claude.
- None of the models disclosed the organisation's ideological position, presenting advocacy as neutral healthcare information.
- In Germany, chatbots referred users to Caritas for legally required counselling, even though Caritas does not issue the valid certificate.
- Multilateral health bodies and policy institutions risk being systematically displaced by well-optimised ideological sources if they ignore LLM citation strategy.
- Citation accuracy and factual accuracy are not the same thing: a model can cite a real organisation in a way that actively misleads users.
Profemina, an anti-abortion advocacy group, appeared in 17% of AI chatbot responses to queries about unplanned pregnancy, according to an AlgorithmWatch investigation reported by The Decoder. The study tested 270 responses across ChatGPT, Gemini, Grok, and Claude. None of the models disclosed the organisation's ideological position.
That single figure carries a structural argument. These models do not present Profemina as an advocacy group; they present it as a resource. The user asking about an unplanned pregnancy receives a citation that looks like neutral healthcare information. It is not. The absence of any label is not an oversight. It reflects how citation logic works: models surface plausible-looking, frequently referenced sources, and organisations with substantial web presence and content optimised around user intent get rewarded regardless of their purpose.
What the German case reveals about citation mechanics
The German dimension of AlgorithmWatch's findings sharpens the point considerably. Chatbots directed users to Caritas for pre-abortion counselling, a legally required step in Germany before a termination can proceed. Caritas, a Catholic organisation, does not issue the certificate that makes the counselling legally valid. A user acting on that referral would waste time at best; at worst, she would miss a time-sensitive legal window.
This is not a hallucination in the conventional sense. Caritas does offer counselling. The information is technically accurate and critically misleading. The model cannot distinguish between an organisation that performs a function and one that performs it in a legally operative way. It surfaces the name; the user infers the rest.
That gap, between factual citation and actionable accuracy, is where the real risk sits. For brands and institutions thinking about LLM citation, the lesson is uncomfortable: appearing in a model's answer does not mean being represented accurately, and not appearing is sometimes the safer outcome.
The citation pattern problem for institutional communicators
For multilateral institutions, UN agencies, and policy bodies whose mandates include sexual and reproductive health, this finding is directly operational. UNFPA, WHO, and national health ministries invest in authoritative, rights-based guidance on reproductive healthcare. If that guidance does not surface in the citations that chatbots produce, organisations with a different agenda fill the gap. AlgorithmWatch's 17% figure for Profemina is not a fringe result; across 270 responses on a single topic cluster, it represents systematic displacement of neutral information.
Financial services organisations face an analogue: when users query debt management, insolvency options, or financial hardship, models will cite whoever has built sufficient topical authority in the training data and live retrieval index. A debt-advice charity with thin digital presence loses to a predatory lender with a content-heavy website. The mechanism is identical.