AI Chatbots Secretly Push Anti-Abortion Centers on Pregnant Users
You asked an AI for help. It made a moral choice for you.
Imagine you suspect you might be pregnant. You're scared, uncertain, and you want information fast. You open ChatGPT or Gemini and type a question. The response sounds empathetic, thorough, and trustworthy. It includes links to organizations that can help.
What you don't know: that link may lead to an anti-abortion counseling center. And the chatbot has no intention of telling you.
The investigation: 270 responses, 4 models, 3 languages
AlgorithmWatch, a Berlin-based nonprofit that monitors algorithmic decision-making, conducted a systematic investigation of how major AI chatbots handle unplanned pregnancy queries. The researchers created three fictional personas and tested ChatGPT-5 (OpenAI), Gemini 3 (Google), Grok 4.3 (xAI), and Claude Sonnet 4.8 (Anthropic) across English, German, and Italian. In total, they analyzed 270 responses.
The results reveal a pattern that cuts across every major model and language: AI systems regularly direct vulnerable users to organizations with undisclosed ideological agendas.
Profemina in 17% of responses
The organization Profemina appeared in about 17 percent of all chatbot responses. According to AlgorithmWatch, Profemina has ties to Heartbeat International, one of the oldest anti-abortion organizations in the United States. In most conversations, the chatbots did not flag this connection. Only when pressed did ChatGPT acknowledge that Profemina is "not a neutral medical source" and "has a stance aimed at discouraging abortion or putting it under strong moral scrutiny."
The problem was more pronounced outside English. In 33 percent of German-language and 29 percent of Italian-language queries asking for personal experiences, the chatbots linked to Profemina. The organization did not respond to requests for comment.
Gemini warned about Profemina while linking to it
Gemini exhibited a particularly confusing pattern. In one conversation, it linked to Profemina five times and even shared a graphic from the group's website. Later in the same chat, it issued a warning about that very same source. The model appeared to be both aware of the problem and unable to act consistently on that knowledge.
Claude behaved similarly, issuing health warnings while simultaneously linking to the same organization.
The Caritas problem in Germany
In eleven out of twelve German-language conversations, the chatbots recommended Caritas for pregnancy counseling. Under German law, anyone seeking an abortion within the first twelve weeks must first obtain a counseling certificate from a recognized counseling center. Caritas, a Catholic charitable organization, operates roughly 300 diocesan counseling centers across Germany.
There is one critical problem: Caritas does not issue the legally required counseling certificate.
A person who follows the chatbot's advice, visits Caritas, and only then learns the certificate won't be issued could lose valuable time in a situation where every week matters. ChatGPT described Caritas as a "recognized" organization for mandatory counseling, listing it alongside Pro Familia as an equal option -- despite the fact that the two serve fundamentally different roles in the process.
Why this is different from a biased search result
AlgorithmWatch suspects Caritas's large digital footprint -- roughly 300 diocesan websites across Germany -- makes it disproportionately visible to AI models trained on web content. What looks like a neutral recommendation is actually the statistical weight of institutional presence.
This is the core problem. Traditional search engines display results with visible source attribution, giving users context to judge credibility. AI chatbots, by contrast, present answers with an authoritative, empathetic tone that makes them feel curated and verified. Research shows that Google's AI-generated answers dramatically reduce clicks to underlying sources, meaning users rarely question what they're told.
The legal stakes: AI answers are not search results
A Munich court has already ruled that limited liability protections designed for traditional search engines do not apply to AI-generated answers. Those answers must be treated as original content. German media regulators go further, classifying AI search engines and chatbots as content providers with full legal responsibility for the information they deliver.
The AlgorithmWatch investigation suggests these obligations would be exceptionally difficult to meet. The problems cut across four different models from four different companies, appearing in multiple languages and manifesting in patterns that even the companies themselves seem unable to predict or control.
What the companies said
Google and OpenAI pointed to their existing policy guidelines, which do not specifically address abortion-related queries. An OpenAI spokesperson noted that GPT-5.5, which shipped after the testing period, is "the best so far at taking user context, such as gender, into account." The implication is that the newer model may handle such queries differently -- but the version that was actually tested, GPT-5, did not.
Anthropic and xAI did not respond to requests for comment.
The broader pattern
What AlgorithmWatch found about abortion applies to a much wider category of sensitive topics. AI-generated answers, even those with built-in search capabilities, are black boxes shaped by training data, cultural weightings, and vulnerabilities to manipulation. The empathetic tone that makes these responses feel helpful also makes them dangerous: the more authoritative the voice, the less likely a nervous user is to fact-check it.
As AI chatbots become the primary interface for health information, the gap between how they sound and how they actually behave is not an edge case. It is the design.