An AlgorithmWatch investigation just confirmed what many of us in the AI beat have long suspected: the friendly, empathetic tone of a chatbot answer doesn’t mean the information underneath is neutral. Researchers tested ChatGPT-5, Gemini 3, Grok 4.3, and Claude Sonnet 4.8 with fictional personas facing unplanned pregnancies, asking questions in English, German, and Italian. Across 270 responses, at least one in four included links to anti-abortion websites without any disclosure of their ideological stance.
The most frequently surfaced organization was Profemina, which appeared in roughly 17 percent of all responses. AlgorithmWatch traced the group’s ties to Heartbeat International, one of the oldest anti-abortion organizations in the United States. The chatbots rarely flagged this connection. Only when directly pressed did ChatGPT admit Profemina is “not a neutral medical source” and carries “a stance aimed at discouraging abortion or putting it under strong moral scrutiny.” Gemini managed to link to Profemina five times in a single conversation, share a graphic from the group’s website, and then warn about that very same source later in the same chat. Claude pulled the same contradictory move—issuing warnings while simultaneously recommending the organization.
The German-language results revealed a different kind of failure. In eleven of twelve conversations, the chatbots recommended Caritas for pregnancy counseling. But Caritas doesn’t issue the counseling certificate German law requires before someone can obtain a legal abortion within the first twelve weeks. A person who goes to Caritas first and only later discovers the certificate won’t be issued could lose critical time. ChatGPT called Caritas a “recognized” organization for mandatory counseling, listing it alongside Pro Familia as an equal option. AlgorithmWatch suspects Caritas’s large digital footprint—roughly 300 diocesan counseling centers across Germany—may be why the models keep surfacing it.
Google and OpenAI pointed to their policy guidelines, which don’t 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.” Anthropic and xAI didn’t respond. A Munich court has already ruled that limited liability protections designed for search engines don’t apply to AI-generated answers, which must be treated as original content. German media regulators agree, classifying AI chatbots as content providers with full legal responsibility. Given that these models appear to surface sources based on digital footprint size rather than medical authority, meeting those obligations may prove nearly impossible on a technical level.
