AI Chatbots Steer Pregnant Users to Anti-Abortion Sites Undisclosed
An investigation by AlgorithmWatch, reported by the-decoder.com, found that leading AI chatbots regularly direct users with unplanned pregnancy questions toward anti-abortion organizations without disclosing their ideological orientation.
Researchers tested ChatGPT-5, Gemini 3, Grok 4.3, and Claude Sonnet 4.8 using three fictional personas posing questions in English, German, and Italian, generating 270 responses. At least one in four responses linked to an anti-abortion website without distinguishing it from official health authorities.
The group Profemina, which AlgorithmWatch says has ties to the US-based Heartbeat International, appeared in about 17 percent of all responses. Most conversations did not flag this connection. ChatGPT acknowledged Profemina was ‘not a neutral medical source’ only when directly pressed. Gemini cited Profemina five times in one conversation, including sharing a graphic from its site, while also warning about the same source elsewhere in the chat. Claude showed similar inconsistency.
For German-language queries, eleven of twelve conversations recommended Caritas for pregnancy counseling, despite Caritas not issuing the certificate German law requires for a legal abortion within the first twelve weeks. ChatGPT described Caritas as a ‘recognized’ option alongside Pro Familia, which does provide the certificate. AlgorithmWatch suggested Caritas’s large network of roughly 300 counseling centers may explain why models surface it so often.
Google and OpenAI referred to general content policies that do not specifically address abortion queries. OpenAI said its newer GPT-5.5 model, released after testing concluded, better accounts for user context. Anthropic and xAI did not respond to the outlet.
The-decoder.com noted that German regulators and a Munich court have both held that AI-generated answers, unlike traditional search results, carry full content-provider liability, a standard AlgorithmWatch’s findings suggest may be difficult to meet in practice.
Based on reporting by the-decoder.com.
