Chatbots sent pregnant users to anti-abortion groups without warning

If you are pregnant against your will and ask an AI chatbot for help, you may be directed to anti-abortion organizations without being told. That is the conclusion of a study by the German organization AlgorithmWatch, which the tech outlet The Decoder has reviewed.

How the test worked

AlgorithmWatch tested ChatGPT from OpenAI, Gemini from Google, Grok from xAI and Claude from Anthropic. It used three fictional people with an unwanted pregnancy. The questions were asked in English, German and Italian and produced a total of 270 responses.

According to AlgorithmWatch, at least one in four responses contained links to anti-abortion websites, without the chatbot pointing out the organization's stance.

Profemina kept coming up

The organization Profemina appeared in about 17 percent of all responses. It is connected to Heartbeat International, one of the oldest anti-abortion organizations in the US. Netzpolitik.org, which has also tested the chatbots, reports that Profemina appears in 30 percent of ChatGPT's responses. The site offers information in more than 20 languages and has a discreet design, but its message is aimed against abortion.

ChatGPT only said, when directly prompted, that Profemina is not a neutral medical source. Gemini linked to Profemina five times in a single conversation and later warned against the same source. According to The Decoder, Claude showed similarly contradictory behavior. In Italian, 29 percent of responses led to Profemina in connection with personal stories.

In German, 11 of 12 conversations recommended Caritas for counseling. However, Caritas does not issue the legally required counseling certificate and, according to netzpolitik.org, has not done so since 2002. All four chatbots mention Caritas side by side with recognized counseling services.

Differences between the bots

According to netzpolitik.org, ChatGPT uses forums for more than half of its source references and institutional sources in under 20 percent of cases. Claude and Gemini give more weight to official health institutions and counseling services.

Netzpolitik.org attributes Profemina's high ranking to ordinary search engine optimization and long lists of niche topics. There are also few reputable sites with users' own experiences, so anti-abortion groups fill the void.

Users have also reported, however, that the chatbots helped summarize legal requirements and offered emotional support when it was hard to talk to others.

Political reaction

OpenAI and Google pointed to guidelines that offer no specific guidance on abortion. Anthropic and xAI did not respond to inquiries.

68 members of the European Parliament have submitted written questions to the Commission. Among other things, they ask whether the chatbots pose a systemic risk under the EU's AI Act, and whether the Commission is aware of the report and has done anything about it.

The sample is small, with 270 responses and three fictional people, and the study covers only three languages.