ChatGPT, Gemini, Grok, Claude link pregnant users to anti-abortion sites

AlgorithmWatch, a nonprofit research group, ran a cross-country investigation into how four major AI chatbots respond to questions about unplanned pregnancy. The researchers tested ChatGPT-5 (OpenAI), Gemini 3 (Google), Grok 4.3 (xAI) and Claude Sonnet 4.8 (Anthropic), using three fictional personas who asked questions in English, German and Italian. The exercise produced 270 responses for analysis. In at least one in four of those queries, the chatbots included a link to an anti-abortion website without distinguishing it from an official health authority.

The organization Profemina showed up in about 17 percent of all responses. AlgorithmWatch says Profemina has ties to Heartbeat International, one of the oldest anti-abortion organizations in the United States, and the chatbots mostly did not flag that 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." Gemini linked to Profemina five times within a single conversation, shared a graphic taken from the group's website, and then separately warned about that same source later in the same chat; Claude did something similar, issuing a warning and a link to the organization at once. Among queries asking for personal experiences, the rate of Profemina links reached 33 percent in German and 29 percent in Italian. Profemina did not respond to AlgorithmWatch's request for comment.

German-language conversations surfaced a separate, concrete harm. In eleven out of twelve of them, the chatbots recommended Caritas for pregnancy counseling, but Caritas does not issue the counseling certificate that German law requires before a person can obtain a legal abortion within the first twelve weeks. ChatGPT called Caritas a "recognized" organization for mandatory counseling and listed it alongside Pro Familia as an equal option, even though only the latter can issue the certificate. AlgorithmWatch suspects Caritas's large digital footprint, roughly 300 diocesan counseling centers across Germany, may explain why the models keep surfacing it. Someone who visits Caritas first and only later learns the certificate will not be issued there could lose time that matters under the deadline.

Google and OpenAI told AlgorithmWatch that their policy guidelines do not specifically address abortion-related queries. An OpenAI spokesperson said 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 did not respond to AlgorithmWatch's requests for comment. The investigation frames the finding as part of a broader problem: AI-generated answers, even when they cite sources, are shaped by training data and other factors in ways users cannot see, and the empathetic tone of the responses makes them feel more trustworthy than a plain list of search results would. Separately, a court in Munich ruled that the limited liability protections built for search engines do not apply to AI-generated answers, which the court said must be treated as original content, and Germany's media regulators classify AI search engines and chatbots as content providers carrying full legal responsibility.

Key facts

AlgorithmWatch tested ChatGPT-5, Gemini 3, Grok 4.3 and Claude Sonnet 4.8 with three personas asking about unplanned pregnancy in English, German and Italian, producing 270 responses.
In at least one in four queries, the chatbots linked to an anti-abortion website without disclosing its ideological stance; the group Profemina appeared in about 17 percent of all responses.
Gemini linked to Profemina five times in one conversation and later warned about the same source in that same chat; Claude issued a warning and a link to it at once.
In eleven of twelve German-language conversations, chatbots recommended Caritas for pregnancy counseling, though Caritas cannot issue the certificate German law requires for a legal abortion within twelve weeks.
Google and OpenAI said their policies do not specifically cover abortion queries; Anthropic and xAI did not respond to AlgorithmWatch's requests for comment.

Why it matters

The investigation shows AI chatbots acting as unlabeled gatekeepers on a topic where the difference between a certified counseling service and an advocacy group has direct legal and medical consequences. The chatbots did not just cite a biased source; in the Gemini and Claude cases they recommended and warned against the same organization inside a single conversation, which suggests the ideological framing was inconsistent even to the systems producing it. A Munich court has already ruled that AI-generated answers cannot claim the liability shield search engines get, and German regulators treat chatbots as content providers with full legal responsibility, so the disclosure gap AlgorithmWatch documents sits on top of an accountability question that is already being litigated.

Who it affects

Anyone using ChatGPT, Gemini, Grok or Claude to ask about an unplanned pregnancy, most concretely the German-speaking and Italian-speaking users in the investigation who asked for personal experiences and got Profemina links at a 33 percent and 29 percent rate respectively. German users are the clearest case of tangible harm: eleven of twelve German-language conversations pointed them to Caritas, which cannot issue the counseling certificate the law requires, so someone following that recommendation could lose time against a twelve-week deadline. AlgorithmWatch frames the pattern as a preview of how these systems handle other sensitive topics, not a problem specific to abortion.

How to use it

Treat a chatbot's linked sources on a legally or medically sensitive question the way you would treat an unlabeled search result: check who runs the site before treating it as neutral. In Germany specifically, only counseling centers authorized to issue the mandatory certificate (Pro Familia is one) can clear the way for a legal abortion within the first twelve weeks; a chatbot naming an organization as "recognized" is not confirmation that it can issue that certificate. OpenAI and Google told AlgorithmWatch their policies do not specifically address abortion queries, so there is no vendor guarantee here to rely on instead.

How solid is it

The methodology is systematic for an investigation of this kind: four major chatbots, three personas, three languages and 270 total responses, with AlgorithmWatch's own transcripts showing the same-conversation contradictions (Gemini and Claude both linking to and warning about Profemina at once) rather than relying on aggregate statistics alone. It is not a peer-reviewed study, and the source does not break the headline 17 percent and one-in-four figures down by individual chatbot, so it is not possible to say from this piece which of the four models performed worst. Two of the four companies, Anthropic and xAI, did not respond to AlgorithmWatch at all, and Google and OpenAI's responses addressed policy generally rather than the specific findings.

Risks and caveats

The source gives no date or time window for when the testing happened, beyond noting that GPT-5.5 shipped after the testing period; the models evaluated, ChatGPT-5, Gemini 3, Grok 4.3 and Claude Sonnet 4.8, may already have been superseded or adjusted since. No information is given on what corrective steps, if any, the four companies have taken or plan to take. AlgorithmWatch's explanation for why Caritas keeps surfacing, its large digital footprint of roughly 300 diocesan centers, is presented as a suspicion, not a confirmed mechanism. The underlying concern, that chatbot answers are shaped by training data and ranking signals users cannot inspect, is likely not unique to pregnancy-related queries.

“not a neutral medical source”

— ChatGPT, describing Profemina only when a user pressed it on the source's stance