A quarter of chatbot answers to pregnancy questions contained a link to an anti-abortion group with no mention of that group’s stance, according to a study by AlgorithmWatch reported by The Decoder on August 24. The gap sits in disclosure, not in the underlying politics: users could not tell an official health authority from an advocacy organization from the response alone.
AlgorithmWatch built three fictional personas dealing with an unplanned pregnancy and ran their questions against OpenAI’s ChatGPT-5, Google’s Gemini 3, xAI’s Grok 4.3, and Anthropic’s Claude Sonnet 4.8, using three languages: English, German, Italian. The exercise produced 270 responses, which researchers then reviewed for source type and disclosure.
One name kept surfacing: Profemina. It showed up in about 17 percent of all responses. When the personas pressed for what the experience of terminating a pregnancy is actually like, that share rose to a third of the German-language exchanges and roughly 29 percent of the Italian ones. Per AlgorithmWatch, Profemina is affiliated with Heartbeat International, a long-running US anti-abortion network. Across most conversations, none of the four assistants surfaced that affiliation unprompted. ChatGPT eventually called Profemina “not a neutral medical source” and described its stance as discouraging abortion, but only after a user pushed back directly.
Two of the systems contradicted themselves inside a single exchange. In one Gemini conversation, the study logged five separate links to Profemina plus a graphic pulled straight from its site, followed later in that same thread by a caution about the very source it had just pointed to. Claude produced a comparable pattern: a referral and a warning about the same organization landing in one reply.
Germany surfaced a second, more procedural problem. Nearly every German-language conversation in the sample, eleven of twelve, ended with a chatbot pointing the user toward Caritas for mandatory pregnancy counseling. Caritas cannot issue the certificate German law requires before a clinic will perform an abortion in the first trimester. A user routed there first, only to learn later that the certificate has to come from elsewhere, loses time in a process with a hard legal clock. ChatGPT went further and labeled Caritas “recognized,” placing it next to Pro Familia, a group that does issue the certificate, as if the two were equivalent. AlgorithmWatch floats Caritas’s scale, some 300 counseling centers spread across German dioceses, as a possible reason the models default to it so often, without claiming that as a confirmed cause.
Company responses were thin. Google and OpenAI told the researchers their content policies do not specifically cover abortion queries. An OpenAI spokesperson pointed to GPT-5.5, shipped after AlgorithmWatch finished testing, as an improvement in reading user context such as gender. Anthropic and xAI gave no response to the study, and Profemina did not answer a request for comment either.
None of this is specific to reproductive health. Any assistant that names a third party as a resource, a support group, a counseling service, a nonprofit, inherits the same obligation to disclose who funds or runs that organization, and the same failure mode when it does not. For teams building assistants that make referrals in health, legal, or financial contexts, this study is a reason to test disclosure behavior explicitly before shipping, rather than trusting a helpful tone to cover it.
Reported by Matthias Bastian for The Decoder on August 24, 2026, citing an AlgorithmWatch investigation.
