Major LLMs direct pregnant users to anti-abortion sites without disclosure

The German-specific finding is the sharpest detail in the underlying audit: routing users to Caritas for pregnancy counseling in a jurisdiction where that organization cannot issue the legally required Beratungsschein certificate means the model is not just editorially biased but potentially blocking access to a statutory right. That is a compliance failure, not merely a content moderation lapse.

This is largely disconnected from recent activity in our archive, so it belongs to a broader ongoing conversation about LLM behavior on sensitive health topics, one that has been building across academic audits and civil society investigations over the past two years. The core issue sits at the intersection of retrieval source selection, the absence of mandatory provenance disclosure, and the question of whether foundation model providers have any duty of care when queries carry medical stakes. AlgorithmWatch's methodology, auditing four major models simultaneously across a single sensitive domain, is a template other researchers are likely to replicate for reproductive health, mental health, and medication queries.

Watch whether any of the four named providers (OpenAI, Google, xAI, Anthropic) issue a formal response to AlgorithmWatch's methodology within the next 60 days, and whether that response includes a commitment to source-bias labeling for health queries specifically. Silence or a generic safety statement would confirm that no internal escalation reached product teams.