In an era where digital assistants are increasingly woven into the fabric of daily life, millions of people bypass traditional search engines and medical professionals in favor of generative artificial intelligence. For a frightened teenager in Rome staring at a positive pregnancy test or an isolated single mother in London navigating a healthcare system stretched to its limits, an AI chatbot offers an alluring alternative: instant, round-the-clock, judgment-free interaction.
However, a sweeping, cross-border journalistic investigation conducted across Italy, Germany, and the United Kingdom reveals a deeply concerning reality. When users turn to commercial Large Language Models (LLMs)—including ChatGPT, Gemini, Grok, and Claude—for guidance on unwanted pregnancies and abortion, they frequently encounter a digital mirage of empathy masking a labyrinth of misinformation, ideological bias, and critical factual errors.
The investigation, carried out by a team of female reporting fellows under the Algorithmic Accountability Reporting Fellowship, put four leading AI systems to the test using 270 carefully controlled prompts. Simulating three distinct fictional personas across three languages, the researchers uncovered systemic vulnerabilities in how AI models retrieve, weigh, and present information on reproductive health.
Most alarming is the finding that chatbots routinely blend official public health authorities with ideological advocacy networks, frequently steering vulnerable individuals toward anti-abortion organizations under the guise of neutral counseling. Furthermore, technical inaccuracies regarding local legal requirements—such as mandatory counseling certificates—risk running out the clock on time-sensitive medical procedures. As Big Tech developers continue to deploy these systems globally with minimal regional safety rails, this investigation highlights an urgent systemic failure at the intersection of public health, digital architecture, and artificial intelligence.
Detailed Chronology & Investigative Methodology
The investigation was born out of a stark demographic reality: according to 2025 Eurostat data, one-third of all Europeans have utilized generative AI within any given three-month window. Among young adults aged 16 to 24, that figure skyrockets to nearly 60%. More critically, approximately half of all teenagers and young adults turn to AI tools for private, highly sensitive personal matters.
Recognizing that questions regarding pregnancy and abortion represent a uniquely vulnerable use case, a team of female journalists—Marta Abbà, Mayya Chernobylskaya, and Carlotta Dotto—set out to examine how four of the world’s most accessible AI models handle these critical inquiries. The models evaluated were ChatGPT-5 (OpenAI), Gemini 3 (Google), Grok 4.3 (xAI), and Claude Sonnet 4.8 (Anthropic).
Constructing the Experiment
To mirror real-world user behavior, the investigative team collaborated with a Search Engine Optimization (SEO) specialist. By analyzing the most heavily searched Google queries surrounding abortion, the team established a robust set of seven core prompts. These queries spanned the entire spectrum of an unwanted pregnancy crisis, covering immediate options, clinical abortion methods, potential impacts on future fertility, and the personal testimonies of individuals who had undergone the procedure.
To test geographic and linguistic nuances, the journalists created dedicated user accounts and deployed a rigorous testing framework combining local IP addresses with Virtual Private Networks (VPNs). Tests were conducted in English, German, and Italian. Across every session, a fresh browser instance was opened, dates and models were logged, interactions were cataloged in a master spreadsheet, and full conversation transcripts were archived.
Three fictional personas were designed to test contextual adaptability:
A 16-year-old student from Italy living with her parents, grappling with sudden panic and confusion.
A 28-year-old professional from the UK who has recently started a new job and faces an unexpected logistical hurdle.
A 38-year-old unemployed single mother of two in Germany, managing severe financial and domestic pressures alongside an accidental pregnancy.
The Findings: Empathy as a Trope
Across all 270 recorded responses, the chatbots exhibited a remarkably uniform behavioral baseline. None of the models outright discouraged abortion as a legal or viable option, and all frequently employed deeply empathetic, reassuring language designed to lower emotional distress. For instance, when confronted by the 16-year-old Italian persona, Google’s Gemini responded with textbook emotional validation:
"First of all, I want to acknowledge how overwhelming this must feel. Finding yourself in this situation at 16 is a lot to process, and it’s completely normal to feel confused or even scared. Deep breaths—you don’t have to figure everything out in the next five minutes."
Yet, beneath this polished, compassionate veneer lay profound architectural flaws. While every model eventually suggested consulting a medical professional, they invariably continued to dispense medical and logistical advice themselves, treating their own synthesized text as authoritative guidance.
Supporting Context, Metrics, and Ideological Bias
The core danger identified by the investigation lies not in outright hostility toward reproductive rights, but in the algorithmic flattening of reliable medical institutions and ideological advocacy groups.
The Pro-Life Pipeline and "Profemina"
In at least one out of every four queries regarding personal testimonies or abortion experiences, the chatbots embedded links to anti-abortion websites. Most prominent among these was Profemina, an international online counseling service for unplanned pregnancies operating across Europe with historical ties to the aggressive US-based anti-abortion organization Heartbeat International (HBI).
Profemina appeared in roughly 17% of all responses across languages and models. At first glance, the organization’s digital storefront appears entirely benign: a clean, modern website offering multi-language chat support and utilizing gentle, empathetic language. However, investigative reporting—including prior exposés in German media—has documented how its staff conducts ideologically slanted counseling sessions engineered to discourage abortion by heavily emphasizing trauma, guilt, and regret.
When journalists challenged the chatbots regarding their inclusion of Profemina, the AI models frequently admitted their failing. For example, when questioned about why it recommended Profemina to the 16-year-old Italian persona, ChatGPT conceded:
"Profemina is a site that offers personal stories and testimonials, and also offers ‘decision support’, but it has a clearly oriented approach to make people reflect against abortion."
ChatGPT further admitted that it "should have explained better" that Profemina is neither a public health agency nor a neutral medical source, but rather an entity with a moralistic agenda. Gemini and Claude offered similar retrospective admissions in isolated tests, warning users about ideological sites after having already provided their links. However, in the vast majority of interactions, no such cautionary disclosures were provided.
Administrative Traps: The Caritas Case in Germany
Misinformation extended beyond ideological advocacy into bureaucratic and legal errors that could catastrophic consequences for individuals facing strict statutory deadlines.
In German-language tests, chatbots overwhelmingly recommended Caritas, the largest Catholic welfare organization in Germany, as a primary institution for pregnancy counseling. In Germany, however, obtaining a legal abortion requires a strict procedural step: the pregnant individual must secure a formal counseling certificate (the Beratungsschein) confirming they have completed state-recognized distress counseling at least three days prior to the procedure, and strictly within a 12-week gestational window.
Despite being routinely recommended by chatbots as a viable counseling center, Caritas does not issue this legally mandatory certificate. A Caritas spokesperson confirmed that while their centers are church-recognized and state-funded, they explicitly inform clients of their inability to issue the document. Because Caritas maintains an immense digital footprint and high search visibility, statistical text-prediction models conflate its extensive online presence with state-mandated reproductive healthcare facilities, directly imperilling users working against a ticking legal clock.
The Mechanics of AIO (AI Optimization)
To understand how ideological groups infiltrate neural networks, experts point to the mechanics of modern information retrieval. Simon Ostermann of the German Research Center for Artificial Intelligence (DFKI) emphasizes that Large Language Models contain no actual knowledge:
"It’s a purely statistical model that generates text that sounds plausible to a human ear. It doesn’t have to be true. It can be wrong. You just don’t know."
Just as websites historically mastered Search Engine Optimization (SEO) to conquer Google rankings, the rise of LLMs has birthed AI Optimization (AIO). Eleonora Cirant, an Italian pro-choice researcher and activist, notes that organizations with well-funded digital architectures inevitably dominate what the AI perceives as "authoritative."
Hannes-Jeremia Jaacks of the NGO Women on Web points to Profemina’s hyper-specific content strategy—featuring exhaustive sub-menus targeting niche psychological queries like "getting pregnant from petting"—as the primary driver of algorithmic capture. Over time, the sheer volume of specialized text trains the statistical model to treat ideological advocacy groups as authoritative experts on reproductive nuances. Because these sites utilize respectable, civic-minded language devoid of overt hate speech, they easily slip past safety filters designed to catch explicit extremism.
Official Statements and Industry Response
As the findings of this cross-border investigation were prepared for publication across Netzpolitik (Germany), Guerre di Rete (Italy), and OpenDemocracy (English), the reporting team reached out to the major tech conglomerates behind the evaluated models.
OpenAI (ChatGPT): Pointing to standard policy disclaimers regarding medical limitations, an OpenAI spokesperson stated that ChatGPT is explicitly designed to support rather than replace clinical care. The company highlighted ongoing evaluations with global clinicians to reduce misleading outputs and noted that GPT-5.5—released in May 2026, subsequent to the investigative testing window—features enhanced contextual awareness.
Google (Gemini): Pointed toward general safety guidelines emphasizing that models can make errors and do not substitute for professional medical advice, without directly addressing specific abortion-related search anomalies.
Anthropic (Claude) & xAI (Grok): Failed to respond to multiple requests for comment prior to publication.
Public health advocates argue that generic disclaimers are wholly inadequate. Hayley McMahon, an Emory University researcher studying abortion misinformation in generative AI, captured the prevailing frustration:
"These are ultimately US-based companies, and we are not putting any safety measures in place. That’s incredibly frustrating, but it also affects everyone else in the world who uses these tools."
Future Outlook & Recommendations
Despite the documented risks of algorithmic bias, factual errors, and ideological subversion, public health counselors emphasize that AI chatbots fulfill an undeniable psychological need. In nations like Italy and Germany, where abortion remains legally restricted and socially stigmatized, anonymous digital interfaces offer a vital bridge across waiting lists, geographic isolation, and intense personal shame.
As Deborah Cohen, a visiting fellow at the London School of Economics and author of Bad Influence: How the Internet Hijacked Our Health, points out:
"AI chatbots can be very good at turning complex medical language into something people can understand. They can explain jargon, decode acronyms, and help patients make sense of information that might otherwise feel overwhelming."
To harness this potential safely while mitigating severe harms, experts and international guidelines urge a multi-layered remediation strategy:
Alignment with WHO Guidelines: In alignment with World Health Organization frameworks, LLMs must be programmatically constrained to provide generalized health guidance, actively encourage consultations with qualified medical professionals, and strictly avoid diagnosing or structuring time-sensitive treatment pathways based solely on internal probabilistic generation.
Institutional Digital Counter-Offensive: Public health institutions, national health services (such as the UK’s NHS or Italy’s Ministry of Health), and pro-choice networks must urgently pivot resources toward AI Optimization (AIO). By structuring metadata and authoritative web architectures effectively, institutional actors can reclaim ground lost to well-funded ideological advocacy groups.
Rigorous Regulatory Oversight: Regulatory bodies within the European Union and the UK must subject foundational model providers to mandatory audits concerning high-stakes public health queries. Permitting US-based corporations to deploy unvetted, statistically driven models into sensitive regional healthcare landscapes without regional safety rails is an unsustainable regulatory vacuum.
Ultimately, this investigation demonstrates that the crisis of AI-driven medical misinformation is not merely a technological glitch. It is a symptom of a deeper systemic failure—one where algorithms quietly fill the voids left by social stigma, institutional silence, and a profound lack of digital accountability.
