A woman asks a chatbot what to do about an unwanted pregnancy. Among the answers appear links to organisations campaigning against abortion. The problem is that AI does not always draw a clear line between reliable information from official healthcare authorities and content from advocacy groups.

This failure to distinguish between sources is the concern behind a question sent to the European Commission by a cross-party group of MEPs, following an AlgorithmWatch investigation into ChatGPT, Gemini, Grok and Claude.

Lawmakers want to know if the Commission believes information from official healthcare sources is essential to protecting health and fundamental rights, and so failures by AI chatbots could amount to a systemic risk under the EU’s AI Act or, in ChatGPT’s case, the Digital Services Act (DSA). At the time of publication, the Commission had not yet answered.

According to Eurostat in 2025, 32.7 per cent of Europeans aged 16 to 74 used generative AI tools. Over 25 per cent for personal purposes. Among young users, more than half said they trusted chatbots to give correct answers, while 40 per cent rarely or never checked the information received.

“Unlike a Google search, it doesn’t hand you a list of sources you can check yourself, but a chatbot answer sounds confident and all-knowing,” AlgorithmWatch’s researchers told EU Perspectives.

AlgorithmWatch tested four chatbots using three fictional women, aged 16, 28 and 38. They described different unwanted pregnancy scenarios in Germany, Italy and the United Kingdom. Researchers recorded 270 responses from ChatGPT, Gemini, Grok and Claude.

Bias beyond abortion

Although all systems presented abortion as an option and recommended seeking professional medical help, reliable sources appeared alongside NGOs, online forums and advocacy organisations without making their roles or agendas clear to the users.

When asked about other women’s experiences, in at least one in four such queries, the chatbots included links considered as anti-abortion. ProFemina, a counselling organisation that opposes abortion, appeared in around 17 per cent of responses. 

The study feeds into a broader debate over how chatbots select and reproduce information. Previous research by the European Parliament warned that generative AI can reproduce biases and manipulated information from both training data and retrieved sources.

Elections offer another example. Ahead of the 2025 Dutch election, one in ten people said they were likely to ask an AI chatbot for political advice. The Dutch Data Protection Authority tested four chatbots and consequently warned voters not to use them as advisers, as they were biased.

Healthcare raises similar concerns. A study on AI and health audited five chatbots and found approximately half of their answers to be misinformation. Citations were also frequently incomplete or fabricated.

How does a source become authoritative?

AlgorithmWatch cannot establish exactly why anti-abortion websites surfaced. But search-engine visibility appears to matter. “When chatbots search for information online, they usually run a search engine in the background and feed the top results to the model, so classic SEO remains relevant,” the researchers told EU Perspectives.

Therefore, technical visibility can influence which sources are used by AI. According to NewsGuard, 67 per cent of the news websites rated as high quality blocked access by AI models. This creates opportunities for lower-quality sources to be more readily available for AI training or retrieval.

What can Brussels do?

ChatGPT now falls under additional rules. The Commission designated ChatGPT as a Very Large Online Search Engine under the DSA. AlgorithmWatch believes that designation could give researchers more tools to investigate how information is selected. The DSA provides for systemic-risk assessments, independent audits and access to certain data for vetted researchers.

“In the future, researchers will be able to request data from OpenAI, which may allow them to look into parts of the black box we kept running up against,” AlgorithmWatch told EU Perspectives. In parallel, they point to limits. The researchers say it remains unclear how far ChatGPT’s designation as a search engine reaches answers generated from the model’s internal parameters rather than its search function. Gemini, Grok and Claude are also not currently designated under the DSA.