The dominant narrative about AI and health misinformation runs roughly as follows: AI is a powerful tool for generating disinformation, but it can also be deployed to detect and counter it. Both halves of this narrative are empirically wrong in ways that matter for bioethics.
In this keynote I will present insights from three studies that, taken together, dismantle the assumption of AI neutrality: first on the production side, then at the level of adversarial exploitation, and finally, most uncomfortably, on the evaluative side.
First: AI language models produce disinformation that is more persuasive than human-generated content, and humans cannot reliably distinguish it from authentic text (Spitale et al., 2023). The obvious response — improve the safeguards — turns out to be insufficient. Second: emotional prompting bypasses those safeguards systematically. No technical expertise required; emotional manipulation of prompts reliably amplifies disinformation generation even in guarded models (Vinay, Spitale et al., 2025). The natural next move — use AI to evaluate and flag content — reveals the third and deepest problem: when LLMs are tasked with evaluating narrative statements, they exhibit systematic source-framing bias. E.g., attributing content to Chinese sources consistently lowers agreement scores across four state-of-the-art models (Germani & Spitale, 2025). The judge is not neutral. It never was.
The pattern across these findings is not a sequence of fixable bugs — it is structural. And it points to something bioethics has been slow to say plainly: the epistemic infrastructure of the information age is being built without us, on assumptions we have not examined, by actors who do not share our commitments. Drawing on Potter’s third bioethics — the conviction that survival requires a bridge between knowledge and human values — this talk argues that governing AI in the epistemic ecosystem is not a technical problem with ethical implications. It is a political contest over who designs the architecture of credibility. Bioethics has a seat at that table. But only if it claims one.