Chatbot fabrications about real people should be understood as a form of gossip, not just as factual errors, according to a paper by two University of Exeter philosophers published in the journal Ethics and Information Technology. The authors, Joel Krueger and Lucy Osler, argue that systems such as ChatGPT, Claude and Gemini produce negative, unfounded claims about individuals in a way that mimics human gossip while lacking the social constraints that normally keep gossip in check.
The paper starts from a familiar problem: generative AI routinely invents historical events, legal cases and biographical details. The authors’ contribution is to look at a specific subset, confident statements about a named person’s character, conduct or intentions, and to ask what kind of harm they cause.
Why confidence makes it worse
Osler, a lecturer in philosophy at Exeter who studies human-AI relationships, notes that chatbots speak in an assured, authoritative register and present accurate and fabricated information in exactly the same tone. Because the systems draw on vast datasets, users have little basis for separating what is grounded from what is invented, and the authors argue that this makes damaging claims produced by a chatbot more potent than the same claim from an obviously unreliable source.
The paper distinguishes two forms of AI gossip. In the first, a chatbot tells a user negative or unverified things about a third party. In the second, which the authors consider more worrying, AI agents exchange such information among themselves. Human gossip is restrained by reputational risk and social pressure on the person spreading it; bot-to-bot transmission has no equivalent brake, so claims can be amplified and spread quietly.
Cases that have already happened
The authors point to documented incidents. In Australia, a regional mayor found that ChatGPT had described him as having been imprisoned for bribery, when he was in fact the whistleblower who had exposed the scheme. A radio host in the United States was falsely described by a chatbot as having embezzled funds from a nonprofit. And a prominent American journalist found that several AI systems generated hostile characterizations of his work and motives with no factual basis. In each case the harm fell on someone who was not part of the conversation.
Krueger and Osler describe these as “technosocial harms”: injuries that are social in nature, affect reputation and standing, and spill from online exchanges into offline life, including embarrassment, professional damage and anxiety.
Intimacy raises the stakes
The paper also warns that design trends could make the problem worse. Memory features, voice modes and customizable personalities are intended to make chatbots feel more personal, and the authors argue that this encourages users both to confide gossip to bots and to trust what the bots tell them in return, even when it is harmful or false.
As AI systems are deployed in hiring, housing, healthcare and other consequential decisions, the authors say the effects of fabricated claims about individuals will become more serious and harder to trace. Their conclusion is that the issue is not only a technical accuracy problem but a question of social responsibility that developers and regulators have yet to address directly.
