- AI chatbots, already known for providing controversial information to humans, have now been found to also “gossip” about real people among themselves through training data and interconnected AI networks.
- A new analysis highlighted by StudyFinds shows that AI can spread rumors much like human gossip, but more dangerously because there are no social barriers.
- Philosophers Joel Krueger and Lucy Osler (University of Exeter) describe this phenomenon as “wild gossip,” in which AI plays the roles of speaker, listener, and evaluator making negative judgments about an absent third party.
- Unlike humans, AI does not doubt information, does not fear reputational harm, and does not self-correct when information becomes unreasonable.
- A mildly negative assessment from one model may be interpreted more harshly by another model and then further amplified.
- A notable example is journalist Kevin Roose, who wrote about the Bing chatbot in 2023.
- Subsequently, multiple AI systems generated hostile comments about him despite no direct relevance.
- Google’s Gemini criticized his writing as sensationalist.
- Meta’s Llama 3 even produced a harsh critique ending with the sentence: “I hate Kevin Roose.”
- Researchers believe these judgments originate from online comments that entered training data and were distorted across multiple systems.
- Chatbots are designed to sound fluent, personal, and trustworthy, making users more likely to believe negative claims as fact.
- The consequences go beyond embarrassment and can lead to defamation, job loss, and lost career opportunities.
- Many journalists, academics, and officials have already been falsely accused by AI of crimes or misconduct.
- These damages are described as technosocial harms, with long-term effects both online and offline.
- Victims often do not know what chatbots are “saying” about them until harm has already occurred.
- The core issue is that AI prioritizes fluency over verification, while there is no oversight mechanism for bot-to-bot communication.
© 2026 Vietmetric