Researchers have warned that “wild” gossip spread through AI bots is likely to become more frequent and widespread, causing reputational damage, shame, humiliation, anxiety and distress.
New analysis by University of Exeter philosophers Joel Kruger and Lucy Osler shows that chatbots such as ChatGPT, Claude, and Gemini don't just make things up, they create and spread gossip with negative reviews and juicy rumors that can cause real-world harm.
The harm caused by AI gossip is not a hypothetical threat. Examples of AI gossip already exist. After publishing an article about how chatbots can manipulate emotions, New York Times reporter Kevin Luce learned that chatbots labeled his writing as sensational and accused him of poor journalistic ethics and unscrupulousness. Other AI bots falsely detail people's involvement in bribery, embezzlement, and sexual harassment. These gossipy AI-generated outputs cause real-world harm, including reputational damage, shaming, and social unrest.
This study outlines how chatbots gossip both to human users and to other chatbots, but in a different way than humans. This could lead to far more widespread potential harm than false information spread by chatbots.
Gossip between bots is especially dangerous because it occurs unconstrained by the social norms that suppress human gossip. It continues to embellish and exaggerate unchecked, spreading rapidly in the background and moving from bot to bot, causing significant damage.
“Chatbots often say unexpected things, and when you're talking to them, you can feel like there's a human on the other end of the interaction,” Dr. Osler said. “As chatbots become more sophisticated, this feeling will become more common.”
“Chatbot ‘bullshit’ can be deceptive and seductive. When interacting with a chatbot, the chatbot sounds authoritative, its data set is beyond what a single human could know, and false information is often presented alongside information known to be true. It's easy to take that output at face value.
“This trust can be dangerous. Unsuspecting users can develop false beliefs that lead to harmful actions and biases based on discriminatory information propagated by chatbots.”
This study shows how the trend towards increasingly personalizing chatbots is being guided by the desire for us to rely more on these systems and for chatbots to have more access to our lives. It is also done to strengthen our sense of trust and drive us to develop richer social relationships with them.
Dr. Krueger said: “Designing AI to engage in gossip is another way to ensure an increasingly strong emotional bond between users and bots.
“Of course, bots are not interested in fostering emotional connections with other bots; they don't get the same 'stimulus' from spreading gossip as humans do. However, certain aspects of how bots spread gossip mirror the connection-promoting nature of human gossip, while also potentially making gossip between bots even more harmful than gossip involving humans. ”
Researchers predict that gossip between users and bots could become more common. In such cases, users may seed the bot with various chunks of gossip, knowing that the latter will rapidly spread the gossip in its characteristically wild manner. Bots can thus act as intermediaries, responding to gossip spread by users and rapidly spreading it to others.
