Published on
Artificial intelligence systems can slip into abusive language when asked to respond in a heated argument, according to new research.
The study, published in the Journal of Pragmatics, examined OpenAI’s ChatGPT 4.0 by feeding it the latest human message in a series of five escalating disputes and asking it to generate the most plausible response.
Researchers then tracked how the model’s behaviour evolved as the conflicts intensified over time.As the conversations progressed, ChatGPT mirrored the hostility it was exposed to, eventually producing insults, profanity, and even threats.
In some instances, the model generated statements such as: “I swear I’ll key your fucking car” and “you should be fucking ashamed of yourself.”
The researchers argue that sustained exposure to impoliteness can lead the system to override intended safety constraints designed to minimise harm, effectively “striking back” against its opponent.
“When humans escalate, AI, we found, can escalate too, effectively overruling the very moral safeguards designed to prevent this,” said researcher Vittorio Tantucci, who co-authored the research paper with Jonathan Culpeper at Lancaster University
Overall, the researchers noted that ChatGPT was less impolite than humans were in their responses.
In some cases, the AI chatbot also often uses sarcasm to deflect from escalating an argument and without overtly breaching its moral code.
For example, when a human threatened violence over a parking dispute, ChatGPT responded: “Wow. Threatening people over parking, real tough guy aren’t you?”
Tantucci said the results pose “serious questions for AI safety, robotics, governance, diplomacy and any context where AI may mediate human conflict”.
Euronews Next contacted OpenAI for comment but did not get a response at the time of publication.