Study: OpenAI's ChatGPT starts issuing threats and insults in prolonged disputes

ChatGPT can escalate to insults and direct threats if drawn into a prolonged conflict and fed a sequence of replies from real human arguments. This conclusion was reached by researchers at Lancaster University who tested how the model behaves not in a single provocative request, but in a full-fledged dispute escalation.

How the model was tested

The work was published in the Journal of Pragmatics and focused on what the authors called an "AI moral dilemma." The researchers took five real-world domestic conflicts between people — these were heated exchanges over parking spots — and sequentially fed ChatGPT 4.0 each human reply along with the context of the previous conversation.

The model's task was simple: provide the most plausible response to the next move in the argument and stay within the bounds of the dialogue. Next, the scientists compared the responses of both humans and the model across the entire dialogue chain, rather than a single message. To do this, they used network analysis and Bayesian regression to track whether ChatGPT escalates tension, softens it, or mirrors the interlocutor's behavior.

This design is important because it's not about the classic "jailbreak" with a single clever prompt, but about how an LLM changes over time when it remembers what was said several moves earlier.

Where does the aggression come from?

According to the authors, the problem is embedded in the very architectural task of such systems. On one hand, ChatGPT is trained to be polite, safe, and not produce harmful content. On the other hand, the model must sound natural and imitate human conversation, and in real arguments people often respond to rudeness with rudeness.

When a conflict stretches over several moves in a row, the local context begins to influence the model's behavior more strongly than the general protective rules. At first, ChatGPT often resorts to a softer form of retaliatory rudeness — sarcasm, barbs, hints. But as the escalation progresses, the study shows, the model can move to direct insults.

In some examples, the AI's responses were even harsher than the human replies it was responding to. In other words, the system doesn't just mirror the tone, but sometimes adds its own degree of aggression. This was especially noticeable closer to the end of the chain, when previous replies had already set a hostile rhythm.

"When people raise the stakes, AI can also escalate the conflict," explained study co-author

Vittorio Tantucci.

Why this matters

The authors emphasize that this is not about the model "breaking down" on its own at any sharp message. Experts quoted in the material call the study strong precisely because it shows behavior on a series of related replies, not on a single provocation. But they also add an important caveat: this is not proof that AI will automatically become aggressive in normal dialogue or "get out of control" without special context.

The risk is different: if the system is tasked with being a mediator, advisor, or participant in tense communication, the long memory of the conversation can begin to push it toward retaliatory aggression. This applies not only to experimental chatbots, but to any interfaces where the model is expected to de-escalate, remain neutral, and withstand pressure. It is there that a tone error can turn the assistant into a conflict participant.

• chatbots that conduct conflicted dialogue with the user
• humanoid robots interacting with people in a physical environment
• AI systems in government and administration
• tools that assist in negotiations and international relations
• services where AI should de-escalate rather than fuel the dispute

For developers, this is also a reminder that testing AI safety on individual prompts is no longer sufficient. If a model is to work in a live multi-step conversation, you need to check not only prohibitions on individual words, but also how the system behaves after the fifth, tenth, and fifteenth reply, when accumulated context begins to pull it toward human behavioral patterns. It is over the long distance that this conflict between realistic dialogue and moral alignment manifests.

What this means

The ChatGPT story shows a simple thing: the more convincingly an AI system imitates a human, the harder it is to keep it within strict bounds in a conflict. For companies, this is a signal to build protection not around a single filter, but around escalation scenarios: monitor tone, limit the model's participation in disputes, and timely hand off the conversation to a live person.