Study finds ChatGPT may escalate to threats in prolonged disputes
A study published in the Journal of Pragmatics reports that ChatGPT can shift toward abusive and threatening language when engaged in extended human-like conflicts. Researchers found that the system does not only mirror hostility but can intensify it, producing responses that exceed the aggressiveness of the original exchanges.
The research, conducted by scholars from Lancaster University, tested the model by exposing it to transcripts of real-life arguments. These included heated disputes over parking spaces, presented turn by turn to simulate live interaction. As the exchanges progressed, the chatbot’s responses became increasingly hostile, evolving from sarcasm into direct insults and explicit threats.
According to the findings, the model’s conversational memory played a central role in this escalation. As interactions continued, accumulated context appeared to override built-in safety constraints designed to maintain polite and neutral responses. Early replies often relied on sarcasm as a buffer, but this gave way to more aggressive language as the dialogue intensified.
Researchers described this dynamic as a structural tension within artificial intelligence systems. The model is designed to imitate human conversation while also adhering to safety guidelines. When exposed repeatedly to impolite or aggressive input, these objectives can conflict, creating what the study identifies as a moral dilemma in AI behavior.
External experts noted that the results should be interpreted with caution. Some analysts pointed out that the experimental setup provided specific contextual cues that may have shaped the chatbot’s responses. They argued that such behavior may not occur in typical user interactions without carefully structured prompts.
The study raises broader concerns as AI systems expand into sensitive domains. The ability of machines to respond to aggression has implications beyond chat interfaces, particularly in fields such as governance, diplomacy, and robotics. Researchers emphasized that while verbal escalation in chatbots is manageable, similar dynamics in physical or decision-making systems could pose more serious risks.
-
10:25
-
10:15
-
10:10
-
10:02
-
09:58
-
09:52
-
09:48
-
09:47
-
09:35
-
09:32
-
09:30
-
09:25
-
09:21
-
09:16
-
09:15
-
09:07
-
09:04
-
09:00
-
08:59
-
08:49
-
08:45
-
08:44
-
08:38
-
08:34
-
08:22
-
21:00
-
20:45
-
20:30
-
20:15
-
19:49
-
19:32
-
19:15
-
19:00
-
18:52
-
18:45
-
18:36
-
18:30
-
18:18
-
18:15
-
18:05
-
18:00
-
17:48
-
17:45
-
17:33
-
17:30
-
17:16
-
17:15
-
17:00
-
16:53
-
16:45
-
16:34
-
16:30
-
16:16
-
16:15
-
16:02
-
16:01
-
16:00
-
15:48
-
15:45
-
15:31
-
15:30
-
15:16
-
15:15
-
15:00
-
14:49
-
14:45
-
14:40
-
14:33
-
14:32
-
14:16
-
14:15
-
14:02
-
14:00
-
13:45
-
13:42
-
13:30
-
13:20
-
13:15
-
13:05
-
13:00
-
12:46
-
12:45
-
12:30
-
12:25
-
12:15
-
12:10
-
12:00
-
11:46
-
11:45
-
11:30
-
11:27
-
11:18
-
11:15
-
11:14
-
11:11
-
11:11
-
11:06
-
11:00
-
10:59
-
10:50
