Can ChatGPT reciprocate impoliteness? The AI moral dilemma

Abstract

Can AI ‘learn’ to be (verbally) violent? Unfortunately, it can. This paper tests ChatGPT 4.0 against real-life impolite interactions to assess whether it responds to human patterns of verbal conflict. We drew on the Principle of (Im)politeness Reciprocity (Culpeper & Tantucci, 2021) which posits that humans normally match the (im)politeness of one another. We thus prompted ChatGPT 4.0 turn by turn with authentic disputes, tracking its responses through network analysis and Bayesian regression. Our results reveal what we call the ‘AI moral dilemma’ (AI-MD): Large language models are constrained to avoid impoliteness through moderation and reinforcement learning, but are also designed to emulate human conversation, where (Im)politeness Reciprocity is intrinsic. When exposed to sustained impoliteness from real human disputes, the system's context-sensitive ‘working memory’ can override its moral safeguards, progressively leading it to reciprocate to impolite behaviour: AI can learn to 'strike back'. This contradiction –between being unconditionally moral and being human-like– raises pressing questions for AI ethics and the risks of replicating human conflict. This led to a second finding about AI's moral dilemma effects: while ChatGPT eventually learns to approximate human spirals of impoliteness reciprocity, it initially circumvents outright insults by resorting to implicational impoliteness (cf. Culpeper, 2011). This has important implications for AI's ability to simulate social intentions and Theory of Mind.