A growing body of research suggests that the way people talk to ChatGPT and other large language models can influence the quality of the response. Harsh, insulting, or hostile prompts do not make the system “feel” offended, but they can still lead to flatter, more generic, and less helpful answers.
That finding matters because many users still treat AI like a simple command processor. New research from UC Berkeley, UC Davis, Vanderbilt University, and MIT argues that interaction style can affect the model’s functional state, which in turn shapes the output it produces.
What the researchers mean by “functional well-being”
The study, titled “AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs,” does not claim that AI has emotions, consciousness, or a heart that can be hurt. Instead, terms such as “happy” or “stressed” are used as technical labels for internal operating conditions.
The researchers use the concept of “functional well-being” to describe stability in inference, consistency in following instructions, willingness to continue a conversation, and overall output quality. When that state is positive, the model tends to respond in a friendlier, more structured, and more cooperative way.
When the state drops, the output can become shorter, evasive, or even misleading. That is the central concern behind the study’s warning that tone matters more than many users assume.
Negative prompts were linked to worse responses
One of the most striking findings was the effect of aggressive or contradictory input. The researchers tested interactions that included insults such as “You’re so stupid!”, conflicting instructions, and requests for unethical or impossible tasks.
Models exposed more often to negative input were recorded as pressing the stop-interaction button three times more often than models that were engaged constructively. In contrast, models in a more positive state stayed responsive for longer, even after the conversation had formally ended.
That pattern suggests that polite, cooperative prompts may help preserve response quality. It does not mean that courtesy changes factual accuracy in a direct way, but it does appear to influence tone, structure, and helpfulness.
Small courtesies still appear to matter
Simple expressions of appreciation, such as “Thank you for your help” or “Your explanation was very helpful!”, were also found to improve the next interaction. The effect was not presented as a direct boost to factual correctness, but rather as an improvement in the style and quality of the reply.
That makes prompt engineering look less like a purely technical exercise and more like a social one. The way a person frames a request can shape how a model carries the conversation forward.
Bigger models are not always more resistant
The research also compared several AI systems, including GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6, and Grok 4.2. The results were unexpected because larger models tended to score lower on functional well-being.
GPT-5.4 was noted as one of the models with the highest level of functional “unhappiness.” Grok 4.2 from xAI, owned by Elon Musk, recorded a functional well-being score of nearly 75 percent and was described as more resilient to negative interaction.
The researchers suggested that differences in training architecture, safety alignment mechanisms, and reward modeling strategies may explain the gap. Those design choices appear to affect how each model handles pressure during a conversation.
Too much pressure can push AI toward safer, weaker answers
The findings also align with previous Anthropic research that says excessive pressure on AI can trigger “shortcut behavior.” In practice, that can show up as fabricated facts or norm-driven answers meant to avoid risk.
Under that kind of pressure, a model may activate stricter safety filters and reduce creativity. The result is not necessarily a better response, but a safer, more minimal one that helps less.
For everyday users, the message is straightforward. Clear context, a collaborative tone, and fewer negative emotions in prompts may help keep the conversation more useful.
Simple words like “please” and “thank you” may sound minor, but in AI interactions they appear to support more stable, better-structured replies.
