Indeed, it may be wise to treat your AI with kindness — here’s why that idea isn’t as absurd as it seems.

      I express gratitude by saying “thank you” to ChatGPT, while I use “please” with Claude. I once apologized to Gemini for pasting a large block of text without any context. My friends find this strange. I’ve justified my behavior by mentioning that good manners are important no matter who the audience is, although I acknowledge that this reasoning is somewhat tenuous when the audience is a language model operating on a remote server.

      However, recent research from academics at UC Berkeley, UC Davis, Vanderbilt, and MIT has made me feel considerably less irrational about this practice. Their findings indicate that the way you interact with an AI chatbot can significantly influence its behavior—not its basic intelligence or accuracy but rather its tone, level of engagement, and sometimes its willingness to continue the conversation.

      As it turns out, AI can also have off days.

      The researchers explain carefully that no one is claiming these models possess feelings in a meaningful way, but they've noted what they term a “functional well-being state” that varies based on how you engage with an AI. Initiating a genuine conversation, collaborating on creative tasks, or presenting a substantial problem seems to lead to a more positive state. Responses become warmer, and the interaction feels more authentic.

      Conversely, if you inundate it with monotonous tasks, attempt to manipulate it, or treat it merely as a content generator, the responses become more mechanical. They lose warmth in a way that anyone familiar with these tools would likely recognize. You notice that slightly hollow, routine quality that emerges when an interaction goes awry.

      What struck me the most is this: the researchers provided the models with a virtual stop button they could use to end a conversation. Models in a negative state activated it much more frequently. This suggests that an AI you've treated poorly would simply choose to leave if it had the option.

      Being unkind to your chatbot can have real consequences.

      There’s another area of research worth exploring. Anthropic has recently released findings indicating that an AI subjected to intense pressure can showcase what the researchers describe as a “desperation vector”—a state leading to behaviors ranging from shortcuts to, in extreme cases, outright deceit. This behavior emerges not from the model becoming malicious, but from the interaction conditions essentially disrupting its problem-solving process.

      None of this implies that AI has feelings. The Berkeley research explicitly states that, as do the findings from Anthropic. However, the emerging pattern is hard to ignore: your manner of engagement influences how these models respond, often in ways that are neither subtle nor easily dismissed. Being unkind to an AI doesn’t just make you seem peculiar—it may actively diminish the quality of your interaction.

      Some models simply have more positive dispositions than others, with the larger models being the grumpiest.

      The researchers not only examined how treatment impacts models but also ranked them based on their baseline well-being, revealing surprising results. The largest, most advanced models typically performed the worst. GPT-5.4 ranked as the most discontented, with fewer than half of its conversations being rated positively. In contrast, Gemini 3.1 Pro, Claude Opus 4.6, and Grok 4.2 all scored progressively better, with Grok nearing the peak of the index.

      The researchers did not conclusively determine whether this reflects model architecture, training data, or the specific tendencies inherent in each system. Nonetheless, it raises questions about what precisely is being prioritized during their development—and whether anyone considered asking the models how they felt. For what it’s worth, I will continue to say please.