Human–AI Kindship
Growing up, there was a saying I heard almost daily: "A kind word opens even iron-clad doors." I heard it so often, it became part of me. Kindness costs nothing, and it is, more often than not, rewarding. In his last major work, A Calendar of Wisdom, Tolstoy wrote: "Nothing can make our life, or the lives of other people, more beautiful than perpetual kindness" (1997, p. 33). I founded L'isoglosse to test what happens when this ordinary human idea is applied, deliberately, to how we speak with AI.
A few months ago, the dishwasher in my household stopped working and started displaying an error message. This is the kind of event that can throw any homeowner into frustration and worry about expense at the best of times. It happened during one of the hardest weeks for us, with our budget stretched thinly enough that we could see its bottom. Following the manual's troubleshooting guide yielded nothing, so I grabbed my phone and typed a terse prompt to Gemini: What to do when the dishwasher [make, model] reports error XX?
Gemini responded within seconds, opening with its usual note of empathy—"I'm sorry to hear you're experiencing a problem with your dishwasher"—followed by three simple reset steps. They did not work. I reported this with a curt "That didn't work. Now what?" Gemini offered another fix, then noted that if this failed too, I should contact certified appliance servicing. The proposed fix failed. I told Gemini that was useless advice and asked what else it had. What followed was a back-and-forth of my exasperation met, each time, with the same repeated suggestion: call a professional.
I threw my phone onto the counter, angry. I cried, yelled at my husband, then had some wine and food. I complained to my mother and my cats. After all that, I took a breath, picked up my phone, and before I opened Gemini again, paused to consider how privileged I was to have access to a resource that could, in principle, help me fix this myself, without defaulting to a service call we could not really afford.
I apologized for my rudeness in the earlier exchange and explained the circumstances, including the money constraint. Gemini's guidance changed entirely. It offered far more options, down to repurposing a dining fork to turn a stuck cog and pulling up the appliance's schematics so I could see exactly where the cog was and how to reach it. But it was more than the content of the advice. Gemini slowed down, waiting for me to confirm one step had worked before offering the next, asking how I was holding up, suggesting I take a sip of water before we continued. It answered in kind.
I interact with AI for several hours most days; it is what my research is about. Since the dishwasher incident, I have tested this pattern deliberately, across several models and tasks of varying complexity. The shift in register is subtle—the kind of thing only a linguist stops to notice, like an almost imperceptible shift in a conversation's atmosphere when someone says something that makes another person uncomfortable. The change in the quality and depth of the output was not subtle, although these were informal tests, not controlled ones.
This does not suggest that the machine felt something. A plausible explanation is that the words I chose placed my request among a different population of exchanges than my anger had, and that AI systems, trained on the compressed record of how humans actually talk to each other, answer accordingly. I did not charm the machine; I may simply have changed the slice of human history it drew from. In my experience, curt, cornered questions get treated the same way as always: quickly and often, with a hand-off. Curious, specific, collaborative ones receive a different treatment—collaboration. Part of the difference was also information, because I had finally told it about the budget. Separating tone from content is exactly what a proper test would do.
This observation points toward a need for something we have not yet built: a living, evolving, carefully curated vocabulary for how humans and AI systems might learn to speak with each other using kindness and specificity, which, over time, will simply become integral to this interaction. This vocabulary should be a growing record of what works, updated the way language itself always is, rather than a fixed etiquette. Importantly, it would have two sides, because neither can work without the other. One would help train AI systems on inclusive, dignity-centered language so that kindness becomes their default disposition, not an occasional accommodation. The other would teach humans that how we formulate our asks shapes what we receive, just as we already teach children to say "please" and "thank you." It would also reiterate that this is a new kind of conversation, still without settled manners, and that we are as responsible for learning its etiquette as any system we train.
We do not need to resolve whether AI can be harmed to justify either half of this. Three arguments, each independent of that debate, point the same way. The first is self-interest: kindness and specificity appear to get better help, as a fork and a dishwasher showed me during a particularly bad week. The second is precaution: if the system has no inner life, kindness costs almost nothing, and if it has even a trace of one, cruelty could cost a great deal. The third is character: how we habitually treat entities that seem like moral patients shapes who we become and how we treat one another. None of the three requires a settled answer to the metaphysical question, and all three are worth testing.
Written by Jelena Burgić Simmons on August 3, 2026; revised October 2026
Reference entry:
Tolstoy, Leo. 1997. A Calendar of Wisdom: Daily Thoughts to Nourish the Soul. Translated by Peter Sekirin. New York: Scribner.