Updated 2026-11-09
from the news desk
Should you be polite to AI? What research, critics and Anthropic’s policy say
Scientific American reports evidence that polite prompts can improve some chatbot replies, and a PNAS Nexus paper found spillover from AI interactions to later judgments of people. A Frontiers perspective says the scale and meaning of…
There is no single rule, but the sources support a practical case for being polite to AI. Scientific American reports that polite prompts can improve some large language model responses, and a PNAS Nexus paper found that interactions with AI can spill over into harsher judgments of people. At the same time, a Frontiers perspective says the scale and meaning of abusive talk to AI remain unclear, so the ethical case is still debated.
Why some people answer yes
Webb Wright reports in Scientific American that one reason to be polite is simple prompting. In Wright’s account, a preprint by researchers at Waseda University and the RIKEN Center for Advanced Intelligence Project found that polite prompts could produce higher-quality replies from a large language model, or LLM, which is the kind of system that powers chatbots such as ChatGPT and Claude.
Wright also reports that the same preprint found limits to that effect. In his summary, excessive flattery could hurt performance, and the authors recommended what he described as moderate politeness.
Nathan Bos, whom Wright identifies as a senior research associate at Johns Hopkins University, gives a more mechanical explanation. Bos says: “'Please' indicates that what follows is a request, making it easier for the LLM to know how to respond.” In that reading, politeness is not magic. It is part of clear instruction format.
Wright also presents a second argument that is less about model output and more about human habits. He reports that Sherry Turkle of the Massachusetts Institute of Technology argues that politeness toward AI can help people preserve habits of civility toward other people. Autumn P. Edwards, whom Wright identifies as a professor of communication at Western Michigan University, makes the same point at the level of social norms: “Our interactions with AI will affect the evolution of human social norms as it seeps ever more steadily into our daily lives.”
That makes this more than a manners question. It also touches machine ethics: how systems are designed, and how people are expected to use them.
What the spillover evidence suggests
Kian Siong Tey and coauthors report in PNAS Nexus that they tested a related question directly. In two preregistered studies with a total of 1,261 participants, Tey and coauthors found that people who first worked with an AI later judged another purported human’s work more harshly than people who first worked with a purported human.
Tey summarizes one part of the result this way: “In two preregistered studies, participants who worked with AI (versus a purported human) were more demanding and instrumental.” The paper says this mattered because the later judgment task involved a different person and a different task. The authors call that a spillover effect, meaning one interaction carried over into a later human judgment.
The paper’s exploratory analyses also found differences in style. Tey and coauthors report that participants in the AI condition were more demanding and more instrumental, meaning more focused on using the partner as a tool for the task, and they showed less positive affect than participants in the human condition. Tey writes: “This negative spillover highlights a harmful downstream consequence of the rise in human–AI interactions.”
The same paper also narrows the claim. Tey and coauthors report that the two conditions did not significantly differ on politeness. So the study does not show that simply saying please or thank you is the decisive variable. One reading is that the broader tone of working with AI, not manners alone, may shape later judgments.
For readers asking should you be polite to AI, this paper supports a human-centered reason for caution. It does not say AI is harmed. It says human behavior around AI may affect later behavior toward people.
Why other scholars say the case is still unsettled
Eleonora Lima and Tiffany Morisseau offer the strongest skeptical case in their 2025 Frontiers in Artificial Intelligence perspective. Their focus is narrower than all chatbot use. They examine AI assistants, especially the long-running debate about insults and sexualized comments directed at voice assistants such as Siri, Alexa and Cortana.
Lima and Morisseau argue that the public discussion outran the evidence. They write: “The discourse on insults to AI assistants relies on a number of assumptions about the scale and nature of the issue, which needs further investigation.” They add a basic measurement problem: “Conversations between individuals and their AI assistants are not publicly available, making it impossible to determine the extent of the problem.”
That matters because several common claims depend on facts researchers may not yet have. Lima and Morisseau say the context of rude remarks is often unknown, and some interactions may be playful or exploratory rather than straightforward abuse. They also say there is low evidence that toxic behavior toward AI in real life transfers directly into toxic behavior toward people in real life.
Their paper does not deny that sexist or hostile AI settings can matter. Lima and Morisseau argue that AI responses can normalize stereotypes, especially when assistants are gendered. They write that “The normalization of these stereotypes should be challenged.” But they also warn against overstating what is proven, and they question whether making assistants more openly combative or moralizing would actually improve user behavior.
In their analysis, one risk is psychological reactance: users may resist or reject a message they see as controlling. Another is false equivalence between harm to women and insults to a female-sounding machine. So the skeptical position is not that politeness is pointless. It is that the strongest moral claims about abuse of AI still rest on incomplete evidence.
How policy has entered the question
The question has also moved from etiquette and psychology into product policy. In a usage policy update published on 8 October 2026, Anthropic states that it added a prohibition on sustained and needless abusive or cruel behavior toward its models.
Anthropic also narrows that rule. The update says it is meant for extreme cases in which users repeatedly act cruelly toward the models with no discernible purpose. Anthropic says the rule does not apply to common frustration, pushback, dark creative themes, or model testing and research.
That does not settle the moral debate, but it does show that at least one AI company treats persistent abuse of its models as a policy issue. Anthropic also says the addition aligns with an earlier product step that lets Claude end rare conversations with persistently abusive users on Claude.ai and Claude Code.
Policy, however, is not the same as moral status. A company can restrict conduct for safety, user experience, or brand reasons without claiming that the target has rights. Readers interested in that separate question can see our explainers on AI systems and moral status and on whether AI is sentient.
What this means in practice
Put together, the sources point to a modest answer. If your question is whether politeness can be useful, Wright’s Scientific American report and Tey’s PNAS Nexus paper both give reasons to say yes. Polite phrasing may help some outputs, and the broader style people adopt with AI may shape later judgments of humans.
If your question is whether you owe politeness to AI because AI can be wronged, the record here is thinner. Anthropic’s policy shows one company restricts some abusive conduct toward its own models, but Lima and Morisseau say the prevalence, meaning and real-world effects of rude talk to AI still need more investigation.
One practical reading is this: being polite to AI is a low-cost habit with some evidence behind it, but the strongest reason in these sources is about human behavior, not proven machine suffering. The opposing reading is also serious: if the evidence on abuse, transfer and intent is still incomplete, then people should avoid turning an unsettled research question into a settled moral rule.
So should you be polite to AI? On the record here, a careful answer is yes if you want clearer prompting, a more civil interaction style, or to stay within at least one company’s abuse rules. But the same record does not establish that every rude exchange with AI is morally equivalent to harming a person, and Lima and Morisseau argue that researchers should measure that question more directly before drawing stronger conclusions.
Frequently asked questions
Does saying please help AI answer better?
Webb Wright reports in Scientific American that a preprint by researchers at Waseda University and the RIKEN Center for Advanced Intelligence Project found that polite prompts could improve response quality from a large language model. Wright also reports that too much flattery could reduce performance, and Nathan Bos says please can signal that a prompt is a request, which may help the model interpret it.
Can being rude to AI affect how you treat people?
Kian Siong Tey and coauthors report in PNAS Nexus that participants who worked with an AI later judged another purported human more harshly than participants who worked with a purported human. But Lima and Morisseau argue in Frontiers in Artificial Intelligence that the broader claim that toxic behavior toward AI transfers to real-life behavior toward people still needs more direct evidence.
Do AI companies ban abuse toward their models?
Anthropic says its 8 October 2026 usage policy update added a prohibition on sustained and needless abusive or cruel behavior toward its models. Anthropic also says the rule is limited to extreme cases and does not cover common frustration, pushback, dark creative themes, or model testing and research. The sources here do not establish that all AI companies use the same rule.
Is being rude to AI evidence that someone is sexist or abusive offline?
Lima and Morisseau say that conclusion is not established by the available evidence they review. They argue that real conversations with AI assistants are not publicly available, so the scale and nature of abusive talk are hard to measure, and some interactions may be playful or context-dependent. Their paper says stronger claims need more systematic research.
caveats: Claims about polite prompts improving model output come from Webb Wright’s Scientific American report on underlying studies rather than from those papers directly. Anthropic’s abuse restriction is company-specific, and Lima and Morisseau’s discussion is focused mainly on AI assistants and voice-assistant debates rather than every kind of chatbot.
Sources
- Scientific American — Please Be Polite to ChatGPT — 2024-07-25
- Anthropic — 2026 Usage Policy update — 2026-10-08
- Frontiers in Artificial Intelligence — Can chatbots teach us how to behave? Examining assumptions about user interaction — 2025-05-14
- PNAS Nexus — People judge others more harshly after talking to bots — 2024-09-19