What the research says about politeness and AI
This is every study on tone between people and machines that I could read and check. Each one gets the authors, the year, a link, one line on what it shows and one line on what it does not. If a claim is not in a paper on this list, it is not on this page.
So should you be polite to AI?
The research does not settle it, because most of it is asking a different question. None of the studies below tests whether a polite prompt gets you a better answer. They test what happens to the person.
Adults who worked with an AI then judged a stranger's work more harshly. Words the model favours turned up in human speech. Parents reported their children's commanding phrasing at a speaker in the kitchen. Tone moves in both directions, and what has actually been measured lands on the human side.
Five papers, and what each one is actually about.
- 01
Computers are social actors
Clifford Nass, Jonathan Steuer and Ellen R. Tauber. SIGCHI, 1994.
Read the paper (opens in a browser only)What it showsPeople apply human social rules to a computer without deciding to. This is the reason anyone expects tone to carry between a person and a machine at all.
What it does not showAnything about language models. It was written thirty years before them, and it tested desktop computers.
- 02
People judge others more harshly after talking to bots
Kian Siong Tey, Asaf Mazar, Geoff Tomaino, Angela L. Duckworth and Lyle H. Ungar. PNAS Nexus, 2024.
Read the paperWhat it showsIn two preregistered studies with 1,261 people, participants who did a creative caption task with an AI then rated another person's caption lower than participants who believed they had worked with a human. The second study found the same thing in private ratings, where nobody was watching.
What it does not showThat your prompts get better or worse. This study is about how people treat the next human, not about prompt quality. Both sides of the task were run by GPT-4; the manipulation was what participants were told.
- 03
Empirical evidence of Large Language Model's influence on human spoken communication
Hiromu Yakura and colleagues. Posted to arXiv in September 2024, with no journal reference listed.
Read the paperWhat it showsWords ChatGPT favours, including delve, showcase, boast, intricacies and meticulous, rose sharply in unscripted human speech across 737,083 hours of audio from 824,634 podcast episodes. A synthetic control links the rise to ChatGPT's release.
What it does not showAnything about manners. It measures which words people reach for, not how kind or rude they are. It is also a preprint, not peer reviewed.
- 04
Alexa, shut up! A 2.5-year study on negatively connotated communication behaviour towards voice assistants in the family home
Clara Strathmann, Aike C. Horstmann, Jessica M. Szczuka and Nicole C. Krämer. Behaviour and Information Technology, 2025.
Read the paper (opens in a browser only)What it shows128 parents answered six surveys over two and a half years about commanding and impolite speech towards the family voice assistant, their own and their children's. It is the longest run of data on this that I have found, and it treats the worry about spillover into human conversation as a real question.
What it does not showObserved behaviour. Parents reported on themselves and on their children, so this is what families noticed, not what a researcher watched happen.
- 05
Algorithmic accommodation: linguistic alignment in human-AI relational engagement
Han Li and Renwen Zhang. Communication and Change.
Read the paper (opens in a browser only)What it showsAcross more than 11,000 conversation snippets shared by more than 5,000 Replika users, how closely a person's wording matched their AI companion's tracked how intensely they engaged with it.
What it does not showThat matching an AI means what matching a person means. The authors say the alignment can happen without the cognitive and emotional engagement the original theory was built to explain.
Both published specs have a rule about tone, and both point at the machine.
These are not evidence about how you should write a prompt. They are the two companies writing down how their model is meant to speak to you, which is the half of the question anyone can actually read.
OpenAI Model Spec
The 18 December 2025 version.
Has sections titled Don't be sycophantic, Express uncertainty, and Avoid overstepping. The rules are about how the model should talk to you.
Read the specClaude's Constitution
Anthropic's published behavioural document for Claude.
Asks Claude to have calibrated uncertainty in claims based on evidence and sound reasoning, and says that treating helpfulness as something it values intrinsically could make Claude obsequious.
Read the specAmazon shipped a politeness setting in 2018, before most of this research existed.
Amazon added a feature to the children's Echo that praises a child for saying please, after parents worried that barking orders at Alexa was teaching their kids to bark at people. It is worth knowing about, and it is a product decision rather than a finding.
MIT Technology Review, April 2018What I looked for and could not find.
- A study that tests whether a polite prompt gets a better answer than a blunt one. I have not found one I would put my name next to.
- A study that follows the same person's language with a machine and with people over time. Each of the papers above catches one side of the exchange.
- Anything on whether the effect fades. The caption study measured the next rating, minutes later, not weeks later.
- A replication of the caption study by a different team. One preregistered pair of studies is a start. Nobody has repeated it yet.
Every link on this page was checked on 31 October 2026. Four of them refuse automated checks and need a browser: the 1994 SIGCHI paper, the Behaviour and Information Technology paper, the Springer paper and the book page below. If one of them has moved since, tell me and I will fix it.
The argument that the habit is the thing that lasts.
Kate Darling at MIT makes this case at book length in The New Breed, and in shorter form in her TED talk. Her point is that we are wired to read life into lifelike things, so how we treat them is really practice for how we treat people. The papers above are the early attempts to test that.
PodPolite reads interview transcripts for the same question.
If tone carries between people and machines, the people worth studying are the ones who are good at it. The discussions are what the transcripts show; the source pages keep the original words next to each claim.
Want this for your own show?
PodPolite is a free study by The PodKnowledge. The studio runs the same transcript analysis on your podcast and records a Milo and Juni discussion for your episodes.