Because You Did Not Use the Magic Word

22 Sep, 2026

Politeness changes how language models answer. The reason may have nothing to do with manners and everything to do with the shape of the training data.

In Ghostbusters, Walter Peck of the Environmental Protection Agency walks into the firehouse and asks to see the storage facility where the team keeps its ghosts. Peter Venkman says no, and Peck asks why not. "Because you did not use the magic word," Venkman replies, and when Peck asks what the magic word is, the answer is "Please."

Nowadays researchers test whether the magic word works on language models, and the unsettling answer is that something happens. Change the social form of a request and the machine may change its answer. The popular reading arrived quickly: be nice to the AI and the AI will be nice back. Sam Altman fed the joke in April 2025, when a user on X wondered what all the pleases and thank yous cost OpenAI in electricity and he replied, "tens of millions of dollars well spent." Nobody should mistake that for an audited figure, but it confirms that the habit is common enough to be funny.

Steve Partridge, Please Close The Gate, Thank You, 2007. CC BY-SA 2.0, via Geograph and Wikimedia Commons.

The evidence disagrees with itself

The first serious test came from Ziqi Yin and colleagues in 2024. They wrote prompts at eight levels of politeness in English, Chinese and Japanese and ran them through several models and tasks. Their summary is careful: "impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes," and the best level of politeness differed by language. A year later Om Dobariya and Akhil Kumar ran a smaller experiment that pointed the other way. They took 50 questions in mathematics, science and history, rewrote each in five tones from very polite to very rude, and gave all 250 prompts to ChatGPT 4o. Accuracy rose from 80.8% for the very polite prompts to 84.8% for the very rude ones. One model and 50 questions should not overturn a larger literature, but the result is on the record.

Correctness also turns out to be the least of it. David Elsweiler, Christine Elsweiler and Anna Ziegner, in a paper for CHIIR 2026, simulated 18,000 conversations across five politeness profiles and three open weight models. Engagement seeking prompts drew replies up to 90% longer, with up to 38% more nuggets of information, than terse and efficient prompts did, and impolite input tended to produce wordy answers that delivered less information for the energy spent. Politeness changed the kind of continuation the model produced, not merely whether it was right. A preprint posted in September 2026 by Marina Mitiaeva and Lu Xiao adds a further caution. Comparing 15,881 conversations between people and ChatGPT with 10,784 conversations between people, they found that the linguistic signals of cooperation do not carry over intact. Hedging and softening, which mark accommodation between humans, were associated with reduced alignment when the AI produced them.

If a model rewarded courtesy the way a person does, these findings would be hard to reconcile. A courteous clerk does not serve rude customers better in one shop and worse in the next, in a pattern that varies by language and by employer. The scatter suggests the question has been framed badly. Asking whether the AI appreciates politeness assumes a social mechanism, and a statistical mechanism fits the mess much better.

Manners as metadata

The missing piece comes from a paper that never mentions politeness. In Findings of EMNLP 2023, Hila Gonen, Srini Iyer, Terra Blevins, Noah Smith and Luke Zettlemoyer set out to explain why some prompts work better than others for the same task. Their hypothesis was that "the performance of a prompt is coupled with the extent to which the model is familiar with the language it contains." They measured familiarity with perplexity, which is roughly how surprised a model is by a stretch of text, and across a wide range of tasks the prompts with lower perplexity performed better. Their reasoning ran through training frequency: language that appears more often in the training data, or resembles language that does, should be more familiar to the model. Training corpora usually cannot be searched, so perplexity stands in for the count.

Put the two literatures side by side and a hypothesis falls out. Polite request forms may change model behavior because they are statistically familiar forms in the language the models learned from. "Please explain" and "Could you please explain" are among the most recognizable constructions in English, and in the text a model has absorbed they are followed, overwhelmingly, by explanations, answers and help. For a human listener "please" expresses courtesy. For a language model it may also do a second job, locating the prompt inside a familiar pattern of request followed by assistance. What people regard as manners may look like metadata to a language model.

The hypothesis would also explain why the studies disagree. Different models have different training corpora, different post training and different learned associations, so the same tone lands in a different region of each model's experience. Pretraining is only part of that experience. InstructGPT, the ancestor of the modern chat model, was tuned on prompts written by labelers and prompts submitted by real API users, then shaped by human demonstrations and preference rankings. Familiarity with polite requests could arise at any of those stages, and a closed model offers no way to tell them apart.

None of it has been demonstrated, and the gap deserves a plain statement. One literature shows that familiarity predicts how well a prompt performs. Another shows that politeness changes what a model does. No paper located for this post connects them by showing the full chain from polite construction to greater familiarity to lower perplexity to changed performance.

Explain please photosynthesis

It is, however, a hypothesis that could be tested. The tempting comparison is the wrong one. Setting "Photosynthesis." against "Please explain photosynthesis." changes far more than politeness, because a bare noun could be a heading, a search term or a label, and only the second version states a task. A clean design holds the instruction constant and varies only the courtesy.

Explain photosynthesis.
Please explain photosynthesis.
Explain photosynthesis, please.
Could you please explain photosynthesis?
Explain please photosynthesis.

The last line carries the experiment. It contains exactly the same added word as the second line, placed where no fluent speaker would put it. If the token does the work, every version containing "please" should behave about the same. If familiar structure does the work, the natural constructions should beat the awkward one. Run the five forms, plus a rude one, across 500 or so factual, mathematical and reasoning tasks, measure perplexity alongside accuracy, length and information density, and ask whether the apparent effect of politeness shrinks once perplexity is controlled. The natural platform is OLMo, because Ai2 releases the model, the training recipes and the Dolma corpus of trillions of tokens behind it. The frequency of each construction could be counted in the training data directly instead of inferred.

"Thank you" needs its own design, since gratitude that arrives after an answer cannot improve that answer. It does, however, become part of the context for the next turn, and human conversation is full of sequences in which a request, a satisfactory reply and a word of thanks are followed by more cooperation. The test is to hold a first exchange constant, follow it with "Thank you," with "OK," or with nothing, and then pose an identical second task. The prediction would not be that gratitude makes the model happy. It would be that gratitude changes the statistical context from which the next words are drawn.

The human half of this story is much older than ChatGPT. In 1999 Clifford Nass, Youngme Moon and Paul Carney had people work with a computer and then evaluate its performance. Participants rated the machine more favorably when that same machine asked the questions than when a different computer or a paper questionnaire did. They were sparing its feelings, as they would an interviewer's, and Nass and Byron Reeves had already argued in The Media Equation that people apply social rules to machines automatically, even when they know better. For decades the research question was why humans are polite to computers. Language models add the reverse question, which is why computers respond differently when humans are polite to them. The first belongs to psychology, and the second may belong partly to statistics. There is a loop in it as well: people wrote their manners into the corpus, the machine learned the shape of sentences with manners in them, and everyone was then surprised when manners moved the machine.

Venkman knew exactly why "please" mattered. He wanted Walter Peck to show some manners, and even then he kept the door shut. A language model presents the stranger case. It does not need to be respected, and it has no feelings to protect on this point. It may simply have seen the magic word, and the structure around it, often enough to know what usually comes next. Maybe "please" is not the magic word because the machine learned to appreciate good manners. Maybe it is the magic word because we trained the machine on ourselves.

Further Reading

From this blog

• Want to Go Eat Some Tacos? ¡Claro Que Sí! - another case of a small linguistic cue steering a whole system.
• How I Use AI to Write This Blog - the working method behind these posts, prompts included.

Sources

• Demystifying Prompts in Language Models via Perplexity Estimation - Gonen, Iyer, Blevins, Smith and Zettlemoyer, Findings of EMNLP 2023. Lower perplexity prompts perform better.
• Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance - Yin, Wang, Horio, Kawahara and Sekine, SICon 2024.
• Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy - Dobariya and Kumar, 2025. The 80.8% and 84.8% figures.
• Cooking Up Politeness in Human-AI Information Seeking Dialogue - Elsweiler, Elsweiler and Ziegner, CHIIR 2026. The 18,000 simulated conversations.
• Talking Past the Machine: Morality, Politeness, and Alignment in Human-AI Dialogue - Mitiaeva and Xiao, preprint, September 2026.
• Training language models to follow instructions with human feedback - Ouyang et al., 2022. The InstructGPT training process.
• Nass, Moon and Carney, "Are People Polite to Computers? Responses to Computer-Based Interviewing Systems," Journal of Applied Social Psychology 29(5), 1999.
• Reeves and Nass, The Media Equation, Cambridge University Press, 1996.
• OLMo and Dolma - Ai2's open models and open training corpus.
• Sam Altman on the cost of "please" and "thank you" - TechRadar, April 2025.

#AIData #Algorithms